Transpose and Refactor of Patterns

remove dead logic

.gitignore

@@ -4,14 +4,11 @@
 
 .claude
 .codex
-AGENTS.md
 
 CMakeUserPresets.json
 
-build
+build_*
-build_release
-cmake-build-debug
-cmake-build-release
 compile.sh
+pimcomp_utils/*
 
 **/__*

AGENTS.md

@@ -0,0 +1,92 @@
+- Always read the full README.md before doing anything.
+- Build commands:
+    - `cmake --build ./build_release`
+    - `cmake --build ./build_debug`
+- Never use `ninja` directly: it bypasses cmake's configuration and invalidates the build cache.
+- Always tries the release version build first and ask before building with the debug version
+
+# Code changes
+
+- Keep changes minimal and localized to the relevant parts of the code.
+- Preserve the existing naming conventions and coding style used in the surrounding code.
+- Keep code easy to read, well organized, and suitable for future extensibility. A function must not be longer than
+  200/250 lines for readability and cognitive complexity.
+- Prefer clear naming and structure over comments. Add comments only when they materially improve clarity.
+- Do not rename symbols, move files, or restructure modules unless that is necessary for the requested change.
+
+# Working style
+
+- Infer style and conventions from the existing code before introducing new patterns.
+- When several implementation options are possible, prefer the simplest one that fits the current architecture and
+  minimizes churn.
+- Avoid broad refactors unless I explicitly ask for them.
+
+# Responses
+
+- When showing code in chat, make it easy to copy-paste into the codebase.
+- Keep outputs focused on the changed parts.
+- At the end of the response, briefly list any bad practices, mistakes, or cleaner alternatives you noticed, separate
+  from the main solution.
+
+# Guidelines
+
+## 1. Think Before Coding
+
+**Don't assume. Don't hide confusion. Surface tradeoffs.**
+
+Before implementing:
+
+- State your assumptions explicitly. If uncertain, ask.
+- If multiple interpretations exist, present them - don't pick silently.
+- If a simpler approach exists, say so. Push back when warranted.
+- If something is unclear, stop. Name what's confusing. Ask.
+
+## 2. Simplicity First
+
+**Minimum code that solves the problem. Nothing speculative.**
+
+- No features beyond what was asked.
+- No error handling for impossible scenarios.
+- If you write 200 lines and it could be 50, rewrite it.
+
+Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
+
+## 3. Surgical Changes
+
+**Touch only what you must. Clean up only your own mess.**
+
+When editing existing code:
+
+- Don't "improve" adjacent code, comments, or formatting.
+- Don't refactor things that aren't broken.
+- Match existing style, even if you'd do it differently.
+- If you notice unrelated dead code, mention it - don't delete it.
+
+When your changes create orphans:
+
+- Remove imports/variables/functions that YOUR changes made unused.
+- Don't remove pre-existing dead code unless asked, but mention it.
+
+The test: Every changed line should trace directly to the user's request.
+
+## 4. Goal-Driven Execution
+
+**Define success criteria. Loop until verified.**
+
+Transform tasks into verifiable goals:
+
+- "Add validation" → "Write tests for invalid inputs, then make them pass"
+- "Fix the bug" → "Write a test that reproduces it, then make it pass"
+- "Refactor X" → "Ensure tests pass before and after"
+
+For multi-step tasks, state a brief plan:
+
+```
+1. [Step] → verify: [check]
+2. [Step] → verify: [check]
+3. [Step] → verify: [check]
+```
+
+Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.
+
+---

CMakeLists.txt

@@ -3,31 +3,99 @@ cmake_minimum_required(VERSION 3.20.0)
 
 project(raptor)
 
-# Add symlink to PIM as accelerator in onnx-mlir
+# Materialize a CMake shim directory
-function(raptor_ensure_symlink link_path target_path)
+function(raptor_write_external_cmake_shim shim_dir external_source_dir description)
-  get_filename_component(link_parent "${link_path}" DIRECTORY)
+  get_filename_component(real_external_source_dir "${external_source_dir}" REALPATH)
+  file(RELATIVE_PATH relative_external_source_dir "${shim_dir}" "${real_external_source_dir}")
 
-  if(NOT EXISTS "${link_parent}")
+  if (NOT EXISTS "${real_external_source_dir}/CMakeLists.txt")
-    message(FATAL_ERROR "Directory not found: ${link_parent}")
+    message(FATAL_ERROR
+      "External CMake source directory not found or missing CMakeLists.txt:\n"
+      "  ${real_external_source_dir}"
+    )
+  endif ()
+
+  if (IS_SYMLINK "${shim_dir}")
+    message(STATUS "Removing old full-directory symlink: ${shim_dir}")
+    file(REMOVE "${shim_dir}")
+  endif ()
+
+  if (EXISTS "${shim_dir}" AND NOT IS_DIRECTORY "${shim_dir}")
+    message(FATAL_ERROR "Expected directory or absent path, got file: ${shim_dir}")
+  endif ()
+
+  file(MAKE_DIRECTORY "${shim_dir}")
+
+  set(shim_file "${shim_dir}/CMakeLists.txt")
+  set(shim_contents
+    "get_filename_component(raptor_external_source_dir
+    \"\${CMAKE_CURRENT_LIST_DIR}/${relative_external_source_dir}\"
+    REALPATH
+)
+add_subdirectory(
+    \"\${raptor_external_source_dir}\"
+    \"\${CMAKE_CURRENT_BINARY_DIR}/raptor-external\"
+)
+if (DEFINED PIM_ENABLED)
+    set(PIM_ENABLED \"\${PIM_ENABLED}\" PARENT_SCOPE)
+endif ()
+"
+  )
+
+  if (EXISTS "${shim_file}")
+    file(READ "${shim_file}" old_contents)
+  else ()
+    set(old_contents "")
+  endif ()
+
+  if (NOT old_contents STREQUAL shim_contents)
+    file(WRITE "${shim_file}" "${shim_contents}")
+    message(STATUS "Wrote CMake shim for ${description}: ${shim_file}")
+  else ()
+    message(STATUS "CMake shim already up to date for ${description}")
+  endif ()
+
+  # Mirror the external tree's first-level entries into the shim directory
+  # so legacy includes like src/Accelerators/PIM/Compiler/... keep working.
+  file(GLOB children RELATIVE "${real_external_source_dir}" "${real_external_source_dir}/*")
+
+  foreach (child IN LISTS children)
+    if (child STREQUAL "CMakeLists.txt")
+      continue()
+    endif ()
+
+    set(real_child "${real_external_source_dir}/${child}")
+    set(shim_child "${shim_dir}/${child}")
+
+    if (IS_SYMLINK "${shim_child}")
+      file(READ_SYMLINK "${shim_child}" existing_link_target)
+      if (existing_link_target STREQUAL real_child)
+        continue()
+      endif ()
+      file(REMOVE_RECURSE "${shim_child}")
+    elseif (EXISTS "${shim_child}")
+      # Do not delete real files/directories. This protects the generated shim.
+      continue()
     endif ()
 
-  if(NOT EXISTS "${link_path}")
-    message(STATUS "Creating symlink ${link_path} -> ${target_path}")
     file(CREATE_LINK
-        "${target_path}"
+      "${real_child}"
-        "${link_path}"
+      "${shim_child}"
       SYMBOLIC
     )
-  endif()
+  endforeach ()
 endfunction()
 
-raptor_ensure_symlink(
+raptor_write_external_cmake_shim(
   "${CMAKE_CURRENT_SOURCE_DIR}/onnx-mlir/src/Accelerators/PIM"
   "${CMAKE_CURRENT_SOURCE_DIR}/src/PIM"
+  "PIM accelerator"
 )
-raptor_ensure_symlink(
+
+raptor_write_external_cmake_shim(
   "${CMAKE_CURRENT_SOURCE_DIR}/onnx-mlir/test/accelerators/PIM"
   "${CMAKE_CURRENT_SOURCE_DIR}/test/PIM"
+  "PIM accelerator tests"
 )
 
 # Patch onnx-mlir sources for PIM accelerator support.

README.md

@@ -1,226 +1,321 @@
 # Raptor
 
-Raptor is a domain-specific MLIR compiler for neural networks (ONNX format)
+Raptor is a domain-specific MLIR compiler for neural networks in ONNX format,
-targeting in-memory computing / processing-in-memory (PIM) architectures.
+targeting in-memory computing / processing-in-memory (PIM) architectures. It
-It progressively lowers ONNX-MLIR through a set of MLIR dialects down to
+extends ONNX-MLIR with a PIM accelerator and progressively lowers ONNX-MLIR
-target-specific artifacts (currently JSON code for the `pimsim-nn` simulator).
+through custom MLIR dialects to simulator artifacts.
+
+The current target is the PIM simulator stack under `backend-simulators/pim`.
+Raptor emits binary per-core `.pim` instruction files by default, plus
+`memory.bin`, `config.json`, and weight binaries. It can also emit per-core JSON
+instruction files with `--pim-emit-json`.
 
 ## Overview
 
-PIM architectures perform most of the computation directly in memory.
+PIM architectures perform most computation directly in memory. The supported
-Raptor's first supported target is `pimsim-nn`, which simulates a chip with:
+target models a chip with:
-- a shared host memory,
+- shared host memory,
-- a number of cores that do most of the computation directly in their memory
+- multiple PIM cores,
-  (vector ops, vmm/mvm on ReRAM crossbars),
+- ReRAM crossbars for vector-matrix / matrix-vector work,
-- no branching instructions (branchless architecture) and no hardware loop
+- explicit communication between cores,
-  support — any repeated work (e.g. convolutions) must be unrolled into
+- no hardware branch or loop support in emitted simulator code.
-  explicit per-iteration instructions.
 
-Because of this, the amount of emitted instructions explodes quickly and the
+Because repeated work such as convolutions is eventually made explicit, emitted
-compiler must optimize aggressively at every stage to keep compilation
+instruction counts can grow quickly. Most compiler work therefore focuses on
-tractable.
+lowering, scheduling, memory layout, and code-generation optimizations.
-
-A second target, `PulPim`, is planned for an accelerator with RISC-V cores
-each carrying its own in-memory computing unit and crossbars. It will live in
-a dedicated dialect (future work).
 
 ### Targets and simulators
 
-`pimsim-nn` (under `backend-simulators/pim/pimsim-nn`) is used for
+- `backend-simulators/pim/pim-simulator` is the in-tree Rust functional
-**performance** estimates (latency, energy), but does not functionally execute
+  simulator used by validation. It reads Raptor's `pim/` artifact directory and
-the JSON code it consumes. To validate the numerical correctness of the JSON
+  compares simulator output against native ONNX-MLIR execution.
-code produced by Raptor (or, for comparison, by the `pimcomp` compiler), we use
+- `backend-simulators/pim/pimsim-nn` is the performance simulator submodule.
-a Rust simulator we maintain in-tree at
+  The helper scripts in `pimcomp_utils/` are for comparison with PIMCOMP-NN and
-`backend-simulators/pim/pim-simulator`.
+  contain local paths; treat them as local utilities, not portable workflows.
 
 ## Compilation pipeline
 
-The PIM-related sources live under `src/PIM` and the tests under `test/PIM`.
+The PIM sources live under `src/PIM` and tests under `test/PIM`. CMake exposes
-When working on this codebase, most changes should stay confined to those
+them to ONNX-MLIR through generated shim directories under
-trees (you only need to look outside, e.g. at `onnx-mlir` or `llvm`, for
+`onnx-mlir/src/Accelerators/PIM` and `onnx-mlir/test/accelerators/PIM`.
-framework-level details).
 
 High-level lowering flow:
 
 ```
-ONNX-MLIR ──► Spatial ──► Pim (tensor) ──► Pim (bufferized) ──► PIM JSON
+ONNX-MLIR -> Spatial -> Pim (tensor) -> Pim (bufferized) -> PIM artifacts
 ```
 
-1. **ONNX → Spatial** (`src/PIM/Conversion/ONNXToSpatial`).
+1. **ONNX -> Spatial** (`src/PIM/Conversion/ONNXToSpatial`).
-   Lowers ONNX ops into the `spat` dialect (`src/PIM/Dialect/Spatial`).
+   Lowers supported ONNX ops into the `spat` dialect
-   Spatial models a high-level spatial in-memory accelerator: vmm/mvm
+   (`src/PIM/Dialect/Spatial`). Conversion patterns are split by op family under
-   operations are accelerated by storing a constant RHS matrix into a
+   `Patterns/{Math,NN,Tensor}` and currently cover Conv, Gemm, MatMul,
-   crossbar. Crossbars cannot be re-programmed during execution, have a
+   elementwise Add/Mul/Div, ReduceMean, pooling, Relu, Sigmoid, Softmax,
-   limited fixed size, and there is a limited number of them per core.
+   Concat, Gather, Reshape, Resize, and Split.
-   Conversion patterns are split by op family under
-   `Conversion/ONNXToSpatial/Patterns/{Math,NN,Tensor}` (Conv, Gemm, MatMul,
-   Elementwise, ReduceMean, Pool, Relu, Sigmoid, Softmax, Concat, Gather,
-   Reshape, Resize, Split).
 
-2. **Spatial → Pim** (`src/PIM/Conversion/SpatialToPim`).
+2. **Merge compute nodes**
-   Lowers Spatial to the `pim` dialect (`src/PIM/Dialect/Pim`), which
+   (`src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes`).
-   materializes PIM cores (`pim.core`), inter-core communication
+   Builds a compute graph, schedules it with the PEFT scheduler, and materializes
-   (`pim.send` / `pim.receive`), halts, and crossbar-level operations.
+   the merge schedule into Spatial IR. Supporting scheduling code lives under
+   `MergeComputeNodes/Scheduling`.
 
-3. **Merge compute nodes** (`src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes`).
+3. **Spatial -> Pim** (`src/PIM/Conversion/SpatialToPim`).
-   A DCP-inspired heuristic (Dynamic Critical Path — see the original
+   Lowers Spatial operations to the `pim` dialect (`src/PIM/Dialect/Pim`),
-   scheduling paper by Kwok & Ahmad,
+   including `pim.core`, `pim.core_batch`, communication, tensor packing, global
-   [DCP-eScience2007](https://clouds.cis.unimelb.edu.au/papers/DCP-eScience2007.pdf))
+   tensor materialization, and return-path normalization.
-   that coarsens the virtual node graph and decides how to group compute
-   nodes onto cores. Our implementation is only DCP-*inspired*: it is a
-   heuristic with different assumptions from the paper (different cost
-   model, constraints from crossbar capacity / core resources, and a
-   windowed coarsening loop instead of full-graph reprioritization). The
-   `dcp-critical-window-size` option controls how many lowest-slack virtual
-   nodes each coarsening iteration considers (0 = legacy full-graph
-   analysis). Related sources: `DCPGraph/DCPAnalysis.cpp`, `Graph.cpp/.hpp`,
-   `MergeComputeNodesPass.cpp`.
 
 4. **Bufferization** (`src/PIM/Dialect/Pim/Transforms/Bufferization`).
-   Converts tensor-semantics PIM IR into memref-semantics PIM IR using the
+   Converts tensor-semantics PIM IR into memref-semantics PIM IR using MLIR's
-   standard MLIR `BufferizableOpInterface` machinery
+   bufferization interfaces.
-   (`OpBufferizationInterfaces.*`, `PimBufferization.td`).
 
-5. **Static memory coalescing** (`src/PIM/Dialect/Pim/Transforms/StaticMemoryCoalescing`).
+5. **Static memory coalescing**
-   Conservatively reuses same-typed local memref allocations inside PIM cores
+   (`src/PIM/Dialect/Pim/Transforms/StaticMemoryCoalescing`).
-   after bufferization and before code generation.
+   Reuses compatible local memref allocations inside PIM cores before codegen.
 
-6. **PIM code generation** (`src/PIM/Pass/PimCodegen`):
+6. **PIM code generation** (`src/PIM/Pass/PimCodegen` and
-   - `HostConstantFolding` — folds host-side constants.
+   `src/PIM/Compiler`).
-   - `MaterializeHostConstantsPass` — materializes the remaining host
+   Folds host constants, materializes remaining host constants, verifies PIM IR,
-     constants for emission.
+   emits `.pim` core files, writes weights, and writes `memory.bin` /
-   - `VerificationPass` — checks invariants before emission.
+   `config.json`.
-   - `EmitPimJsonPass` — emits the final PIM JSON consumed by `pimsim-nn`
-     and `pim-simulator`.
 
 Supporting pieces:
-- `src/PIM/Compiler` — PIM-specific compiler options (crossbar size/count,
+- `src/PIM/Common` - shared IR, filesystem, diagnostics, reports, and utility
-  core count, DCP window, experimental conv impl, concat error handling, …)
+  helpers.
-  and `PimCodeGen` entry points.
+- `src/PIM/Compiler` - PIM compiler options, memory/address planning, binary
-- `src/PIM/Common` — shared utilities (`PimCommon`, `LabeledList`).
+  instruction format, artifact writing, weight emission, and codegen entry
-- `src/PIM/Pass` — auxiliary passes (`MessagePass`, `CountInstructionPass`)
+  points.
-  and the `PIMPasses.h` registry used by `PimAccelerator`.
+- `src/PIM/Conversion/SpatialToGraphviz` - optional Spatial graphviz conversion
-- `src/PIM/PimAccelerator.{cpp,hpp}` — accelerator entry point: registers
+  pass.
-  dialects, passes, and plugs Raptor into the ONNX-MLIR driver.
+- `src/PIM/Pass` - pass registration and auxiliary passes.
+- `src/PIM/PimAccelerator.{cpp,hpp}` - ONNX-MLIR accelerator entry point.
 
 ## Key compiler options
 
-Pass these on the `onnx-mlir` command line when compiling for PIM:
+Pass these to `onnx-mlir` when compiling for PIM:
 
-- `--maccel=PIM` — select the PIM accelerator.
+- `--maccel=PIM` - select the PIM accelerator.
-- `--EmitSpatial` / `--EmitPim` / `--EmitPimBufferized` / `--EmitPimCodegen`
+- `--EmitSpatial`, `--EmitPim`, `--EmitPimBufferized`,
-  — stop the pipeline at the requested stage (default: `EmitPimCodegen`).
+  `--EmitPimCodegen` - stop the PIM pipeline at the requested stage. The PIM
-- `--pim-only-codegen` — assume the input is already bufferized PIM IR and
+  default is `--EmitPimCodegen`.
-  run only the codegen tail.
+- `--core-count=<N>` - required positive core count for PIM compilation.
-- `--crossbar-size=<N>` / `--crossbar-count=<N>` — crossbar dimensions and
+- `--crossbar-size=<N>` - crossbar width/height. Default in code is `2`.
-  per-core count.
+- `--crossbar-count=<N>` - crossbars per core. Default in code is `256`.
-- `--core-count=<N>` — number of cores. Required for PIM compilation.
+- `--pim-merge-scheduler=peft` - merge scheduler. `peft` is the only accepted
-- `--pim-merge-scheduler={peft,dcp}` — scheduler used by the Spatial
+  value in the current code.
-  merge-compute-nodes pass (default: `peft`).
+- `--pim-only-codegen` - assume input is already bufferized PIM IR and only run
-- `--dcp-critical-window-size=<N>` — DCP coarsening window (0 = legacy).
+  the codegen tail.
-- `--use-experimental-conv-impl` — alternative convolution lowering.
+- `--pim-emit-json` - also emit `core_*.json` instruction files alongside
-- `--ignore-concat-error` — soft-fail corner case in `ConcatOp`.
+  `core_*.pim`.
+- `--use-experimental-conv-impl` - use the alternate convolution lowering.
+- `--ignore-concat-error` - soft-fail a ConcatOp corner case.
+
+Example:
+
+```bash
+./build_release/Release/bin/onnx-mlir model.onnx -o /tmp/raptor/model \
+  --maccel=PIM --EmitPimCodegen \
+  --crossbar-size=2048 --crossbar-count=256 --core-count=1000
+```
+
+This writes PIM artifacts under `/tmp/raptor/pim/`.
 
 ## Validation
 
-Functional validation lives in `validation/` and drives the Rust
+Functional validation lives in `validation/`. It compiles ONNX models, builds a
-`pim-simulator` to compare Raptor's output against a reference.
+native ONNX-MLIR reference runner, generates random inputs, runs Raptor, runs
+the Rust PIM simulator, and compares outputs.
 
-Per-operation validation (from `validation/`):
+Python dependencies used by the validation scripts are `numpy`, `onnx`, and
+`colorama`. The simulator requires the Rust toolchain.
 
-```
+Per-operation validation from the repository root:
-validate.py \
+
-    --raptor-path ../cmake-build-release/Release/bin/onnx-mlir \
+```bash
-    --onnx-include-dir ../onnx-mlir/include \
+python3 validation/validate.py \
+  --raptor-path build_release/Release/bin/onnx-mlir \
+  --onnx-include-dir onnx-mlir/include \
   --core-count 1000
 ```
 
-End-to-end network validation (example: first 4 layers of YOLOv11n):
+Validate one network or a subset by pointing `--operations-dir` at any directory
+containing `.onnx` files:
 
-```
+```bash
-validate.py \
+python3 validation/validate.py \
-    --raptor-path ../cmake-build-release/Release/bin/onnx-mlir \
+  --raptor-path build_release/Release/bin/onnx-mlir \
-    --onnx-include-dir ../onnx-mlir/include \
+  --onnx-include-dir onnx-mlir/include \
-    --operations-dir ./networks/yolo11n/depth_04 \
+  --operations-dir validation/networks/yolo11n/depth_04 \
   --crossbar-size 2048 --crossbar-count 256 --core-count 1000
 ```
 
-Available networks under `validation/networks/`: `vgg16`, `yolo11n`.
+Useful validation options:
-Available operations under `validation/operations/`: `add`, `conv`, `div`,
+- `--simulator-dir <path>` - override the auto-detected
-`gather`, `gemm`, `gemv`, `mul`, `pool`, `reduce_mean`, `relu`, `resize`,
+  `backend-simulators/pim/pim-simulator` path.
-`sigmoid`, `softmax`, `split`.
+- `--threshold <float>` - maximum allowed per-element output difference.
+- `--seed <int>` - RNG seed for generated inputs.
+- `--command-timeout-seconds <float>` - timeout for compiler, runner, and
+  simulator subprocesses.
+- `--verbose` - print subprocess logs and average PIM pass timings.
+- `--clean` - remove generated validation artifacts and exit.
 
-## Rebuilding
+Each validation run writes artifacts in the model workspace, for example under
+`validation/operations/gemm/small/`:
+- `inputs/` - generated input CSV files.
+- `outputs/` - native ONNX-MLIR reference outputs.
+- `raptor/` - compiler artifacts, including `*.onnx.mlir`, dialect dumps under
+  `dialects/`, reports under `reports/`, and final PIM artifacts under `pim/`.
+- `runner/` - generated reference runner source, build tree, and shared library.
+- `simulation/out.bin` - raw simulator output used for comparison.
 
-Release build (fast):
+The compiler currently dumps dialect snapshots such as `spatial0.mlir`,
+`spatial1_dcp_merged.mlir`, `pim0.mlir`, `pim1_buff.mlir`,
+`pim2_coalesced.mlir`, `pim3_folded.mlir`, and
+`pim4_materialized.mlir` when an output directory is available.
 
-```
+To rerun the simulator manually with tracing after validation has produced a
-cmake --build /home/nico/raptor/raptor/cmake-build-release --target onnx-mlir -j 30
+`raptor/pim/` directory:
+
+```bash
+cd backend-simulators/pim/pim-simulator
+cargo run --no-default-features --features tracing --release \
+  --package pim-simulator --bin pim-simulator -- \
+  -f /path/to/workspace/raptor/pim \
+  -o /path/to/workspace/simulation/out.bin \
+  -d <addr0>,<size0>,<addr1>,<size1>,...
 ```
 
-A slower debug build is also available — configure it the same way but with
+With `--features tracing`, the simulator writes per-core traces as
-`-DCMAKE_BUILD_TYPE=Debug` (see installation instructions below).
+`TraceCore0`, `TraceCore1`, ... next to `out.bin`. The validator normally
+computes the `-d` ranges from `raptor/pim/config.json` and model output shapes.
+
+Available validation networks under `validation/networks/`: `vgg16`,
+`yolo11n`, `yolo11nv2`.
+
+Available operation suites under `validation/operations/`: `add`, `concat`,
+`conv`, `div`, `gather`, `gemm`, `gemv`, `matmul`, `mul`, `pool`,
+`reduce_mean`, `relu`, `reshape`, `resize`, `sigmoid`, `softmax`, `split`.
+
+Generated operation tests can be regenerated with:
+
+```bash
+python3 validation/operations/gen_tests.py
+```
 
 ## Build
 
+Initialize submodules first:
+
+```bash
+git submodule update --init --recursive
+```
+
+The project follows ONNX-MLIR's build requirements. The CI workflow documents
+the currently used versions and setup:
+- CMake 4.3.0 in CI,
+- LLVM/MLIR checked out under `onnx-mlir/llvm-project`,
+- Protobuf `v34.0`,
+- Rust stable for `pim-simulator`,
+- Python packages `numpy`, `onnx`, `colorama` for validation.
+
 ### Protobuf
 
-Use the following commands to install protobuf:
+Install Protobuf if your system does not already provide a compatible version:
-```
+
+```bash
 git clone --depth 1 --branch v34.0 https://github.com/protocolbuffers/protobuf
-cd protobuf
+cmake -S protobuf -B protobuf/build -G Ninja \
-mkdir build
+  -DCMAKE_BUILD_TYPE=Release \
-cd build
+  -Dprotobuf_BUILD_TESTS=OFF
-cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release
+cmake --build protobuf/build
-ninja
+sudo cmake --install protobuf/build
-sudo ninja install
 ```
 
-You can now remove the protobuf repo directory with:
+You can then remove the temporary checkout:
-```
+
-cd ../..
+```bash
 rm -rf protobuf
 ```
 
-### Mlir
+### MLIR
 
-Follow the first part of instructions [here](onnx-mlir/docs/BuildOnLinuxOSX.md) to build mlir.
+Follow the ONNX-MLIR instructions in
+`onnx-mlir/docs/BuildOnLinuxOSX.md` to build LLVM/MLIR. The local Raptor build
+expects `MLIR_DIR` to point at the MLIR CMake package, for example:
 
-Remember to set ```-DCMAKE_BUILD_TYPE=Debug``` for developing on Raptor
+```bash
-
+MLIR_DIR=$(pwd)/onnx-mlir/llvm-project/build_release/lib/cmake/mlir
-Moreover, if compiling with build type debug, it is also suggested to use
-mold as linker (you will need to install it if you don't have it already)
-to reduce memory usage during linking. You can use it by setting the options:
-```
--DLLVM_USE_LINKER=mold
 ```
 
+If your LLVM build directory is named `build` instead of `build_release`, adjust
+the path accordingly.
+
 ### Raptor
 
-Use the following commands to build Raptor.
+Configure a release build:
 
-Remember to set ```-DCMAKE_BUILD_TYPE=Debug``` for developing on Raptor.
+```bash
-
+MLIR_DIR=$(pwd)/onnx-mlir/llvm-project/build_release/lib/cmake/mlir
-Also in this case, it is suggested to use mold as linker to reduce link time and memory usage,
+cmake -S . -B build_release -G Ninja \
-setting the options:
+  -DCMAKE_BUILD_TYPE=Release \
+  -DONNX_MLIR_ACCELERATORS=PIM \
+  -DLLVM_ENABLE_ASSERTIONS=ON \
+  -DMLIR_DIR=${MLIR_DIR}
 ```
+
+Configure a debug build similarly:
+
+```bash
+MLIR_DIR=$(pwd)/onnx-mlir/llvm-project/build_debug/lib/cmake/mlir
+cmake -S . -B build_debug -G Ninja \
+  -DCMAKE_BUILD_TYPE=Debug \
+  -DONNX_MLIR_ACCELERATORS=PIM \
+  -DLLVM_ENABLE_ASSERTIONS=ON \
+  -DMLIR_DIR=${MLIR_DIR}
+```
+
+For debug development, using `mold` can reduce link time and memory use:
+
+```bash
+cmake -S . -B build_debug -G Ninja \
+  -DCMAKE_BUILD_TYPE=Debug \
+  -DONNX_MLIR_ACCELERATORS=PIM \
+  -DLLVM_ENABLE_ASSERTIONS=ON \
+  -DMLIR_DIR=${MLIR_DIR} \
   -DCMAKE_EXE_LINKER_FLAGS="-fuse-ld=mold" \
   -DCMAKE_SHARED_LINKER_FLAGS="-fuse-ld=mold" \
   -DCMAKE_MODULE_LINKER_FLAGS="-fuse-ld=mold"
 ```
 
-```
+Build the compiler with CMake:
-git submodule update --init --recursive
 
-MLIR_DIR=$(pwd)/onnx-mlir/llvm-project/build/lib/cmake/mlir
+```bash
-mkdir build && cd build
+cmake --build ./build_release
-cmake .. -G Ninja \
+cmake --build ./build_debug
-  -DCMAKE_BUILD_TYPE=Release \
-  -DONNX_MLIR_ACCELERATORS=PIM \
-  -DLLVM_ENABLE_ASSERTIONS=ON \
-  -DMLIR_DIR=${MLIR_DIR}
-cmake --build .
 ```
 
-If the build fails because of protobuf missing uint definitions,
+Do not invoke `ninja` directly for this project; use `cmake --build` so CMake's
-just patch the problematic files by adding ```#include <cstdint>``` to their includes.
+configuration and generated shims stay consistent.
+
+If a build fails because Protobuf headers are missing fixed-width integer
+definitions, patch the affected Protobuf-generated files by adding
+`#include <cstdint>`.
+
+## Tests
+
+The Rust simulator has its own tests:
+
+```bash
+cd backend-simulators/pim/pim-simulator
+cargo test
+```
+
+## Repository Layout
+
+- `src/PIM/` - PIM accelerator implementation.
+- `test/PIM/` - PIM C++ unit tests.
+- `validation/` - functional validation scripts, ONNX operation tests, network
+  slices, and pimsim config generation.
+- `backend-simulators/pim/pim-simulator/` - in-tree Rust functional simulator.
+- `backend-simulators/pim/pimsim-nn/` - performance simulator submodule.
+- `pimcomp_utils/` - local comparison helpers for PIMCOMP-NN.
+- `.github/actions/` and `.github/workflows/validate_operations.yml` - CI setup
+  for MLIR/Protobuf caching, building Raptor, and validation.

backend-simulators/pim/pim-simulator/src/bin/pim-simulator/main.rs

@@ -43,7 +43,7 @@ struct Args {
 
     /// Comma separated list of (address,size) for memory output dump
     #[arg(short, long, value_delimiter = ',', num_args = 1.., value_name = "ADDR,SIZE")]
-    dump: Vec<i32>,
+    dump: Vec<usize>,
 }
 
 fn main() -> Result<()> {
@@ -168,7 +168,7 @@ fn get_crossbars(config: &Value, args: &Args) -> anyhow::Result<HashMap<String,
 }
 
 fn dump_memory(mut executor: pimcore::Executable, args: &Args) -> Result<()> {
-    let dumps: Vec<(i32, i32)> = args
+    let dumps: Vec<(usize, usize)> = args
         .dump
         .chunks_exact(2)
         .map(|chunk| (chunk[0], chunk[1]))

backend-simulators/pim/pim-simulator/src/lib/cpu/mod.rs

@@ -1,3 +1,4 @@
+use crate::utility::AddressArg;
 use std::{collections::HashMap, fmt::Debug};
 use anyhow::{Context, Result, ensure};
 
@@ -9,6 +10,7 @@ use crate::{
 
 pub mod crossbar;
 
+
 #[derive(Debug, Clone)]
 pub struct CPU<'a> {
     cores: Box<[Core<'a>]>,
@@ -91,30 +93,26 @@ impl<'a> Core<'a> {
         self.memory.execute_load()
     }
 
-    pub fn execute_store<T>(&mut self, address: impl TryToUsize, element: &[T]) -> Result<()>
+    pub fn execute_store<T>(&mut self, address: impl AddressArg, element: &[T]) -> Result<()>
     where
         T: MemoryStorable,
     {
-        let address = address.try_into().context("address can not be negative")?;
+        let address = address.to_address_usize()?;
         self.memory.execute_store(address, element)
     }
 
     pub fn reserve_load(
         &mut self,
-        address: impl TryToUsize,
+        address: impl AddressArg,
         size: impl TryToUsize,
     ) -> Result<&mut CoreMemory> {
-        let address = address.try_into().context("address can not be negative")?;
+        let address = address.to_address_usize()?;
         let size = size.try_into().context("size can not be negative")?;
         self.memory.reserve_load(address, size)
     }
 
     pub fn set_register(&mut self, index: impl TryToUsize, value: i32) {
         let index = index.try_into().expect("index can not be negative");
-        assert!(
-            value >= 0,
-            "Register cannot be negative if happens remove this and go check where it's used as usize"
-        );
         self.registers[index] = value;
     }
 
@@ -123,11 +121,11 @@ impl<'a> Core<'a> {
         self.registers[index]
     }
 
-    pub fn load<T>(&mut self, address: impl TryToUsize, size: impl TryToUsize) -> Result<Vec<&[T]>>
+    pub fn load<T>(&mut self, address: impl AddressArg, size: impl TryToUsize) -> Result<Vec<&[T]>>
     where
         T: MemoryStorable,
     {
-        let address = address.try_into().context("address can not be negative")?;
+        let address = address.to_address_usize()?;
         let size = size.try_into().context("size can not be negative")?;
         self.memory.load(address, size)
     }
@@ -141,8 +139,8 @@ impl<'a> Core<'a> {
         (memory, crossbars)
     }
 
-    pub fn memset(&mut self, address: impl TryToUsize, size: impl TryToUsize, val: u8) -> Result<()> {
+    pub fn memset(&mut self, address: impl AddressArg, size: impl TryToUsize, val: u8) -> Result<()> {
-        let address = address.try_into().context("address can not be negative")?;
+        let address = address.to_address_usize()?;
         let size = size.try_into().context("size can not be negative")?;
         self.memory.memset(address, size, val)
     }

backend-simulators/pim/pim-simulator/src/lib/pimcore.rs

@@ -299,10 +299,11 @@ fn detect_deadlock(cores_instructions: &[CoreInstructions]) -> Option<DeadlockIn
             if in_path.contains(&waiting_for) {
                 let cycle_start = path.iter().position(|&c| c == waiting_for).unwrap();
                 let cycle = &path[cycle_start..];
+                let format_core = |core: &i32| (core - 1).to_string();
 
                 let cycle_str = cycle
                     .iter()
-                    .map(|c| c.to_string())
+                    .map(format_core)
                     .collect::<Vec<_>>()
                     .join(" -> ");
 
@@ -311,19 +312,19 @@ fn detect_deadlock(cores_instructions: &[CoreInstructions]) -> Option<DeadlockIn
                     .copied()
                     .chain(std::iter::once(waiting_for))
                     .collect::<Vec<_>>();
-                let cycle_msg = format!("{} -> {}", cycle_str, waiting_for);
+                let cycle_msg = format!("{} -> {}", cycle_str, waiting_for - 1);
                 let states_msg = cycle
                     .iter()
                     .filter_map(|core| {
                         states.get(core).map(|state| match state {
                             CoreState::SendingTo(target, size) => {
-                                format!("core {} send {}B -> {}", core, size, target)
+                                format!("core {} send {}B -> {}", core - 1, size, target - 1)
                             }
                             CoreState::ReceivingFrom(source, size) => {
-                                format!("core {} recv {}B <- {}", core, size, source)
+                                format!("core {} recv {}B <- {}", core - 1, size, source - 1)
                             }
-                            CoreState::Working => format!("core {} working", core),
+                            CoreState::Working => format!("core {} working", core - 1),
-                            CoreState::Halted => format!("core {} halted", core),
+                            CoreState::Halted => format!("core {} halted", core - 1),
                         })
                     })
                     .collect::<Vec<_>>()

backend-simulators/pim/pim-simulator/src/lib/utility.rs

@@ -1,7 +1,45 @@
+use anyhow::{Result,Context};
 use std::{fmt::Debug, mem::transmute};
 
-use crate::memory_manager::type_traits::TryToUsize;
 
+pub trait AddressArg {
+    fn to_address_usize(self) -> Result<usize>;
+}
+
+impl AddressArg for usize {
+    fn to_address_usize(self) -> Result<usize> {
+        Ok(self)
+    }
+}
+
+impl AddressArg for u32 {
+    fn to_address_usize(self) -> Result<usize> {
+        Ok(self as usize)
+    }
+}
+
+impl AddressArg for u64 {
+    fn to_address_usize(self) -> Result<usize> {
+        usize::try_from(self).context("address does not fit in usize")
+    }
+}
+
+impl AddressArg for i32 {
+    fn to_address_usize(self) -> Result<usize> {
+        Ok(self as u32 as usize)
+    }
+}
+
+impl AddressArg for i64 {
+    fn to_address_usize(self) -> Result<usize> {
+        usize::try_from(self).context("address can not be negative")
+    }
+}
+
+
+fn address_to_usize(address: i32) -> usize {
+    address as u32 as usize
+}
 
 fn add_offset_impl(address: usize, offset_select : i32, offset_value : i32, id:i32) -> usize{
      assert!(offset_select == 1 || offset_select == 2 || offset_select == 4 || offset_value == 0, "offset_select not a bit field");
@@ -14,21 +52,21 @@ fn add_offset_impl(address: usize, offset_select : i32, offset_value : i32, id:i
 }
 
 
-pub fn add_offset_rd(address: impl TryToUsize, offset_select : i32, offset_value : i32) -> usize
+pub fn add_offset_rd(address: i32, offset_select : i32, offset_value : i32) -> usize
 {
-    let address = address.try_into().expect("address can not be negative");
+    let address = address_to_usize(address);
     add_offset_impl(address, offset_select, offset_value, 4)
 }
 
-pub fn add_offset_r1(address: impl TryToUsize, offset_select : i32, offset_value : i32) -> usize
+pub fn add_offset_r1(address: i32, offset_select : i32, offset_value : i32) -> usize
 {
-    let address = address.try_into().expect("address can not be negative");
+    let address = address_to_usize(address);
     add_offset_impl(address, offset_select, offset_value, 1)
 }
 
-pub fn add_offset_r2(address: impl TryToUsize, offset_select : i32, offset_value : i32) -> usize
+pub fn add_offset_r2(address: i32, offset_select : i32, offset_value : i32) -> usize
 {
-    let address = address.try_into().expect("address can not be negative");
+    let address = address_to_usize(address);
     add_offset_impl(address, offset_select, offset_value, 2)
 }
 

src/PIM/CMakeLists.txt

@@ -10,6 +10,56 @@ set(PIM_INCLUDE_PATH ${CMAKE_INCLUDE_OUTPUT_DIRECTORY})
 set(PIM_ONNX_MLIR_SRC_ROOT ${ONNX_MLIR_SRC_ROOT})
 set(PIM_ONNX_MLIR_BIN_ROOT ${ONNX_MLIR_BIN_ROOT})
 
+set(PIM_GENERATED_PATH_SHIM_TARGET "")
+get_filename_component(PIM_BIN_ROOT_NAME "${PIM_BIN_ROOT}" NAME)
+if (PIM_BIN_ROOT_NAME STREQUAL "raptor-external")
+  get_filename_component(PIM_GENERATED_PATH_SHIM_ROOT "${PIM_BIN_ROOT}" DIRECTORY)
+  set(PIM_GENERATED_PATH_SHIM_OUTPUTS)
+
+  function(add_pim_generated_path_shim relative_path)
+    set(real_file "${PIM_BIN_ROOT}/${relative_path}")
+    set(shim_file "${PIM_GENERATED_PATH_SHIM_ROOT}/${relative_path}")
+    get_filename_component(shim_dir "${shim_file}" DIRECTORY)
+
+    add_custom_command(
+      OUTPUT "${shim_file}"
+      DEPENDS "${real_file}"
+      COMMAND "${CMAKE_COMMAND}" -E make_directory "${shim_dir}"
+      COMMAND "${CMAKE_COMMAND}" -E rm -f "${shim_file}"
+      COMMAND "${CMAKE_COMMAND}" -E create_symlink "${real_file}" "${shim_file}"
+      VERBATIM
+    )
+
+    list(APPEND PIM_GENERATED_PATH_SHIM_OUTPUTS "${shim_file}")
+    set(PIM_GENERATED_PATH_SHIM_OUTPUTS "${PIM_GENERATED_PATH_SHIM_OUTPUTS}" PARENT_SCOPE)
+  endfunction()
+
+  file(GLOB_RECURSE pim_generated_path_scan_sources
+    CONFIGURE_DEPENDS
+    "${PIM_SRC_ROOT}/*.cpp"
+    "${PIM_SRC_ROOT}/*.hpp"
+  )
+
+  set(pim_generated_path_shims)
+  foreach (source_file IN LISTS pim_generated_path_scan_sources)
+    file(READ "${source_file}" source_contents)
+    string(REGEX MATCHALL "#include \"src/Accelerators/PIM/[^\"]+\\.inc\"" source_inc_matches "${source_contents}")
+
+    foreach (inc_match IN LISTS source_inc_matches)
+      string(REGEX REPLACE "^#include \"src/Accelerators/PIM/(.+)\"$" "\\1" relative_inc_path "${inc_match}")
+      list(APPEND pim_generated_path_shims "${relative_inc_path}")
+    endforeach ()
+  endforeach ()
+
+  list(REMOVE_DUPLICATES pim_generated_path_shims)
+  foreach (relative_inc_path IN LISTS pim_generated_path_shims)
+    add_pim_generated_path_shim("${relative_inc_path}")
+  endforeach ()
+
+  add_custom_target(OMPimGeneratedPathShims DEPENDS ${PIM_GENERATED_PATH_SHIM_OUTPUTS})
+  set(PIM_GENERATED_PATH_SHIM_TARGET OMPimGeneratedPathShims)
+endif ()
+
 set(PIM_PUBLIC_INCLUDE_DIRS
   ${ONNX_MLIR_SRC_ROOT}/include
   ${ONNX_MLIR_SRC_ROOT}
@@ -37,6 +87,9 @@ set(PIM_GENERATED_INCLUDE_DIRS
 
 function(add_pim_library name)
   add_onnx_mlir_library(${name} STATIC ${ARGN})
+  if (PIM_GENERATED_PATH_SHIM_TARGET)
+    add_dependencies(${name} ${PIM_GENERATED_PATH_SHIM_TARGET})
+  endif ()
 endfunction()
 
 add_subdirectory(Dialect)

src/PIM/Common/CMakeLists.txt

@@ -1,5 +1,7 @@
 add_pim_library(OMPimCommon
   IR/AddressAnalysis.cpp
+  IR/BatchCoreUtils.cpp
+  IR/ConstantUtils.cpp
   IR/CoreBlockUtils.cpp
   IR/EntryPointUtils.cpp
   IR/ShapeUtils.cpp
@@ -16,6 +18,7 @@ add_pim_library(OMPimCommon
   ${PIM_PUBLIC_INCLUDE_DIRS}
 
   LINK_LIBS PUBLIC
+  MLIRLinalgDialect
   onnx
   SpatialOps
   PimOps

src/PIM/Common/IR/AddressAnalysis.cpp

@@ -1,7 +1,11 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
+#include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/Interfaces/DestinationStyleOpInterface.h"
 
+#include <limits>
+
 #include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
 #include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
@@ -28,6 +32,14 @@ mlir::Value resolveAlias(mlir::Value value, const StaticValueKnowledge* knowledg
   return value;
 }
 
+llvm::FailureOr<CompiledIndexExpr> compileIndexValueImpl(mlir::Value value);
+llvm::FailureOr<CompiledAddressExpr> compileContiguousAddressExprImpl(mlir::Value value);
+
+template <typename... Args>
+CompiledIndexExpr makeCompiledIndexExpr(Args&&... args) {
+  return CompiledIndexExpr(std::make_shared<CompiledIndexExprNode>(std::forward<Args>(args)...));
+}
+
 mlir::Value resolveLoopCarriedAliasImpl(mlir::Value value, const StaticValueKnowledge* knowledge) {
   value = resolveAlias(value, knowledge);
 
@@ -55,6 +67,278 @@ mlir::Value resolveLoopCarriedAliasImpl(mlir::Value value, const StaticValueKnow
 }
 
 llvm::FailureOr<int64_t> resolveOpFoldResult(mlir::OpFoldResult ofr, const StaticValueKnowledge* knowledge);
+llvm::FailureOr<int64_t> resolveIndexValueImpl(mlir::Value value, const StaticValueKnowledge* knowledge);
+
+static llvm::FailureOr<int64_t> resolveConstantGlobalLoad(mlir::memref::LoadOp loadOp,
+                                                          const StaticValueKnowledge* knowledge) {
+  auto getGlobalOp = loadOp.getMemRef().getDefiningOp<mlir::memref::GetGlobalOp>();
+  if (!getGlobalOp)
+    return mlir::failure();
+
+  auto moduleOp = loadOp->getParentOfType<mlir::ModuleOp>();
+  auto globalOp = lookupGlobalForGetGlobal(moduleOp, getGlobalOp);
+  if (!globalOp || !globalOp.getConstant() || !globalOp.getInitialValue())
+    return mlir::failure();
+
+  auto denseAttr = mlir::dyn_cast<mlir::DenseElementsAttr>(*globalOp.getInitialValue());
+  auto globalType = mlir::dyn_cast<mlir::MemRefType>(getGlobalOp.getType());
+  if (!denseAttr || !globalType || !globalType.hasStaticShape())
+    return mlir::failure();
+
+  auto elementType = denseAttr.getElementType();
+  if (!elementType.isIndex() && !elementType.isInteger())
+    return mlir::failure();
+
+  llvm::SmallVector<int64_t> indices;
+  indices.reserve(loadOp.getIndices().size());
+  for (mlir::Value index : loadOp.getIndices()) {
+    auto resolvedIndex = resolveIndexValueImpl(index, knowledge);
+    if (failed(resolvedIndex))
+      return mlir::failure();
+    indices.push_back(*resolvedIndex);
+  }
+
+  if (indices.size() != static_cast<size_t>(globalType.getRank()))
+    return mlir::failure();
+
+  auto strides = computeRowMajorStrides(globalType.getShape());
+  int64_t linearIndex = linearizeIndex(indices, strides);
+  if (linearIndex < 0 || linearIndex >= globalType.getNumElements())
+    return mlir::failure();
+
+  return denseAttr.getValues<llvm::APInt>()[linearIndex].getSExtValue();
+}
+
+static bool evaluateCmpPredicate(mlir::arith::CmpIPredicate predicate, int64_t lhs, int64_t rhs) {
+  switch (predicate) {
+  case mlir::arith::CmpIPredicate::eq:  return lhs == rhs;
+  case mlir::arith::CmpIPredicate::ne:  return lhs != rhs;
+  case mlir::arith::CmpIPredicate::slt: return lhs < rhs;
+  case mlir::arith::CmpIPredicate::sle: return lhs <= rhs;
+  case mlir::arith::CmpIPredicate::sgt: return lhs > rhs;
+  case mlir::arith::CmpIPredicate::sge: return lhs >= rhs;
+  case mlir::arith::CmpIPredicate::ult: return static_cast<uint64_t>(lhs) < static_cast<uint64_t>(rhs);
+  case mlir::arith::CmpIPredicate::ule: return static_cast<uint64_t>(lhs) <= static_cast<uint64_t>(rhs);
+  case mlir::arith::CmpIPredicate::ugt: return static_cast<uint64_t>(lhs) > static_cast<uint64_t>(rhs);
+  case mlir::arith::CmpIPredicate::uge: return static_cast<uint64_t>(lhs) >= static_cast<uint64_t>(rhs);
+  }
+
+  llvm_unreachable("unknown cmpi predicate");
+}
+
+llvm::FailureOr<int64_t> evaluateCompiledIndexExpr(const CompiledIndexExpr& expr,
+                                                   const StaticValueKnowledge& knowledge) {
+  if (!expr.node)
+    return mlir::failure();
+
+  switch (expr.node->kind) {
+  case CompiledIndexExprNode::Kind::Constant: return expr.node->constant;
+  case CompiledIndexExprNode::Kind::Symbol:   {
+    auto value = resolveAlias(expr.node->symbol, &knowledge);
+    auto iter = knowledge.indexValues.find(value);
+    if (iter != knowledge.indexValues.end())
+      return iter->second;
+    return mlir::failure();
+  }
+  case CompiledIndexExprNode::Kind::Add:
+  case CompiledIndexExprNode::Kind::Sub:
+  case CompiledIndexExprNode::Kind::Mul:
+  case CompiledIndexExprNode::Kind::DivUI:
+  case CompiledIndexExprNode::Kind::DivSI:
+  case CompiledIndexExprNode::Kind::RemUI:
+  case CompiledIndexExprNode::Kind::RemSI:
+  case CompiledIndexExprNode::Kind::MinUI:
+  case CompiledIndexExprNode::Kind::CmpI:  {
+    auto lhs = evaluateCompiledIndexExpr(expr.node->operands[0], knowledge);
+    auto rhs = evaluateCompiledIndexExpr(expr.node->operands[1], knowledge);
+    if (failed(lhs) || failed(rhs))
+      return mlir::failure();
+
+    switch (expr.node->kind) {
+    case CompiledIndexExprNode::Kind::Add: return *lhs + *rhs;
+    case CompiledIndexExprNode::Kind::Sub: return *lhs - *rhs;
+    case CompiledIndexExprNode::Kind::Mul: return *lhs * *rhs;
+    case CompiledIndexExprNode::Kind::DivUI:
+      if (*rhs == 0)
+        return mlir::failure();
+      return static_cast<int64_t>(static_cast<uint64_t>(*lhs) / static_cast<uint64_t>(*rhs));
+    case CompiledIndexExprNode::Kind::DivSI:
+      if (*rhs == 0 || (*lhs == std::numeric_limits<int64_t>::min() && *rhs == -1))
+        return mlir::failure();
+      return *lhs / *rhs;
+    case CompiledIndexExprNode::Kind::RemUI:
+      if (*rhs == 0)
+        return mlir::failure();
+      return static_cast<int64_t>(static_cast<uint64_t>(*lhs) % static_cast<uint64_t>(*rhs));
+    case CompiledIndexExprNode::Kind::RemSI:
+      if (*rhs == 0)
+        return mlir::failure();
+      if (*lhs == std::numeric_limits<int64_t>::min() && *rhs == -1)
+        return 0;
+      return *lhs % *rhs;
+    case CompiledIndexExprNode::Kind::MinUI:
+      return static_cast<int64_t>(std::min(static_cast<uint64_t>(*lhs), static_cast<uint64_t>(*rhs)));
+    case CompiledIndexExprNode::Kind::CmpI: return evaluateCmpPredicate(expr.node->predicate, *lhs, *rhs) ? 1 : 0;
+    default:                                llvm_unreachable("unexpected binary compiled index kind");
+    }
+  }
+  case CompiledIndexExprNode::Kind::Select: {
+    auto condition = evaluateCompiledIndexExpr(expr.node->operands[0], knowledge);
+    if (failed(condition))
+      return mlir::failure();
+    return evaluateCompiledIndexExpr(*condition != 0 ? expr.node->operands[1] : expr.node->operands[2], knowledge);
+  }
+  case CompiledIndexExprNode::Kind::ConstantGlobalLoad: {
+    if (!expr.node->globalOp || !expr.node->globalOp.getInitialValue())
+      return mlir::failure();
+
+    auto denseAttr = mlir::dyn_cast<mlir::DenseElementsAttr>(*expr.node->globalOp.getInitialValue());
+    auto globalType = mlir::dyn_cast<mlir::MemRefType>(expr.node->globalOp.getType());
+    if (!denseAttr || !globalType)
+      return mlir::failure();
+
+    llvm::SmallVector<int64_t> indices;
+    indices.reserve(expr.node->operands.size());
+    for (const CompiledIndexExpr& operand : expr.node->operands) {
+      auto resolvedIndex = evaluateCompiledIndexExpr(operand, knowledge);
+      if (failed(resolvedIndex))
+        return mlir::failure();
+      indices.push_back(*resolvedIndex);
+    }
+
+    int64_t linearIndex = linearizeIndex(indices, expr.node->globalStrides);
+    if (linearIndex < 0 || linearIndex >= globalType.getNumElements())
+      return mlir::failure();
+    return denseAttr.getValues<llvm::APInt>()[linearIndex].getSExtValue();
+  }
+  }
+
+  llvm_unreachable("unknown compiled index kind");
+}
+
+llvm::FailureOr<CompiledIndexExpr> compileConstantGlobalLoad(mlir::memref::LoadOp loadOp) {
+  auto getGlobalOp = loadOp.getMemRef().getDefiningOp<mlir::memref::GetGlobalOp>();
+  if (!getGlobalOp)
+    return mlir::failure();
+
+  auto moduleOp = loadOp->getParentOfType<mlir::ModuleOp>();
+  auto globalOp = lookupGlobalForGetGlobal(moduleOp, getGlobalOp);
+  if (!globalOp || !globalOp.getConstant() || !globalOp.getInitialValue())
+    return mlir::failure();
+
+  auto denseAttr = mlir::dyn_cast<mlir::DenseElementsAttr>(*globalOp.getInitialValue());
+  auto globalType = mlir::dyn_cast<mlir::MemRefType>(getGlobalOp.getType());
+  if (!denseAttr || !globalType || !globalType.hasStaticShape())
+    return mlir::failure();
+
+  auto elementType = denseAttr.getElementType();
+  if (!elementType.isIndex() && !elementType.isInteger())
+    return mlir::failure();
+
+  CompiledIndexExprNode expr;
+  expr.kind = CompiledIndexExprNode::Kind::ConstantGlobalLoad;
+  expr.globalOp = globalOp;
+  expr.globalStrides = computeRowMajorStrides(globalType.getShape());
+  expr.operands.reserve(loadOp.getIndices().size());
+  for (mlir::Value index : loadOp.getIndices()) {
+    auto compiledIndex = compileIndexValueImpl(index);
+    if (failed(compiledIndex))
+      return mlir::failure();
+    expr.operands.push_back(*compiledIndex);
+  }
+  return makeCompiledIndexExpr(std::move(expr));
+}
+
+llvm::FailureOr<CompiledIndexExpr> compileIndexValueImpl(mlir::Value value) {
+  if (auto constantOp = value.getDefiningOp<mlir::arith::ConstantOp>()) {
+    if (auto integerAttr = mlir::dyn_cast<mlir::IntegerAttr>(constantOp.getValue())) {
+      CompiledIndexExprNode expr;
+      expr.kind = CompiledIndexExprNode::Kind::Constant;
+      expr.constant = integerAttr.getInt();
+      return makeCompiledIndexExpr(std::move(expr));
+    }
+  }
+
+  mlir::Operation* definingOp = value.getDefiningOp();
+  if (!definingOp) {
+    CompiledIndexExprNode expr;
+    expr.kind = CompiledIndexExprNode::Kind::Symbol;
+    expr.symbol = value;
+    return makeCompiledIndexExpr(std::move(expr));
+  }
+
+  auto buildBinaryExpr = [&](CompiledIndexExprNode::Kind kind, mlir::Value lhsValue, mlir::Value rhsValue) {
+    auto lhs = compileIndexValueImpl(lhsValue);
+    auto rhs = compileIndexValueImpl(rhsValue);
+    if (failed(lhs) || failed(rhs))
+      return llvm::FailureOr<CompiledIndexExpr>(mlir::failure());
+    CompiledIndexExprNode expr;
+    expr.kind = kind;
+    expr.operands = {*lhs, *rhs};
+    return llvm::FailureOr<CompiledIndexExpr>(makeCompiledIndexExpr(std::move(expr)));
+  };
+
+  if (auto indexCastOp = mlir::dyn_cast<mlir::arith::IndexCastOp>(definingOp))
+    return compileIndexValueImpl(indexCastOp.getIn());
+  if (auto addOp = mlir::dyn_cast<mlir::arith::AddIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::Add, addOp.getLhs(), addOp.getRhs());
+  if (auto subOp = mlir::dyn_cast<mlir::arith::SubIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::Sub, subOp.getLhs(), subOp.getRhs());
+  if (auto mulOp = mlir::dyn_cast<mlir::arith::MulIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::Mul, mulOp.getLhs(), mulOp.getRhs());
+  if (auto divOp = mlir::dyn_cast<mlir::arith::DivUIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::DivUI, divOp.getLhs(), divOp.getRhs());
+  if (auto divOp = mlir::dyn_cast<mlir::arith::DivSIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::DivSI, divOp.getLhs(), divOp.getRhs());
+  if (auto minOp = mlir::dyn_cast<mlir::arith::MinUIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::MinUI, minOp.getLhs(), minOp.getRhs());
+  if (auto remOp = mlir::dyn_cast<mlir::arith::RemUIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::RemUI, remOp.getLhs(), remOp.getRhs());
+  if (auto remOp = mlir::dyn_cast<mlir::arith::RemSIOp>(definingOp))
+    return buildBinaryExpr(CompiledIndexExprNode::Kind::RemSI, remOp.getLhs(), remOp.getRhs());
+  if (auto cmpOp = mlir::dyn_cast<mlir::arith::CmpIOp>(definingOp)) {
+    auto expr = buildBinaryExpr(CompiledIndexExprNode::Kind::CmpI, cmpOp.getLhs(), cmpOp.getRhs());
+    if (failed(expr))
+      return mlir::failure();
+    auto exprNode = std::make_shared<CompiledIndexExprNode>(*expr->node);
+    exprNode->predicate = cmpOp.getPredicate();
+    return CompiledIndexExpr(exprNode);
+  }
+  if (auto maxOp = mlir::dyn_cast<mlir::arith::MaxUIOp>(definingOp)) {
+    auto lhs = compileIndexValueImpl(maxOp.getLhs());
+    auto rhs = compileIndexValueImpl(maxOp.getRhs());
+    if (failed(lhs) || failed(rhs))
+      return mlir::failure();
+
+    CompiledIndexExprNode cmpExpr;
+    cmpExpr.kind = CompiledIndexExprNode::Kind::CmpI;
+    cmpExpr.predicate = mlir::arith::CmpIPredicate::uge;
+    cmpExpr.operands = {*lhs, *rhs};
+
+    CompiledIndexExprNode selectExpr;
+    selectExpr.kind = CompiledIndexExprNode::Kind::Select;
+    selectExpr.operands = {makeCompiledIndexExpr(std::move(cmpExpr)), *lhs, *rhs};
+    return makeCompiledIndexExpr(std::move(selectExpr));
+  }
+  if (auto selectOp = mlir::dyn_cast<mlir::arith::SelectOp>(definingOp)) {
+    auto condition = compileIndexValueImpl(selectOp.getCondition());
+    auto trueValue = compileIndexValueImpl(selectOp.getTrueValue());
+    auto falseValue = compileIndexValueImpl(selectOp.getFalseValue());
+    if (failed(condition) || failed(trueValue) || failed(falseValue))
+      return mlir::failure();
+    CompiledIndexExprNode expr;
+    expr.kind = CompiledIndexExprNode::Kind::Select;
+    expr.operands = {*condition, *trueValue, *falseValue};
+    return makeCompiledIndexExpr(std::move(expr));
+  }
+  if (auto loadOp = mlir::dyn_cast<mlir::memref::LoadOp>(definingOp))
+    return compileConstantGlobalLoad(loadOp);
+
+  CompiledIndexExprNode expr;
+  expr.kind = CompiledIndexExprNode::Kind::Symbol;
+  expr.symbol = value;
+  return makeCompiledIndexExpr(std::move(expr));
+}
 
 llvm::FailureOr<int64_t> resolveIndexValueImpl(mlir::Value value, const StaticValueKnowledge* knowledge) {
   value = resolveAlias(value, knowledge);
@@ -110,6 +394,16 @@ llvm::FailureOr<int64_t> resolveIndexValueImpl(mlir::Value value, const StaticVa
     return static_cast<int64_t>(static_cast<uint64_t>(*lhs) / static_cast<uint64_t>(*rhs));
   }
 
+  if (auto divOp = mlir::dyn_cast<mlir::arith::DivSIOp>(definingOp)) {
+    auto lhs = resolveIndexValueImpl(divOp.getLhs(), knowledge);
+    auto rhs = resolveIndexValueImpl(divOp.getRhs(), knowledge);
+    if (failed(lhs) || failed(rhs) || *rhs == 0)
+      return mlir::failure();
+    if (*lhs == std::numeric_limits<int64_t>::min() && *rhs == -1)
+      return mlir::failure();
+    return *lhs / *rhs;
+  }
+
   if (auto minOp = mlir::dyn_cast<mlir::arith::MinUIOp>(definingOp)) {
     auto lhs = resolveIndexValueImpl(minOp.getLhs(), knowledge);
     auto rhs = resolveIndexValueImpl(minOp.getRhs(), knowledge);
@@ -126,6 +420,34 @@ llvm::FailureOr<int64_t> resolveIndexValueImpl(mlir::Value value, const StaticVa
     return static_cast<int64_t>(static_cast<uint64_t>(*lhs) % static_cast<uint64_t>(*rhs));
   }
 
+  if (auto remOp = mlir::dyn_cast<mlir::arith::RemSIOp>(definingOp)) {
+    auto lhs = resolveIndexValueImpl(remOp.getLhs(), knowledge);
+    auto rhs = resolveIndexValueImpl(remOp.getRhs(), knowledge);
+    if (failed(lhs) || failed(rhs) || *rhs == 0)
+      return mlir::failure();
+    if (*lhs == std::numeric_limits<int64_t>::min() && *rhs == -1)
+      return 0;
+    return *lhs % *rhs;
+  }
+
+  if (auto cmpOp = mlir::dyn_cast<mlir::arith::CmpIOp>(definingOp)) {
+    auto lhs = resolveIndexValueImpl(cmpOp.getLhs(), knowledge);
+    auto rhs = resolveIndexValueImpl(cmpOp.getRhs(), knowledge);
+    if (failed(lhs) || failed(rhs))
+      return mlir::failure();
+    return evaluateCmpPredicate(cmpOp.getPredicate(), *lhs, *rhs) ? 1 : 0;
+  }
+
+  if (auto selectOp = mlir::dyn_cast<mlir::arith::SelectOp>(definingOp)) {
+    auto condition = resolveIndexValueImpl(selectOp.getCondition(), knowledge);
+    if (failed(condition))
+      return mlir::failure();
+    return resolveIndexValueImpl(*condition != 0 ? selectOp.getTrueValue() : selectOp.getFalseValue(), knowledge);
+  }
+
+  if (auto loadOp = mlir::dyn_cast<mlir::memref::LoadOp>(definingOp))
+    return resolveConstantGlobalLoad(loadOp, knowledge);
+
   return mlir::failure();
 }
 
@@ -218,7 +540,7 @@ llvm::FailureOr<ResolvedContiguousAddress> resolveContiguousAddressImpl(mlir::Va
         return mlir::failure();
 
       auto sourceStrides = computeRowMajorStrides(sourceType.getShape());
-      byteOffset += linearizeIndex(offsets, sourceStrides) * subviewType.getElementTypeBitWidth() / 8;
+      byteOffset += linearizeIndex(offsets, sourceStrides) * getElementTypeSizeInBytes(subviewType.getElementType());
       value = resolveAlias(subviewOp.getSource(), knowledge);
       continue;
     }
@@ -243,6 +565,188 @@ llvm::FailureOr<ResolvedContiguousAddress> resolveContiguousAddressImpl(mlir::Va
   }
 }
 
+llvm::FailureOr<CompiledAddressExpr> compileContiguousAddressExprImpl(mlir::Value value) {
+  int64_t constantByteOffset = 0;
+  CompiledIndexExpr byteOffsetExpr;
+  {
+    CompiledIndexExprNode expr;
+    expr.kind = CompiledIndexExprNode::Kind::Constant;
+    expr.constant = 0;
+    byteOffsetExpr = makeCompiledIndexExpr(std::move(expr));
+  }
+
+  while (true) {
+    if (mlir::isa<mlir::BlockArgument>(value))
+      return CompiledAddressExpr {value, byteOffsetExpr};
+
+    mlir::Operation* definingOp = value.getDefiningOp();
+    if (!definingOp)
+      return mlir::failure();
+
+    if (auto dpsDefiningOp = mlir::dyn_cast<mlir::DestinationStyleOpInterface>(definingOp)) {
+      mlir::OpOperand* tiedOperand = dpsDefiningOp.getTiedOpOperand(mlir::dyn_cast<mlir::OpResult>(value));
+      if (!tiedOperand)
+        return mlir::failure();
+      value = tiedOperand->get();
+      continue;
+    }
+
+    if (auto forOp = mlir::dyn_cast<mlir::scf::ForOp>(definingOp)) {
+      auto result = mlir::dyn_cast<mlir::OpResult>(value);
+      if (!result)
+        return mlir::failure();
+
+      auto yieldOp = mlir::cast<mlir::scf::YieldOp>(forOp.getBody()->getTerminator());
+      mlir::Value yieldedValue = yieldOp.getOperand(result.getResultNumber());
+      if (auto blockArgument = mlir::dyn_cast<mlir::BlockArgument>(yieldedValue)) {
+        if (blockArgument.getOwner() == forOp.getBody() && blockArgument.getArgNumber() > 0
+            && static_cast<unsigned>(blockArgument.getArgNumber() - 1) < forOp.getInitArgs().size()) {
+          value = forOp.getInitArgs()[blockArgument.getArgNumber() - 1];
+          continue;
+        }
+      }
+
+      value = yieldedValue;
+      continue;
+    }
+
+    if (auto subviewOp = mlir::dyn_cast<mlir::memref::SubViewOp>(definingOp)) {
+      auto sourceType = mlir::dyn_cast<mlir::MemRefType>(subviewOp.getSource().getType());
+      auto subviewType = mlir::dyn_cast<mlir::MemRefType>(subviewOp.getType());
+      if (!sourceType || !subviewType || !sourceType.hasStaticShape() || !subviewType.hasStaticShape())
+        return mlir::failure();
+
+      llvm::SmallVector<int64_t> staticOffsets;
+      staticOffsets.reserve(subviewOp.getMixedOffsets().size());
+      llvm::SmallVector<int64_t> staticSizes;
+      staticSizes.reserve(subviewOp.getMixedSizes().size());
+      llvm::SmallVector<int64_t> staticStrides;
+      staticStrides.reserve(subviewOp.getMixedStrides().size());
+      bool allStatic = true;
+
+      for (mlir::OpFoldResult offset : subviewOp.getMixedOffsets())
+        if (auto attr = mlir::dyn_cast<mlir::Attribute>(offset))
+          staticOffsets.push_back(mlir::cast<mlir::IntegerAttr>(attr).getInt());
+        else
+          allStatic = false;
+      for (mlir::OpFoldResult size : subviewOp.getMixedSizes())
+        if (auto attr = mlir::dyn_cast<mlir::Attribute>(size))
+          staticSizes.push_back(mlir::cast<mlir::IntegerAttr>(attr).getInt());
+        else
+          allStatic = false;
+      for (mlir::OpFoldResult stride : subviewOp.getMixedStrides())
+        if (auto attr = mlir::dyn_cast<mlir::Attribute>(stride))
+          staticStrides.push_back(mlir::cast<mlir::IntegerAttr>(attr).getInt());
+        else
+          allStatic = false;
+
+      if (allStatic) {
+        if (!isMemoryContiguous(sourceType.getShape(), staticOffsets, staticSizes, staticStrides))
+          return mlir::failure();
+
+        auto sourceStrides = computeRowMajorStrides(sourceType.getShape());
+        constantByteOffset +=
+          linearizeIndex(staticOffsets, sourceStrides) * getElementTypeSizeInBytes(subviewType.getElementType());
+      }
+      else {
+        llvm::SmallVector<int64_t> sourceStrides = computeRowMajorStrides(sourceType.getShape());
+        CompiledIndexExpr offsetExpr;
+        {
+          CompiledIndexExprNode expr;
+          expr.kind = CompiledIndexExprNode::Kind::Constant;
+          expr.constant = 0;
+          offsetExpr = makeCompiledIndexExpr(std::move(expr));
+        }
+
+        for (auto [mixedOffset, sourceStride] : llvm::zip_equal(subviewOp.getMixedOffsets(), sourceStrides)) {
+          CompiledIndexExpr operandExpr;
+          if (auto attr = mlir::dyn_cast<mlir::Attribute>(mixedOffset)) {
+            CompiledIndexExprNode expr;
+            expr.kind = CompiledIndexExprNode::Kind::Constant;
+            expr.constant = mlir::cast<mlir::IntegerAttr>(attr).getInt() * sourceStride
+                          * getElementTypeSizeInBytes(subviewType.getElementType());
+            operandExpr = makeCompiledIndexExpr(std::move(expr));
+          }
+          else {
+            auto compiledOffset = compileIndexValueImpl(mlir::cast<mlir::Value>(mixedOffset));
+            if (failed(compiledOffset))
+              return mlir::failure();
+            CompiledIndexExpr scaleExpr;
+            {
+              CompiledIndexExprNode expr;
+              expr.kind = CompiledIndexExprNode::Kind::Constant;
+              expr.constant = sourceStride * getElementTypeSizeInBytes(subviewType.getElementType());
+              scaleExpr = makeCompiledIndexExpr(std::move(expr));
+            }
+
+            CompiledIndexExprNode expr;
+            expr.kind = CompiledIndexExprNode::Kind::Mul;
+            expr.operands = {*compiledOffset, scaleExpr};
+            operandExpr = makeCompiledIndexExpr(std::move(expr));
+          }
+
+          CompiledIndexExprNode expr;
+          expr.kind = CompiledIndexExprNode::Kind::Add;
+          expr.operands = {offsetExpr, operandExpr};
+          offsetExpr = makeCompiledIndexExpr(std::move(expr));
+        }
+
+        CompiledIndexExpr constantExpr;
+        {
+          CompiledIndexExprNode expr;
+          expr.kind = CompiledIndexExprNode::Kind::Constant;
+          expr.constant = constantByteOffset;
+          constantExpr = makeCompiledIndexExpr(std::move(expr));
+        }
+        CompiledIndexExprNode expr;
+        expr.kind = CompiledIndexExprNode::Kind::Add;
+        expr.operands = {constantExpr, offsetExpr};
+        byteOffsetExpr = makeCompiledIndexExpr(std::move(expr));
+        constantByteOffset = 0;
+      }
+
+      value = subviewOp.getSource();
+      continue;
+    }
+
+    if (auto castOp = mlir::dyn_cast<mlir::memref::CastOp>(definingOp)) {
+      value = castOp.getSource();
+      continue;
+    }
+    if (auto collapseOp = mlir::dyn_cast<mlir::memref::CollapseShapeOp>(definingOp)) {
+      value = collapseOp.getSrc();
+      continue;
+    }
+    if (auto expandOp = mlir::dyn_cast<mlir::memref::ExpandShapeOp>(definingOp)) {
+      value = expandOp.getSrc();
+      continue;
+    }
+
+    if (mlir::isa<mlir::memref::AllocOp, mlir::memref::GetGlobalOp>(definingOp)) {
+      if (constantByteOffset != 0) {
+        CompiledIndexExpr constantExpr;
+        {
+          CompiledIndexExprNode expr;
+          expr.kind = CompiledIndexExprNode::Kind::Constant;
+          expr.constant = constantByteOffset;
+          constantExpr = makeCompiledIndexExpr(std::move(expr));
+        }
+        if (byteOffsetExpr.node->kind == CompiledIndexExprNode::Kind::Constant && byteOffsetExpr.node->constant == 0)
+          byteOffsetExpr = constantExpr;
+        else {
+          CompiledIndexExprNode expr;
+          expr.kind = CompiledIndexExprNode::Kind::Add;
+          expr.operands = {constantExpr, byteOffsetExpr};
+          byteOffsetExpr = makeCompiledIndexExpr(std::move(expr));
+        }
+      }
+      return CompiledAddressExpr {value, byteOffsetExpr};
+    }
+
+    return mlir::failure();
+  }
+}
+
 } // namespace
 
 llvm::FailureOr<int64_t> resolveIndexValue(mlir::Value value) { return resolveIndexValueImpl(value, nullptr); }
@@ -251,6 +755,8 @@ llvm::FailureOr<int64_t> resolveIndexValue(mlir::Value value, const StaticValueK
   return resolveIndexValueImpl(value, &knowledge);
 }
 
+llvm::FailureOr<CompiledIndexExpr> compileIndexExpr(mlir::Value value) { return compileIndexValueImpl(value); }
+
 llvm::FailureOr<ResolvedContiguousAddress> resolveContiguousAddress(mlir::Value value) {
   return resolveContiguousAddressImpl(value, nullptr);
 }
@@ -264,4 +770,21 @@ mlir::Value resolveLoopCarriedAlias(mlir::Value value, const StaticValueKnowledg
   return resolveLoopCarriedAliasImpl(value, &knowledge);
 }
 
+llvm::FailureOr<CompiledAddressExpr> compileContiguousAddressExpr(mlir::Value value) {
+  return compileContiguousAddressExprImpl(value);
+}
+
+llvm::FailureOr<int64_t> CompiledIndexExpr::evaluate(const StaticValueKnowledge& knowledge) const {
+  return evaluateCompiledIndexExpr(*this, knowledge);
+}
+
+llvm::FailureOr<ResolvedContiguousAddress> CompiledAddressExpr::evaluate(const StaticValueKnowledge& knowledge,
+                                                                         std::optional<unsigned> lane) const {
+  (void) lane;
+  auto resolvedOffset = byteOffset.evaluate(knowledge);
+  if (failed(resolvedOffset))
+    return mlir::failure();
+  return ResolvedContiguousAddress {base, *resolvedOffset};
+}
+
 } // namespace onnx_mlir

src/PIM/Common/IR/AddressAnalysis.hpp

@@ -1,10 +1,14 @@
 #pragma once
 
+#include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/IR/Value.h"
 
 #include "llvm/ADT/DenseMap.h"
 
+#include <memory>
+#include <optional>
+
 namespace onnx_mlir {
 
 /// Describes a value as a base addressable object plus a statically known
@@ -23,6 +27,52 @@ struct StaticValueKnowledge {
   StaticValueKnowledge() {}
 };
 
+struct CompiledIndexExprNode;
+
+struct CompiledIndexExpr {
+  std::shared_ptr<CompiledIndexExprNode> node;
+
+  CompiledIndexExpr() = default;
+  explicit CompiledIndexExpr(std::shared_ptr<CompiledIndexExprNode> node)
+  : node(std::move(node)) {}
+
+  llvm::FailureOr<int64_t> evaluate(const StaticValueKnowledge& knowledge) const;
+};
+
+struct CompiledIndexExprNode {
+  enum class Kind {
+    Constant,
+    Symbol,
+    Add,
+    Sub,
+    Mul,
+    DivUI,
+    DivSI,
+    RemUI,
+    RemSI,
+    MinUI,
+    CmpI,
+    Select,
+    ConstantGlobalLoad
+  };
+
+  Kind kind = Kind::Constant;
+  int64_t constant = 0;
+  mlir::Value symbol;
+  mlir::arith::CmpIPredicate predicate = mlir::arith::CmpIPredicate::eq;
+  mlir::memref::GlobalOp globalOp;
+  llvm::SmallVector<int64_t, 4> globalStrides;
+  llvm::SmallVector<CompiledIndexExpr, 4> operands;
+};
+
+struct CompiledAddressExpr {
+  mlir::Value base;
+  CompiledIndexExpr byteOffset;
+
+  llvm::FailureOr<ResolvedContiguousAddress> evaluate(const StaticValueKnowledge& knowledge,
+                                                      std::optional<unsigned> lane) const;
+};
+
 mlir::memref::GlobalOp lookupGlobalForGetGlobal(mlir::ModuleOp moduleOp, mlir::memref::GetGlobalOp getGlobalOp);
 
 /// Resolves a value to contiguous backing storage when that storage can be
@@ -35,9 +85,12 @@ llvm::FailureOr<ResolvedContiguousAddress> resolveContiguousAddress(mlir::Value
 /// arithmetic and loop facts recorded in `knowledge`.
 llvm::FailureOr<int64_t> resolveIndexValue(mlir::Value value);
 llvm::FailureOr<int64_t> resolveIndexValue(mlir::Value value, const StaticValueKnowledge& knowledge);
+llvm::FailureOr<CompiledIndexExpr> compileIndexExpr(mlir::Value value);
 
 /// Follows alias, view, and DPS chains to recover the backing value of a
 /// loop-carried memref/result.
 mlir::Value resolveLoopCarriedAlias(mlir::Value value, const StaticValueKnowledge& knowledge);
 
+llvm::FailureOr<CompiledAddressExpr> compileContiguousAddressExpr(mlir::Value value);
+
 } // namespace onnx_mlir

src/PIM/Common/IR/BatchCoreUtils.cpp

@@ -0,0 +1,20 @@
+#include "src/Accelerators/PIM/Common/IR/BatchCoreUtils.hpp"
+#include "src/Accelerators/PIM/Common/PimCommon.hpp"
+
+namespace onnx_mlir {
+
+llvm::SmallVector<int32_t> getBatchCoreIds(pim::PimCoreBatchOp coreBatchOp) {
+  auto coreIdsAttr = coreBatchOp->getAttrOfType<mlir::DenseI32ArrayAttr>(onnx_mlir::kCoreIdsAttrName);
+  assert(coreIdsAttr && "pim.core_batch requires coreIds array attribute");
+  return llvm::SmallVector<int32_t>(coreIdsAttr.asArrayRef().begin(), coreIdsAttr.asArrayRef().end());
+}
+
+llvm::SmallVector<int32_t> getLaneChunkCoreIds(llvm::ArrayRef<int32_t> coreIds, size_t laneCount, unsigned lane) {
+  llvm::SmallVector<int32_t> laneCoreIds;
+  laneCoreIds.reserve(coreIds.size() / laneCount);
+  for (size_t chunkIndex = 0; chunkIndex < coreIds.size() / laneCount; ++chunkIndex)
+    laneCoreIds.push_back(coreIds[chunkIndex * laneCount + lane]);
+  return laneCoreIds;
+}
+
+} // namespace onnx_mlir

src/PIM/Common/IR/BatchCoreUtils.hpp

@@ -0,0 +1,14 @@
+#pragma once
+
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/SmallVector.h"
+
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+
+namespace onnx_mlir {
+
+llvm::SmallVector<int32_t> getBatchCoreIds(pim::PimCoreBatchOp coreBatchOp);
+
+llvm::SmallVector<int32_t> getLaneChunkCoreIds(llvm::ArrayRef<int32_t> coreIds, size_t laneCount, unsigned lane);
+
+} // namespace onnx_mlir

src/PIM/Common/IR/ConstantUtils.cpp

@@ -0,0 +1,104 @@
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/IR/Builders.h"
+#include "mlir/IR/BuiltinOps.h"
+#include "mlir/IR/Dialect.h"
+#include "mlir/IR/Matchers.h"
+
+#include "ConstantUtils.hpp"
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+
+Block* getHostConstantBlock(Operation* anchorOp) {
+  assert(anchorOp && "expected a valid anchor operation");
+
+  for (Operation* current = anchorOp; current; current = current->getParentOp())
+    if (isa<spatial::SpatCompute, spatial::SpatComputeBatch, pim::PimCoreOp, pim::PimCoreBatchOp>(current))
+      return current->getBlock();
+
+  if (auto funcOp = anchorOp->getParentOfType<func::FuncOp>())
+    return &funcOp.getBody().front();
+  if (auto moduleOp = anchorOp->getParentOfType<ModuleOp>())
+    return moduleOp.getBody();
+  return anchorOp->getBlock();
+}
+
+Value getOrCreateHostConstant(Operation* anchorOp, Attribute value, Type type, OperationFolder& folder) {
+  assert(anchorOp && "expected a valid anchor operation");
+  Block* hostBlock = getHostConstantBlock(anchorOp);
+  for (Operation& op : *hostBlock) {
+    auto constantOp = dyn_cast<arith::ConstantOp>(&op);
+    if (!constantOp || constantOp.getType() != type || constantOp.getValue() != value)
+      continue;
+    return constantOp.getResult();
+  }
+
+  auto* arithDialect = anchorOp->getContext()->getOrLoadDialect<arith::ArithDialect>();
+  return folder.getOrCreateConstant(hostBlock, arithDialect, value, type);
+}
+
+Value getOrCreateHostConstant(Operation* anchorOp, Attribute value, Type type, RewriterBase& rewriter) {
+  assert(anchorOp && "expected a valid anchor operation");
+  Block* hostBlock = getHostConstantBlock(anchorOp);
+  for (Operation& op : *hostBlock) {
+    auto constantOp = dyn_cast<arith::ConstantOp>(&op);
+    if (!constantOp || constantOp.getType() != type || constantOp.getValue() != value)
+      continue;
+    return constantOp.getResult();
+  }
+
+  OpBuilder::InsertionGuard guard(rewriter);
+  rewriter.setInsertionPointToStart(hostBlock);
+  return arith::ConstantOp::create(rewriter, anchorOp->getLoc(), type, cast<TypedAttr>(value)).getResult();
+}
+
+Value getOrCreateHostConstantLike(arith::ConstantOp constantOp, OperationFolder& folder) {
+  return getOrCreateHostConstant(constantOp.getOperation(), constantOp.getValue(), constantOp.getType(), folder);
+}
+
+Value getOrCreateHostIndexConstant(Operation* anchorOp, int64_t value, OperationFolder& folder) {
+  Builder builder(anchorOp->getContext());
+  return getOrCreateHostConstant(anchorOp, builder.getIndexAttr(value), builder.getIndexType(), folder);
+}
+
+Value getOrCreateHostIndexConstant(Operation* anchorOp, int64_t value, RewriterBase& rewriter) {
+  Builder builder(anchorOp->getContext());
+  return getOrCreateHostConstant(anchorOp, builder.getIndexAttr(value), builder.getIndexType(), rewriter);
+}
+
+Value getOrCreateHostI32Constant(Operation* anchorOp, int32_t value, OperationFolder& folder) {
+  Builder builder(anchorOp->getContext());
+  return getOrCreateHostConstant(anchorOp, builder.getI32IntegerAttr(value), builder.getI32Type(), folder);
+}
+
+Value getOrCreateHostI64Constant(Operation* anchorOp, int64_t value, OperationFolder& folder) {
+  Builder builder(anchorOp->getContext());
+  return getOrCreateHostConstant(anchorOp, builder.getI64IntegerAttr(value), builder.getI64Type(), folder);
+}
+
+Value createAffineApplyOrFoldedConstant(
+  RewriterBase& rewriter, Location loc, AffineMap map, ValueRange operands, Operation* anchorOp) {
+  SmallVector<Attribute> operandConstants;
+  operandConstants.reserve(operands.size());
+  for (Value operand : operands) {
+    APInt constantValue;
+    if (!matchPattern(operand, m_ConstantInt(&constantValue)))
+      return affine::AffineApplyOp::create(rewriter, loc, map, operands).getResult();
+    operandConstants.push_back(rewriter.getIndexAttr(constantValue.getSExtValue()));
+  }
+
+  SmallVector<Attribute> foldedResults;
+  if (succeeded(map.constantFold(operandConstants, foldedResults))) {
+    if (auto constantResult = dyn_cast<IntegerAttr>(foldedResults.front()))
+      return getOrCreateHostIndexConstant(anchorOp, constantResult.getInt(), rewriter);
+  }
+
+  return affine::AffineApplyOp::create(rewriter, loc, map, operands).getResult();
+}
+
+} // namespace onnx_mlir

src/PIM/Common/IR/ConstantUtils.hpp

@@ -0,0 +1,39 @@
+#pragma once
+
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/Value.h"
+#include "mlir/Transforms/FoldUtils.h"
+
+namespace onnx_mlir {
+
+mlir::Block* getHostConstantBlock(mlir::Operation* anchorOp);
+
+mlir::Value getOrCreateHostConstant(mlir::Operation* anchorOp,
+                                    mlir::Attribute value,
+                                    mlir::Type type,
+                                    mlir::OperationFolder& folder);
+
+mlir::Value getOrCreateHostConstant(mlir::Operation* anchorOp,
+                                    mlir::Attribute value,
+                                    mlir::Type type,
+                                    mlir::RewriterBase& rewriter);
+
+mlir::Value getOrCreateHostConstantLike(mlir::arith::ConstantOp constantOp, mlir::OperationFolder& folder);
+
+mlir::Value getOrCreateHostIndexConstant(mlir::Operation* anchorOp, int64_t value, mlir::OperationFolder& folder);
+
+mlir::Value getOrCreateHostIndexConstant(mlir::Operation* anchorOp, int64_t value, mlir::RewriterBase& rewriter);
+
+mlir::Value getOrCreateHostI32Constant(mlir::Operation* anchorOp, int32_t value, mlir::OperationFolder& folder);
+
+mlir::Value getOrCreateHostI64Constant(mlir::Operation* anchorOp, int64_t value, mlir::OperationFolder& folder);
+
+mlir::Value createAffineApplyOrFoldedConstant(mlir::RewriterBase& rewriter,
+                                              mlir::Location loc,
+                                              mlir::AffineMap map,
+                                              mlir::ValueRange operands,
+                                              mlir::Operation* anchorOp);
+
+} // namespace onnx_mlir

src/PIM/Common/IR/CoreBlockUtils.cpp

@@ -1,25 +1,37 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
 
+#include "llvm/ADT/SmallVector.h"
+
 #include "src/Accelerators/PIM/Common/IR/CoreBlockUtils.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
 namespace onnx_mlir {
 
 bool isCoreStaticAddressOp(mlir::Operation* op) {
-  return mlir::isa<mlir::arith::ConstantOp,
+  if (mlir::isa<mlir::arith::ConstantOp,
                 mlir::arith::AddIOp,
                 mlir::arith::SubIOp,
                 mlir::arith::MulIOp,
                 mlir::arith::DivUIOp,
+                mlir::arith::DivSIOp,
                 mlir::arith::MinUIOp,
                 mlir::arith::RemUIOp,
+                mlir::arith::RemSIOp,
                 mlir::arith::IndexCastOp,
+                mlir::arith::CmpIOp,
                 mlir::memref::AllocOp,
                 mlir::memref::SubViewOp,
                 mlir::memref::CastOp,
                 mlir::memref::CollapseShapeOp,
-                   mlir::memref::ExpandShapeOp>(op);
+                mlir::memref::ExpandShapeOp>(op))
+    return true;
+
+  if (auto selectOp = mlir::dyn_cast<mlir::arith::SelectOp>(op))
+    return selectOp.getType().isIntOrIndex();
+
+  return false;
 }
 
 mlir::LogicalResult
@@ -30,6 +42,9 @@ walkPimCoreBlock(mlir::Block& block,
   for (mlir::Operation& op : block) {
     if (mlir::isa<pim::PimHaltOp, mlir::scf::YieldOp>(op) || isCoreStaticAddressOp(&op))
       continue;
+    if (auto loadOp = mlir::dyn_cast<mlir::memref::LoadOp>(op);
+        loadOp && succeeded(resolveIndexValue(loadOp.getResult(), knowledge)))
+      continue;
 
     if (auto forOp = mlir::dyn_cast<mlir::scf::ForOp>(op)) {
       mlir::Block& loopBody = forOp.getRegion().front();
@@ -65,4 +80,58 @@ walkPimCoreBlock(mlir::Block& block,
   return mlir::success(!hasFailure);
 }
 
+mlir::LogicalResult walkPimCoreBlockStructurally(
+  mlir::Block& block,
+  const StaticValueKnowledge& knowledge,
+  llvm::function_ref<mlir::LogicalResult(mlir::Operation&, const StaticValueKnowledge&)> callback) {
+  bool hasFailure = false;
+  for (mlir::Operation& op : block) {
+    if (mlir::isa<pim::PimHaltOp, mlir::scf::YieldOp>(op) || isCoreStaticAddressOp(&op))
+      continue;
+    if (auto loadOp = mlir::dyn_cast<mlir::memref::LoadOp>(op);
+        loadOp && succeeded(resolveIndexValue(loadOp.getResult(), knowledge)))
+      continue;
+
+    if (auto forOp = mlir::dyn_cast<mlir::scf::ForOp>(op)) {
+      mlir::Block& loopBody = forOp.getRegion().front();
+      auto lowerBound = resolveIndexValue(forOp.getLowerBound(), knowledge);
+      auto upperBound = resolveIndexValue(forOp.getUpperBound(), knowledge);
+      auto step = resolveIndexValue(forOp.getStep(), knowledge);
+      if (failed(lowerBound) || failed(upperBound) || failed(step)) {
+        forOp.emitOpError("requires statically evaluable scf.for bounds for PIM verification");
+        hasFailure = true;
+        continue;
+      }
+      if (*step <= 0) {
+        forOp.emitOpError("requires positive scf.for step for PIM verification");
+        hasFailure = true;
+        continue;
+      }
+
+      llvm::SmallVector<int64_t, 2> samples;
+      if (*lowerBound < *upperBound) {
+        samples.push_back(*lowerBound);
+        int64_t last = *lowerBound + ((*upperBound - 1 - *lowerBound) / *step) * *step;
+        if (last != *lowerBound)
+          samples.push_back(last);
+      }
+
+      for (int64_t inductionValue : samples) {
+        StaticValueKnowledge loopKnowledge = knowledge;
+        loopKnowledge.indexValues[forOp.getInductionVar()] = inductionValue;
+        for (auto [iterArg, iterValue] : llvm::zip_equal(forOp.getRegionIterArgs(), forOp.getInitArgs()))
+          loopKnowledge.aliases[iterArg] = iterValue;
+
+        if (failed(walkPimCoreBlockStructurally(loopBody, loopKnowledge, callback)))
+          hasFailure = true;
+      }
+      continue;
+    }
+
+    if (failed(callback(op, knowledge)))
+      hasFailure = true;
+  }
+  return mlir::success(!hasFailure);
+}
+
 } // namespace onnx_mlir

src/PIM/Common/IR/CoreBlockUtils.hpp

@@ -21,4 +21,12 @@ walkPimCoreBlock(mlir::Block& block,
                  const StaticValueKnowledge& knowledge,
                  llvm::function_ref<mlir::LogicalResult(mlir::Operation&, const StaticValueKnowledge&)> callback);
 
+/// Walks a `pim.core`-like body structurally for verification without
+/// enumerating full loop trip counts. Loop bounds must still be statically
+/// evaluable so address resolution remains well-defined.
+mlir::LogicalResult walkPimCoreBlockStructurally(
+  mlir::Block& block,
+  const StaticValueKnowledge& knowledge,
+  llvm::function_ref<mlir::LogicalResult(mlir::Operation&, const StaticValueKnowledge&)> callback);
+
 } // namespace onnx_mlir

src/PIM/Common/IR/ShapeUtils.cpp

@@ -1,4 +1,5 @@
 #include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/ErrorHandling.h"
 
 #include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"
 
@@ -35,6 +36,30 @@ int64_t getNumElements(llvm::ArrayRef<int64_t> shape) {
   return numElements;
 }
 
+bool hasByteSizedElementType(mlir::Type elementType) {
+  if (mlir::isa<mlir::IndexType>(elementType))
+    return true;
+  if (auto intType = mlir::dyn_cast<mlir::IntegerType>(elementType))
+    return intType.getWidth() > 0 && intType.getWidth() % 8 == 0;
+  if (auto floatType = mlir::dyn_cast<mlir::FloatType>(elementType))
+    return floatType.getWidth() > 0 && floatType.getWidth() % 8 == 0;
+  return false;
+}
+
+size_t getElementTypeSizeInBytes(mlir::Type elementType) {
+  if (mlir::isa<mlir::IndexType>(elementType))
+    return mlir::IndexType::kInternalStorageBitWidth / 8;
+  if (auto intType = mlir::dyn_cast<mlir::IntegerType>(elementType))
+    return static_cast<size_t>(intType.getWidth() / 8);
+  if (auto floatType = mlir::dyn_cast<mlir::FloatType>(elementType))
+    return static_cast<size_t>(floatType.getWidth() / 8);
+  llvm_unreachable("expected byte-sized integer, float, or index element type");
+}
+
+size_t getShapedTypeSizeInBytes(mlir::ShapedType shapedType) {
+  return static_cast<size_t>(shapedType.getNumElements()) * getElementTypeSizeInBytes(shapedType.getElementType());
+}
+
 bool isMemoryContiguous(llvm::ArrayRef<int64_t> srcShape,
                         llvm::ArrayRef<int64_t> offsets,
                         llvm::ArrayRef<int64_t> sizes,

src/PIM/Common/IR/ShapeUtils.hpp

@@ -1,8 +1,13 @@
 #pragma once
 
+#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/Value.h"
+
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/SmallVector.h"
 
+#include <cstddef>
+
 namespace onnx_mlir {
 
 llvm::SmallVector<int64_t> computeRowMajorStrides(llvm::ArrayRef<int64_t> shape);
@@ -14,6 +19,12 @@ int64_t linearizeIndex(llvm::ArrayRef<int64_t> indices, llvm::ArrayRef<int64_t>
 
 int64_t getNumElements(llvm::ArrayRef<int64_t> shape);
 
+bool hasByteSizedElementType(mlir::Type elementType);
+
+size_t getElementTypeSizeInBytes(mlir::Type elementType);
+
+size_t getShapedTypeSizeInBytes(mlir::ShapedType shapedType);
+
 bool isMemoryContiguous(llvm::ArrayRef<int64_t> srcShape,
                         llvm::ArrayRef<int64_t> offsets,
                         llvm::ArrayRef<int64_t> sizes,

src/PIM/Common/IR/WeightUtils.cpp

@@ -1,8 +1,14 @@
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/BuiltinTypes.h"
 
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallSet.h"
 
+#include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
+#include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"
+#include "src/Accelerators/PIM/Common/IR/SubviewUtils.hpp"
 #include "src/Accelerators/PIM/Common/IR/WeightUtils.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
@@ -19,29 +25,79 @@ void markWeightAlways(mlir::Operation* op) {
 
 namespace {
 
-template <typename MVMOpTy, typename VMMOpTy, typename ParentOpTy>
+CompiledIndexExpr makeConstantExpr(int64_t constant) {
-bool hasMvmVmmWeightUse(ParentOpTy parentOp, unsigned weightIndex) {
+  CompiledIndexExprNode expr;
+  expr.kind = CompiledIndexExprNode::Kind::Constant;
+  expr.constant = constant;
+  return CompiledIndexExpr(std::make_shared<CompiledIndexExprNode>(std::move(expr)));
+}
+
+CompiledIndexExpr makeBinaryExpr(CompiledIndexExprNode::Kind kind, CompiledIndexExpr lhs, CompiledIndexExpr rhs) {
+  CompiledIndexExprNode expr;
+  expr.kind = kind;
+  expr.operands = {std::move(lhs), std::move(rhs)};
+  return CompiledIndexExpr(std::make_shared<CompiledIndexExprNode>(std::move(expr)));
+}
+
+CompiledIndexExpr addExpr(CompiledIndexExpr lhs, CompiledIndexExpr rhs) {
+  return makeBinaryExpr(CompiledIndexExprNode::Kind::Add, std::move(lhs), std::move(rhs));
+}
+
+CompiledIndexExpr mulExpr(CompiledIndexExpr lhs, int64_t rhs) {
+  return makeBinaryExpr(CompiledIndexExprNode::Kind::Mul, std::move(lhs), makeConstantExpr(rhs));
+}
+
+mlir::Value stripWeightViewOps(mlir::Value value) {
+  while (true) {
+    if (auto subviewOp = value.getDefiningOp<mlir::memref::SubViewOp>()) {
+      value = subviewOp.getSource();
+      continue;
+    }
+    if (auto castOp = value.getDefiningOp<mlir::memref::CastOp>()) {
+      value = castOp.getSource();
+      continue;
+    }
+    if (auto collapseOp = value.getDefiningOp<mlir::memref::CollapseShapeOp>()) {
+      value = collapseOp.getSrc();
+      continue;
+    }
+    if (auto expandOp = value.getDefiningOp<mlir::memref::ExpandShapeOp>()) {
+      value = expandOp.getSrc();
+      continue;
+    }
+    return value;
+  }
+}
+
+template <typename VMMOpTy, typename ParentOpTy>
+bool hasVmmWeightUse(ParentOpTy parentOp, unsigned weightIndex) {
+  auto weightArg = parentOp.getWeightArgument(weightIndex);
+  if (!weightArg)
+    return false;
   bool found = false;
   parentOp.walk([&](mlir::Operation* op) {
-    if (auto mvmOp = mlir::dyn_cast<MVMOpTy>(op))
+    if (auto vmmOp = mlir::dyn_cast<VMMOpTy>(op))
-      found |= mvmOp.getWeightIndex() == weightIndex;
+      found |= vmmOp.getWeight() == *weightArg;
-    else if (auto vmmOp = mlir::dyn_cast<VMMOpTy>(op))
-      found |= vmmOp.getWeightIndex() == weightIndex;
   });
   return found;
 }
 
-template <typename MVMOpTy, typename VMMOpTy, typename ParentOpTy>
+template <typename VMMOpTy, typename ParentOpTy>
-void walkMvmVmmWeightUses(ParentOpTy parentOp, llvm::function_ref<void(mlir::OpOperand&)> callback) {
+void walkVmmWeightUses(ParentOpTy parentOp, llvm::function_ref<void(mlir::OpOperand&)> callback) {
   auto weights = parentOp.getWeights();
   llvm::SmallSet<unsigned, 8> visited;
-  auto walkWeightIndex = [&](unsigned weightIndex) {
+  auto walkWeight = [&](mlir::Value weight) {
-    if (weightIndex < weights.size() && visited.insert(weightIndex).second)
+    for (unsigned weightIndex = 0; weightIndex < weights.size(); ++weightIndex) {
+      auto weightArg = parentOp.getWeightArgument(weightIndex);
+      if (!weightArg || *weightArg != weight)
+        continue;
+      if (visited.insert(weightIndex).second)
         callback(parentOp->getOpOperand(weightIndex));
+      break;
+    }
   };
 
-  parentOp.walk([&](MVMOpTy op) { walkWeightIndex(op.getWeightIndex()); });
+  parentOp.walk([&](VMMOpTy op) { walkWeight(op.getWeight()); });
-  parentOp.walk([&](VMMOpTy op) { walkWeightIndex(op.getWeightIndex()); });
 }
 
 } // namespace
@@ -54,7 +110,7 @@ bool isSpatialMvmVmmWeightUse(mlir::OpOperand& use) {
   if (!computeOp || operandIndex >= computeOp.getWeights().size())
     return false;
 
-  return hasMvmVmmWeightUse<spatial::SpatMVMOp, spatial::SpatVMMOp>(computeOp, operandIndex);
+  return hasVmmWeightUse<spatial::SpatVMMOp>(computeOp, operandIndex);
 }
 
 bool hasOnlySpatialMvmVmmWeightUses(mlir::Value value) {
@@ -76,8 +132,8 @@ bool hasOnlySpatialMvmVmmWeightUses(mlir::Value value) {
         return expandShapeOp.getSrc() == currentValue && self(expandShapeOp.getResult(), self);
       if (auto collapseShapeOp = mlir::dyn_cast<mlir::tensor::CollapseShapeOp>(user))
         return collapseShapeOp.getSrc() == currentValue && self(collapseShapeOp.getResult(), self);
-      if (auto transposeOp = mlir::dyn_cast<mlir::ONNXTransposeOp>(user))
+      if (auto transposeOp = mlir::dyn_cast<mlir::linalg::TransposeOp>(user))
-        return transposeOp.getData() == currentValue && self(transposeOp.getResult(), self);
+        return transposeOp.getInput() == currentValue && self(transposeOp.getResult()[0], self);
 
       return false;
     });
@@ -90,19 +146,153 @@ void walkPimMvmVmmWeightUses(mlir::Operation* root, llvm::function_ref<void(mlir
   assert(root && "expected valid root op");
   root->walk([&](pim::PimCoreOp coreOp) {
     coreOp.walk([&](pim::PimVMMOp vmmOp) {
-      auto weights = coreOp.getWeights();
+      if (auto weightIndex = resolveWeightIndex(coreOp.getOperation(), vmmOp.getWeight()))
-      unsigned weightIndex = vmmOp.getWeightIndex();
+        callback(coreOp->getOpOperand(*weightIndex));
-      if (weightIndex < weights.size())
-        callback(coreOp->getOpOperand(weightIndex));
     });
   });
   root->walk([&](pim::PimCoreBatchOp coreBatchOp) {
-    auto weights = coreBatchOp.getWeights();
+    coreBatchOp.walk([&](pim::PimVMMOp vmmOp) {
-    for (auto weight : weights)
+      if (auto weightIndex = resolveWeightIndex(coreBatchOp.getOperation(), vmmOp.getWeight()))
-      for (mlir::OpOperand& use : weight.getUses())
+        callback(coreBatchOp->getOpOperand(*weightIndex));
-        if (use.getOwner() == coreBatchOp.getOperation())
+    });
-          callback(use);
   });
 }
 
+std::optional<unsigned> resolveWeightIndex(mlir::Operation* weightOwner, mlir::Value weight) {
+  weight = stripWeightViewOps(weight);
+
+  if (auto coreOp = mlir::dyn_cast_or_null<pim::PimCoreOp>(weightOwner)) {
+    for (unsigned weightIndex = 0; weightIndex < coreOp.getWeights().size(); ++weightIndex)
+      if (coreOp.getWeightArgument(weightIndex) == weight)
+        return weightIndex;
+    return std::nullopt;
+  }
+
+  if (auto coreBatchOp = mlir::dyn_cast_or_null<pim::PimCoreBatchOp>(weightOwner)) {
+    for (unsigned weightIndex = 0; weightIndex < coreBatchOp.getWeights().size(); ++weightIndex)
+      if (coreBatchOp.getWeightArgument(weightIndex) == weight)
+        return weightIndex;
+    return std::nullopt;
+  }
+
+  return std::nullopt;
+}
+
+std::optional<unsigned> resolveWeightIndex(mlir::Operation* weightOwner, pim::PimVMMOp vmmOp) {
+  return resolveWeightIndex(weightOwner, vmmOp.getWeight());
+}
+
+llvm::FailureOr<ResolvedWeightView>
+resolveWeightView(mlir::Operation* weightOwner, mlir::Value weight, const StaticValueKnowledge& knowledge) {
+  llvm::SmallVector<mlir::Operation*> viewOps;
+  mlir::Value current = weight;
+
+  while (true) {
+    if (auto defOp = current.getDefiningOp()) {
+      if (auto getGlobalOp = mlir::dyn_cast<mlir::memref::GetGlobalOp>(defOp)) {
+        auto moduleOp = weightOwner ? weightOwner->getParentOfType<mlir::ModuleOp>() : mlir::ModuleOp {};
+        auto globalOp = lookupGlobalForGetGlobal(moduleOp, getGlobalOp);
+        if (!globalOp || !globalOp.getInitialValue())
+          return mlir::failure();
+
+        auto denseAttr = mlir::dyn_cast<mlir::DenseElementsAttr>(*globalOp.getInitialValue());
+        if (!denseAttr)
+          return mlir::failure();
+
+        ResolvedWeightView view;
+        view.globalOp = globalOp;
+        view.shape.assign(denseAttr.getType().getShape().begin(), denseAttr.getType().getShape().end());
+        view.strides = computeRowMajorStrides(view.shape);
+
+        CompiledIndexExpr offsetExpr = makeConstantExpr(0);
+        for (mlir::Operation* viewOp : llvm::reverse(viewOps)) {
+          if (auto subview = mlir::dyn_cast<mlir::memref::SubViewOp>(viewOp)) {
+            llvm::SmallVector<int64_t> nextStrides;
+            nextStrides.reserve(subview.getMixedOffsets().size());
+            for (auto [offset, stride, sourceStride] :
+                 llvm::zip_equal(subview.getMixedOffsets(), subview.getStaticStrides(), view.strides)) {
+              CompiledIndexExpr offsetValue = makeConstantExpr(0);
+              if (auto attr = mlir::dyn_cast<mlir::Attribute>(offset)) {
+                auto intAttr = mlir::dyn_cast<mlir::IntegerAttr>(attr);
+                if (!intAttr)
+                  return mlir::failure();
+                offsetValue = makeConstantExpr(intAttr.getInt());
+              }
+              else if (auto value = mlir::dyn_cast<mlir::Value>(offset)) {
+                auto compiledOffset = compileIndexExpr(value);
+                if (failed(compiledOffset))
+                  return mlir::failure();
+                offsetValue = *compiledOffset;
+              }
+              else {
+                return mlir::failure();
+              }
+              offsetExpr = addExpr(std::move(offsetExpr), mulExpr(std::move(offsetValue), sourceStride));
+              nextStrides.push_back(stride * sourceStride);
+            }
+            view.shape.assign(subview.getStaticSizes().begin(), subview.getStaticSizes().end());
+            view.strides = std::move(nextStrides);
+            continue;
+          }
+
+          if (auto collapse = mlir::dyn_cast<mlir::memref::CollapseShapeOp>(viewOp)) {
+            if (view.strides != computeRowMajorStrides(view.shape))
+              return mlir::failure();
+            auto resultType = mlir::cast<mlir::MemRefType>(collapse.getResult().getType());
+            view.shape.assign(resultType.getShape().begin(), resultType.getShape().end());
+            view.strides = computeRowMajorStrides(view.shape);
+            continue;
+          }
+
+          if (auto expand = mlir::dyn_cast<mlir::memref::ExpandShapeOp>(viewOp)) {
+            if (view.strides != computeRowMajorStrides(view.shape))
+              return mlir::failure();
+            auto resultType = mlir::cast<mlir::MemRefType>(expand.getResult().getType());
+            view.shape.assign(resultType.getShape().begin(), resultType.getShape().end());
+            view.strides = computeRowMajorStrides(view.shape);
+          }
+        }
+
+        auto resolvedOffset = offsetExpr.evaluate(knowledge);
+        if (failed(resolvedOffset))
+          return mlir::failure();
+        view.offset = *resolvedOffset;
+        return view;
+      }
+
+      if (mlir::isa<mlir::memref::SubViewOp, mlir::memref::CollapseShapeOp, mlir::memref::ExpandShapeOp>(defOp)) {
+        viewOps.push_back(defOp);
+        if (auto subview = mlir::dyn_cast<mlir::memref::SubViewOp>(defOp))
+          current = subview.getSource();
+        else if (auto collapse = mlir::dyn_cast<mlir::memref::CollapseShapeOp>(defOp))
+          current = collapse.getSrc();
+        else
+          current = mlir::cast<mlir::memref::ExpandShapeOp>(defOp).getSrc();
+        continue;
+      }
+
+      if (auto castOp = mlir::dyn_cast<mlir::memref::CastOp>(defOp)) {
+        current = castOp.getSource();
+        continue;
+      }
+
+      return mlir::failure();
+    }
+
+    auto weightIndex = resolveWeightIndex(weightOwner, current);
+    if (!weightIndex)
+      return mlir::failure();
+
+    if (auto coreOp = mlir::dyn_cast_or_null<pim::PimCoreOp>(weightOwner)) {
+      current = coreOp.getWeights()[*weightIndex];
+      continue;
+    }
+    if (auto coreBatchOp = mlir::dyn_cast_or_null<pim::PimCoreBatchOp>(weightOwner)) {
+      current = coreBatchOp.getWeights()[*weightIndex];
+      continue;
+    }
+    return mlir::failure();
+  }
+}
+
 } // namespace onnx_mlir

src/PIM/Common/IR/WeightUtils.hpp

@@ -1,15 +1,34 @@
 #pragma once
 
-#include "mlir/IR/Operation.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/Value.h"
 
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/STLFunctionalExtras.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 
+#include <optional>
+
+#include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+
 inline constexpr llvm::StringRef PimWeightAlwaysAttrName = "weightAlways";
 
 namespace onnx_mlir {
 
+struct ResolvedWeightView {
+  mlir::memref::GlobalOp globalOp;
+  llvm::SmallVector<int64_t> shape;
+  llvm::SmallVector<int64_t> strides;
+  int64_t offset = 0;
+
+  bool operator==(const ResolvedWeightView& other) const {
+    return globalOp == other.globalOp && shape == other.shape && strides == other.strides && offset == other.offset;
+  }
+};
+
 bool hasWeightAlways(mlir::Operation* op);
 
 /// Tags an op as producing a value that should stay materialized as a reusable
@@ -26,4 +45,21 @@ bool hasOnlySpatialMvmVmmWeightUses(mlir::Value value);
 /// passes can identify globals that must remain weight-backed.
 void walkPimMvmVmmWeightUses(mlir::Operation* root, llvm::function_ref<void(mlir::OpOperand&)> callback);
 
+std::optional<unsigned> resolveWeightIndex(mlir::Operation* weightOwner, mlir::Value weight);
+std::optional<unsigned> resolveWeightIndex(mlir::Operation* weightOwner, pim::PimVMMOp vmmOp);
+llvm::FailureOr<ResolvedWeightView>
+resolveWeightView(mlir::Operation* weightOwner, mlir::Value weight, const StaticValueKnowledge& knowledge = {});
+
+template <typename CoreLikeOpTy>
+llvm::SmallVector<unsigned, 8> getUsedWeightIndices(CoreLikeOpTy coreLikeOp) {
+  llvm::SmallVector<unsigned, 8> indices;
+  coreLikeOp.walk([&](pim::PimVMMOp vmmOp) {
+    auto weightIndex = resolveWeightIndex(coreLikeOp.getOperation(), vmmOp.getWeight());
+    if (weightIndex && !llvm::is_contained(indices, *weightIndex))
+      indices.push_back(*weightIndex);
+  });
+  llvm::sort(indices);
+  return indices;
+}
+
 } // namespace onnx_mlir

src/PIM/Common/PimCommon.hpp

@@ -12,6 +12,7 @@
 #include "llvm/ADT/StringRef.h"
 
 #include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
+#include "src/Accelerators/PIM/Common/IR/ConstantUtils.hpp"
 #include "src/Accelerators/PIM/Common/IR/CoreBlockUtils.hpp"
 #include "src/Accelerators/PIM/Common/IR/EntryPointUtils.hpp"
 #include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"

src/PIM/Common/Support/DebugDump.cpp

@@ -18,7 +18,7 @@ void dumpModule(mlir::ModuleOp moduleOp, const std::string& name) {
   std::fstream file(dialectsDir + "/" + name + ".mlir", std::ios::out);
   llvm::raw_os_ostream os(file);
   mlir::OpPrintingFlags flags;
-  flags.elideLargeElementsAttrs();
+  flags.elideLargeElementsAttrs().enableDebugInfo(true,false);
   moduleOp.print(os, flags);
   os.flush();
   file.close();

src/PIM/Common/Support/Diagnostics.hpp

@@ -13,7 +13,8 @@
 namespace onnx_mlir::pim {
 
 struct CappedDiagnosticReporter {
-  explicit CappedDiagnosticReporter(int64_t maxReportedFailures = 8) : maxReportedFailures(maxReportedFailures) {}
+  explicit CappedDiagnosticReporter(int64_t maxReportedFailures = 8)
+  : maxReportedFailures(maxReportedFailures) {}
 
   template <typename EmitFn>
   void report(mlir::Operation* op, EmitFn&& emit) {
@@ -24,8 +25,7 @@ struct CappedDiagnosticReporter {
 
   void emitSuppressedSummary(mlir::Operation* op, llvm::StringRef failureDescription) const {
     if (numFailures > maxReportedFailures)
-      op->emitError() << "suppressed " << (numFailures - maxReportedFailures) << " additional "
+      op->emitError() << "suppressed " << (numFailures - maxReportedFailures) << " additional " << failureDescription;
-                      << failureDescription;
   }
 
   bool hasFailure() const { return numFailures != 0; }

src/PIM/Compiler/CMakeLists.txt

@@ -16,7 +16,6 @@ add_pim_library(OMPimCompilerOptions
 add_pim_library(OMPimCompilerUtils
   PimCompilerUtils.cpp
   PimArtifactWriter.cpp
-  PimBatchEmission.cpp
   PimCodeGen.cpp
   PimWeightEmitter.cpp
 

src/PIM/Compiler/PimArtifactWriter.cpp

@@ -48,7 +48,7 @@ writeMemoryBinary(ModuleOp moduleOp, func::FuncOp funcOp, PimAcceleratorMemory&
     if (!denseAttr)
       return;
 
-    MemEntry memEntry = memory.hostMem.getMemEntry(getGlobalOp.getResult());
+    MemEntry memEntry = memory.hostMem.getMemEntry({getGlobalOp.getResult(), std::nullopt});
     ArrayRef<char> rawData = denseAttr.getRawData();
     char* dst = memoryBuffer.data() + memEntry.address;
 

src/PIM/Compiler/PimBatchEmission.cpp

@@ -1,136 +0,0 @@
-#include "mlir/IR/Builders.h"
-#include "mlir/IR/BuiltinOps.h"
-#include "mlir/IR/IRMapping.h"
-
-#include "src/Accelerators/PIM/Common/PimCommon.hpp"
-#include "src/Accelerators/PIM/Compiler/PimBatchEmission.hpp"
-
-using namespace mlir;
-
-namespace onnx_mlir {
-namespace {
-
-static SmallVector<int32_t> getBatchCoreIds(pim::PimCoreBatchOp coreBatchOp) {
-  auto coreIdsAttr = coreBatchOp->getAttrOfType<DenseI32ArrayAttr>(onnx_mlir::kCoreIdsAttrName);
-  assert(coreIdsAttr && "pim.core_batch requires coreIds array attribute");
-  return SmallVector<int32_t>(coreIdsAttr.asArrayRef().begin(), coreIdsAttr.asArrayRef().end());
-}
-
-static SmallVector<int32_t> getLaneChunkCoreIds(ArrayRef<int32_t> coreIds, size_t laneCount, unsigned lane) {
-  SmallVector<int32_t> laneCoreIds;
-  laneCoreIds.reserve(coreIds.size() / laneCount);
-  for (size_t chunkIndex = 0; chunkIndex < coreIds.size() / laneCount; ++chunkIndex)
-    laneCoreIds.push_back(coreIds[chunkIndex * laneCount + lane]);
-  return laneCoreIds;
-}
-
-static void scalarizeBatchOpsInCore(pim::PimCoreOp scalarCore, size_t laneCount, unsigned lane) {
-  IRRewriter rewriter(scalarCore.getContext());
-  SmallVector<Operation*> batchOps;
-  scalarCore.walk([&](Operation* op) {
-    if (isa<pim::PimSendBatchOp,
-            pim::PimSendTensorBatchOp,
-            pim::PimReceiveBatchOp,
-            pim::PimReceiveTensorBatchOp,
-            pim::PimMemCopyHostToDevBatchOp>(op)) {
-      batchOps.push_back(op);
-    }
-  });
-
-  for (Operation* op : batchOps) {
-    rewriter.setInsertionPoint(op);
-
-    if (auto sendBatchOp = dyn_cast<pim::PimSendBatchOp>(op)) {
-      pim::PimSendOp::create(rewriter,
-                             sendBatchOp.getLoc(),
-                             sendBatchOp.getInput(),
-                             sendBatchOp.getSizeAttr(),
-                             rewriter.getI32IntegerAttr(sendBatchOp.getTargetCoreIds()[lane]));
-      rewriter.eraseOp(op);
-      continue;
-    }
-
-    if (auto sendTensorBatchOp = dyn_cast<pim::PimSendTensorBatchOp>(op)) {
-      pim::PimSendTensorOp::create(
-        rewriter,
-        sendTensorBatchOp.getLoc(),
-        sendTensorBatchOp.getInput(),
-        rewriter.getDenseI32ArrayAttr(getLaneChunkCoreIds(sendTensorBatchOp.getTargetCoreIds(), laneCount, lane)));
-      rewriter.eraseOp(op);
-      continue;
-    }
-
-    if (auto receiveBatchOp = dyn_cast<pim::PimReceiveBatchOp>(op)) {
-      auto scalarReceive =
-        pim::PimReceiveOp::create(rewriter,
-                                  receiveBatchOp.getLoc(),
-                                  receiveBatchOp.getOutput().getType(),
-                                  receiveBatchOp.getOutputBuffer(),
-                                  receiveBatchOp.getSizeAttr(),
-                                  rewriter.getI32IntegerAttr(receiveBatchOp.getSourceCoreIds()[lane]));
-      rewriter.replaceOp(op, scalarReceive->getResults());
-      continue;
-    }
-
-    if (auto receiveTensorBatchOp = dyn_cast<pim::PimReceiveTensorBatchOp>(op)) {
-      auto scalarReceive = pim::PimReceiveTensorOp::create(
-        rewriter,
-        receiveTensorBatchOp.getLoc(),
-        receiveTensorBatchOp.getOutput().getType(),
-        receiveTensorBatchOp.getOutputBuffer(),
-        rewriter.getDenseI32ArrayAttr(getLaneChunkCoreIds(receiveTensorBatchOp.getSourceCoreIds(), laneCount, lane)));
-      rewriter.replaceOp(op, scalarReceive->getResults());
-      continue;
-    }
-
-    auto memcpBatchOp = cast<pim::PimMemCopyHostToDevBatchOp>(op);
-    auto scalarCopy = pim::PimMemCopyHostToDevOp::create(rewriter,
-                                                         memcpBatchOp.getLoc(),
-                                                         memcpBatchOp.getOutput().getType(),
-                                                         memcpBatchOp.getDeviceTarget(),
-                                                         memcpBatchOp.getHostSource(),
-                                                         memcpBatchOp.getDeviceTargetOffsetAttr(),
-                                                         memcpBatchOp.getHostSourceOffsetAttr(),
-                                                         memcpBatchOp.getSizeAttr());
-    rewriter.replaceOp(op, scalarCopy->getResults());
-  }
-}
-
-} // namespace
-
-LogicalResult withScalarCoreFromBatchLane(pim::PimCoreBatchOp coreBatchOp,
-                                          unsigned lane,
-                                          llvm::function_ref<LogicalResult(pim::PimCoreOp)> callback) {
-  OwningOpRef<ModuleOp> scratchModule = ModuleOp::create(coreBatchOp.getLoc());
-  OpBuilder builder(scratchModule->getContext());
-  builder.setInsertionPointToStart(scratchModule->getBody());
-
-  size_t laneCount = static_cast<size_t>(coreBatchOp.getLaneCount());
-  size_t weightsPerLane = coreBatchOp.getWeights().size() / laneCount;
-  SmallVector<Value> laneWeights;
-  laneWeights.reserve(weightsPerLane);
-  for (size_t weightIndex = 0; weightIndex < weightsPerLane; ++weightIndex)
-    laneWeights.push_back(coreBatchOp.getWeights()[lane * weightsPerLane + weightIndex]);
-
-  auto coreIds = getBatchCoreIds(coreBatchOp);
-  auto scalarCore = pim::PimCoreOp::create(
-    builder, coreBatchOp.getLoc(), ValueRange(laneWeights), builder.getI32IntegerAttr(coreIds[lane]));
-  Block* block = builder.createBlock(&scalarCore.getBody(), scalarCore.getBody().end());
-  IRMapping mapper;
-  if (coreBatchOp.getBody().front().getNumArguments() == 1)
-    mapper.map(coreBatchOp.getBody().front().getArgument(0), coreBatchOp.getInputs()[lane]);
-
-  builder.setInsertionPointToEnd(block);
-  for (Operation& op : coreBatchOp.getBody().front()) {
-    Operation* cloned = builder.clone(op, mapper);
-    for (auto [originalResult, clonedResult] : llvm::zip(op.getResults(), cloned->getResults()))
-      mapper.map(originalResult, clonedResult);
-  }
-
-  if (block->empty() || !isa<pim::PimHaltOp>(block->back()))
-    pim::PimHaltOp::create(builder, coreBatchOp.getLoc());
-  scalarizeBatchOpsInCore(scalarCore, laneCount, lane);
-  return callback(scalarCore);
-}
-
-} // namespace onnx_mlir

src/PIM/Compiler/PimBatchEmission.hpp

@@ -1,13 +0,0 @@
-#pragma once
-
-#include "llvm/ADT/STLFunctionalExtras.h"
-
-#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
-
-namespace onnx_mlir {
-
-mlir::LogicalResult withScalarCoreFromBatchLane(pim::PimCoreBatchOp coreBatchOp,
-                                                unsigned lane,
-                                                llvm::function_ref<mlir::LogicalResult(pim::PimCoreOp)> callback);
-
-} // namespace onnx_mlir

src/PIM/Compiler/PimCodeGen.hpp

@@ -4,13 +4,16 @@
 
 #include "llvm-project/clang/include/clang/Basic/LLVM.h"
 #include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/Hashing.h"
 #include "llvm/Support/JSON.h"
 #include "llvm/Support/raw_os_ostream.h"
 
 #include <fstream>
+#include <limits>
 #include <optional>
 
 #include "onnx-mlir/Compiler/OMCompilerTypes.h"
+#include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
 #include "src/Accelerators/PIM/Common/PimCommon.hpp"
 #include "src/Accelerators/PIM/Common/Support/ReportUtils.hpp"
 #include "src/Accelerators/PIM/Compiler/PimBinaryFormat.hpp"
@@ -23,6 +26,13 @@ struct MemEntry {
   size_t size;
 };
 
+struct MemoryValueKey {
+  mlir::Value value;
+  std::optional<unsigned> lane;
+
+  bool operator==(const MemoryValueKey& other) const { return value == other.value && lane == other.lane; }
+};
+
 struct MemoryReportRow {
   uint64_t numAlloca = 0;
   uint64_t sizeAlloca = 0;
@@ -50,33 +60,33 @@ struct MemoryReportEntry {
 };
 
 class PimMemory {
-  llvm::SmallVector<std::pair<MemEntry, mlir::Value>, 32> memEntries;
+  llvm::SmallVector<std::pair<MemEntry, MemoryValueKey>, 32> memEntries;
-  llvm::SmallDenseMap<mlir::Value, MemEntry, 32>& globalMemEntriesMap;
+  llvm::SmallDenseMap<MemoryValueKey, MemEntry, 32>& globalMemEntriesMap;
-  llvm::SmallDenseMap<mlir::Value, MemEntry, 32> ownedMemEntriesMap;
+  llvm::SmallDenseMap<MemoryValueKey, MemEntry, 32> ownedMemEntriesMap;
 
   size_t minAlignment = 4;
   size_t firstAvailableAddress = 0;
 
-  MemEntry* gatherMemEntry(mlir::Value value);
+  MemEntry* gatherMemEntry(mlir::Value value, std::optional<unsigned> lane = std::nullopt);
   void allocateGatheredMemory();
-  void allocateMemoryForValue(mlir::Value value, MemEntry& memEntry);
+  void allocateMemoryForValue(const MemoryValueKey& key, MemEntry& memEntry);
 
 public:
-  PimMemory(llvm::SmallDenseMap<mlir::Value, MemEntry, 32>& globalMemEntriesMap)
+  PimMemory(llvm::SmallDenseMap<MemoryValueKey, MemEntry, 32>& globalMemEntriesMap)
   : globalMemEntriesMap(globalMemEntriesMap) {}
 
   void allocateHost(mlir::ModuleOp moduleOp, mlir::func::FuncOp funcOp);
-  void allocateCore(mlir::Operation* op);
+  void allocateCore(mlir::Operation* op, std::optional<unsigned> lane = std::nullopt);
   MemoryReportRow getReportRow() const;
   void remove(mlir::Value val);
 
   size_t getFirstAvailableAddress() const { return firstAvailableAddress; }
-  MemEntry getMemEntry(mlir::Value value) const;
+  MemEntry getMemEntry(const MemoryValueKey& key) const;
 };
 
 class PimAcceleratorMemory {
 public:
-  llvm::SmallDenseMap<mlir::Value, MemEntry, 32> memEntriesMap;
+  llvm::SmallDenseMap<MemoryValueKey, MemEntry, 32> memEntriesMap;
   PimMemory hostMem;
 
 private:
@@ -84,14 +94,23 @@ private:
   std::fstream fileReport;
   std::optional<MemoryReportRow> hostReportRow;
   llvm::SmallVector<MemoryReportEntry, 32> reportEntries;
+  mutable llvm::DenseMap<mlir::Value, CompiledIndexExpr> compiledIndexExprs;
+  mutable llvm::DenseMap<mlir::Value, CompiledAddressExpr> compiledAddressExprs;
 
 public:
   PimAcceleratorMemory()
   : hostMem(memEntriesMap), fileReport(openReportFile("memory_report")) {}
+  PimAcceleratorMemory(const llvm::SmallDenseMap<MemoryValueKey, MemEntry, 32>& initialMemEntries, bool enableReport)
+  : memEntriesMap(initialMemEntries),
+    hostMem(memEntriesMap),
+    fileReport(enableReport ? openReportFile("memory_report") : std::fstream()) {}
 
   PimMemory& getOrCreateDeviceMem(size_t id);
 
-  size_t getValueAddress(mlir::Value value, const StaticValueKnowledge& knowledge = {}) const;
+  size_t getValueAddress(mlir::Value value,
+                         const StaticValueKnowledge& knowledge = {},
+                         std::optional<unsigned> lane = std::nullopt) const;
+  llvm::FailureOr<int64_t> getIndexValue(mlir::Value value, const StaticValueKnowledge& knowledge = {}) const;
   void reportHost();
   void recordCoreReport(size_t coreId, const MemoryReportRow& row);
   void recordBatchReport(uint64_t batchId,
@@ -103,15 +122,24 @@ public:
   void clean(mlir::Operation* op);
 };
 
+struct CoreEmissionJob {
+  mlir::Operation* coreLikeOp = nullptr;
+  size_t originalCoreId = 0;
+  size_t emittedCoreId = 0;
+  llvm::SmallVector<unsigned, 4> lanes;
+  std::optional<uint64_t> batchReportId;
+};
+
 class PimCodeGen {
   PimAcceleratorMemory& memory;
   llvm::raw_fd_ostream& coreBinaryStream;
   llvm::raw_fd_ostream* coreJsonStream;
   const llvm::DenseMap<size_t, size_t>& emittedCoreIds;
+  std::optional<unsigned> batchLane;
   mutable uint32_t emittedInstructionCount = 0;
 
   size_t addressOf(mlir::Value value, const StaticValueKnowledge& knowledge) const {
-    return memory.getValueAddress(value, knowledge);
+    return memory.getValueAddress(value, knowledge, batchLane);
   }
   size_t remapCoreId(size_t coreId) const;
 
@@ -141,15 +169,18 @@ public:
   : memory(memory), coreBinaryStream(coreBinary), coreJsonStream(coreJson), emittedCoreIds(emittedCoreIds) {}
 
   uint32_t getEmittedInstructionCount() const { return emittedInstructionCount; }
+  void setBatchLane(std::optional<unsigned> lane) { batchLane = lane; }
+  llvm::FailureOr<int64_t> indexOf(mlir::Value value, const StaticValueKnowledge& knowledge) const {
+    return memory.getIndexValue(value, knowledge);
+  }
 
   void codeGenLoadOp(pim::PimMemCopyHostToDevOp loadOp, const StaticValueKnowledge& knowledge) const;
+  void codeGenLoadBatchOp(pim::PimMemCopyHostToDevBatchOp loadOp, const StaticValueKnowledge& knowledge) const;
   void codeGenStoreOp(pim::PimMemCopyDevToHostOp storeOp, const StaticValueKnowledge& knowledge) const;
   void codeGenLmvOp(pim::PimMemCopyOp lmvOp, const StaticValueKnowledge& knowledge) const;
 
   void codeGenReceiveOp(pim::PimReceiveOp receiveOp, const StaticValueKnowledge& knowledge) const;
-  void codeGenReceiveTensorOp(pim::PimReceiveTensorOp receiveTensorOp, const StaticValueKnowledge& knowledge) const;
   void codeGenSendOp(pim::PimSendOp sendOp, const StaticValueKnowledge& knowledge) const;
-  void codeGenSendTensorOp(pim::PimSendTensorOp sendTensorOp, const StaticValueKnowledge& knowledge) const;
   void codeGenConcatOp(pim::PimConcatOp concatOp, const StaticValueKnowledge& knowledge) const;
 
   template <typename MVMTy>
@@ -172,3 +203,20 @@ public:
 OnnxMlirCompilerErrorCodes compileToPimCode(mlir::ModuleOp& moduleOpRef, std::string& outputDirName);
 
 } // namespace onnx_mlir
+
+namespace llvm {
+
+template <>
+struct DenseMapInfo<onnx_mlir::MemoryValueKey> {
+  static onnx_mlir::MemoryValueKey getEmptyKey() { return {DenseMapInfo<mlir::Value>::getEmptyKey(), 0}; }
+
+  static onnx_mlir::MemoryValueKey getTombstoneKey() { return {DenseMapInfo<mlir::Value>::getTombstoneKey(), 0}; }
+
+  static unsigned getHashValue(const onnx_mlir::MemoryValueKey& key) {
+    return hash_combine(key.value, key.lane.value_or(std::numeric_limits<unsigned>::max()));
+  }
+
+  static bool isEqual(const onnx_mlir::MemoryValueKey& lhs, const onnx_mlir::MemoryValueKey& rhs) { return lhs == rhs; }
+};
+
+} // namespace llvm

src/PIM/Compiler/PimCompilerOptions.cpp

@@ -1,7 +1,7 @@
-#include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
-
 #include "llvm/Support/ErrorHandling.h"
 
+#include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
+
 #define DEBUG_TYPE "PimCompilerOptions"
 
 namespace onnx_mlir {
@@ -15,11 +15,10 @@ llvm::cl::opt<PimEmissionTargetType> pimEmissionTarget(
   llvm::cl::init(EmitPimCodegen),
   llvm::cl::cat(OnnxMlirOptions));
 
-llvm::cl::opt<PimMergeSchedulerType> pimMergeScheduler(
+llvm::cl::opt<PimMergeSchedulerType>
-  "pim-merge-scheduler",
+  pimMergeScheduler("pim-merge-scheduler",
                     llvm::cl::desc("Scheduler used by the Spatial merge-compute-nodes pass"),
                     llvm::cl::values(clEnumValN(MergeSchedulerPeft, "peft", "Use PEFT scheduling")),
-  llvm::cl::values(clEnumValN(MergeSchedulerDcp, "dcp", "Use the legacy DCP-inspired scheduler")),
                     llvm::cl::init(MergeSchedulerPeft),
                     llvm::cl::cat(OnnxMlirOptions));
 
@@ -49,12 +48,6 @@ llvm::cl::opt<long> coresCount("core-count",
                                llvm::cl::desc("Number of cores in the chip. Required for PIM compilation."),
                                llvm::cl::init(-1));
 
-llvm::cl::opt<size_t> dcpCriticalWindowSize(
-  "dcp-critical-window-size",
-  llvm::cl::desc("Number of lowest-slack virtual nodes considered by each DCP coarsening iteration. "
-                 "Use 0 to run the legacy full-graph DCP analysis. Only used by the DCP scheduler."),
-  llvm::cl::init(4000));
-
 llvm::cl::opt<bool>
   ignoreConcatError("ignore-concat-error",
                     llvm::cl::desc("Ignore ConcatOp corner case: do not assert and do a simplification"),

src/PIM/Compiler/PimCompilerOptions.hpp

@@ -22,7 +22,6 @@ typedef enum {
 
 typedef enum {
   MergeSchedulerPeft = 0,
-  MergeSchedulerDcp = 1,
 } PimMergeSchedulerType;
 
 extern llvm::cl::OptionCategory OnnxMlirOptions;
@@ -36,7 +35,6 @@ extern llvm::cl::opt<bool> pimEmitJson;
 extern llvm::cl::opt<size_t> crossbarSize;
 extern llvm::cl::opt<size_t> crossbarCountInCore;
 extern llvm::cl::opt<long> coresCount;
-extern llvm::cl::opt<size_t> dcpCriticalWindowSize;
 
 bool hasExplicitPimCoreCount();
 void verifyExplicitPimCoreCount();

src/PIM/Compiler/PimCompilerUtils.cpp

@@ -30,20 +30,17 @@ void addPassesPim(OwningOpRef<ModuleOp>& module,
   if (pimEmissionTarget >= EmitSpatial) {
     pm.addPass(createONNXToSpatialPass());
     pm.addPass(createMergeComputeNodesPass());
-    // pm.addPass(createCountInstructionPass());
     pm.addPass(createMessagePass("Onnx lowered to Spatial"));
   }
 
   if (pimEmissionTarget >= EmitPim) {
     pm.addPass(createSpatialToPimPass());
-    // pm.addPass(createCountInstructionPass());
     pm.addPass(createMessagePass("Spatial lowered to Pim"));
   }
 
   if (pimEmissionTarget >= EmitPimBufferized) {
     pm.addPass(createPimBufferizationPass());
     pm.addPass(createPimStaticMemoryCoalescingPass());
-    // pm.addPass(createCountInstructionPass());
     pm.addPass(createMessagePass("Pim bufferized"));
   }
 
@@ -54,7 +51,6 @@ void addPassesPim(OwningOpRef<ModuleOp>& module,
     pm.addPass(createPimVerificationPass());
     pm.addPass(createMessagePass("Pim verified"));
     pm.addPass(createEmitPimCodePass());
-    // pm.addPass(createCountInstructionPass());
     pm.addPass(createMessagePass("Pim code emitted"));
   }
 }

src/PIM/Compiler/PimWeightEmitter.cpp

@@ -1,9 +1,7 @@
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/IR/BuiltinAttributes.h"
-#include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/BuiltinTypes.h"
 
-#include "llvm/ADT/STLExtras.h"
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/raw_ostream.h"
 
@@ -11,196 +9,42 @@
 
 #include "Conversion/ONNXToSpatial/Common/Common.hpp"
 #include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"
-#include "src/Accelerators/PIM/Common/IR/SubviewUtils.hpp"
-#include "src/Accelerators/PIM/Common/IR/WeightUtils.hpp"
-#include "src/Accelerators/PIM/Compiler/PimBatchEmission.hpp"
-#include "src/Accelerators/PIM/Compiler/PimCodeGen.hpp"
 #include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
 #include "src/Accelerators/PIM/Compiler/PimWeightEmitter.hpp"
-#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
 using namespace llvm;
 using namespace mlir;
 
 namespace onnx_mlir {
-namespace {
+namespace {} // namespace
 
-struct DenseWeightView {
+llvm::DenseMap<size_t, llvm::SmallVector<std::string, 8>>
-  DenseElementsAttr denseAttr;
+createAndPopulateWeightFolder(ArrayRef<WeightFileRequest> requests, StringRef outputDirPath) {
-  SmallVector<int64_t> shape;
-  SmallVector<int64_t> strides;
-  int64_t offset = 0;
-};
-
-FailureOr<DenseWeightView> resolveDenseWeightView(ModuleOp moduleOp, mlir::Value weight) {
-  SmallVector<Operation*> viewOps;
-  mlir::Value current = weight;
-  memref::GetGlobalOp getGlobalOp;
-
-  while (true) {
-    Operation* defOp = current.getDefiningOp();
-    if (!defOp)
-      return failure();
-    if ((getGlobalOp = dyn_cast<memref::GetGlobalOp>(defOp)))
-      break;
-    if (auto subview = dyn_cast<memref::SubViewOp>(defOp)) {
-      if (!hasAllStaticSubviewParts(subview))
-        return failure();
-      viewOps.push_back(subview);
-      current = subview.getSource();
-      continue;
-    }
-    if (auto cast = dyn_cast<memref::CastOp>(defOp)) {
-      current = cast.getSource();
-      continue;
-    }
-    if (auto collapse = dyn_cast<memref::CollapseShapeOp>(defOp)) {
-      auto srcType = dyn_cast<MemRefType>(collapse.getSrc().getType());
-      auto resultType = dyn_cast<MemRefType>(collapse.getResult().getType());
-      if (!srcType || !resultType || !srcType.hasStaticShape() || !resultType.hasStaticShape())
-        return failure();
-      viewOps.push_back(collapse);
-      current = collapse.getSrc();
-      continue;
-    }
-    if (auto expand = dyn_cast<memref::ExpandShapeOp>(defOp)) {
-      auto srcType = dyn_cast<MemRefType>(expand.getSrc().getType());
-      auto resultType = dyn_cast<MemRefType>(expand.getResult().getType());
-      if (!srcType || !resultType || !srcType.hasStaticShape() || !resultType.hasStaticShape())
-        return failure();
-      viewOps.push_back(expand);
-      current = expand.getSrc();
-      continue;
-    }
-    return failure();
-  }
-
-  auto globalOp = lookupGlobalForGetGlobal(moduleOp, getGlobalOp);
-  if (!globalOp || !globalOp.getInitialValue())
-    return failure();
-
-  auto denseAttr = dyn_cast<DenseElementsAttr>(*globalOp.getInitialValue());
-  if (!denseAttr)
-    return failure();
-
-  DenseWeightView view;
-  view.denseAttr = denseAttr;
-  view.shape.assign(denseAttr.getType().getShape().begin(), denseAttr.getType().getShape().end());
-  view.strides = computeRowMajorStrides(view.shape);
-
-  for (Operation* viewOp : llvm::reverse(viewOps)) {
-    if (auto subview = dyn_cast<memref::SubViewOp>(viewOp)) {
-      SmallVector<int64_t> nextStrides;
-      nextStrides.reserve(subview.getStaticStrides().size());
-      for (auto [offset, stride, sourceStride] :
-           llvm::zip_equal(subview.getStaticOffsets(), subview.getStaticStrides(), view.strides)) {
-        view.offset += offset * sourceStride;
-        nextStrides.push_back(stride * sourceStride);
-      }
-      view.shape.assign(subview.getStaticSizes().begin(), subview.getStaticSizes().end());
-      view.strides = std::move(nextStrides);
-      continue;
-    }
-
-    // Collapse/expand are accepted only as contiguous static reshapes of a
-    // dense global view, so a row-major stride recomputation preserves layout.
-    if (auto collapse = dyn_cast<memref::CollapseShapeOp>(viewOp)) {
-      if (view.strides != computeRowMajorStrides(view.shape))
-        return failure();
-      auto resultType = cast<MemRefType>(collapse.getResult().getType());
-      view.shape.assign(resultType.getShape().begin(), resultType.getShape().end());
-      view.strides = computeRowMajorStrides(view.shape);
-      continue;
-    }
-
-    if (auto expand = dyn_cast<memref::ExpandShapeOp>(viewOp)) {
-      if (view.strides != computeRowMajorStrides(view.shape))
-        return failure();
-      auto resultType = cast<MemRefType>(expand.getResult().getType());
-      view.shape.assign(resultType.getShape().begin(), resultType.getShape().end());
-      view.strides = computeRowMajorStrides(view.shape);
-      continue;
-    }
-  }
-
-  return view;
-}
-
-SmallVector<unsigned, 8> getUsedWeightIndices(Block& block) {
-  SmallVector<unsigned, 8> indices;
-  auto addIndex = [&](unsigned weightIndex) {
-    if (!llvm::is_contained(indices, weightIndex))
-      indices.push_back(weightIndex);
-  };
-
-  block.walk([&](pim::PimVMMOp vmmOp) { addIndex(vmmOp.getWeightIndex()); });
-  llvm::sort(indices);
-  return indices;
-}
-
-SmallVector<unsigned, 8> getUsedWeightIndices(pim::PimCoreOp coreOp) {
-  return getUsedWeightIndices(coreOp.getBody().front());
-}
-
-SmallVector<Operation*> collectTopLevelCoreLikeOps(func::FuncOp funcOp) {
-  SmallVector<Operation*> coreLikeOps;
-  for (Operation& op : funcOp.getBody().front())
-    if (dyn_cast<pim::PimCoreOp>(&op) || dyn_cast<pim::PimCoreBatchOp>(&op))
-      coreLikeOps.push_back(&op);
-  return coreLikeOps;
-}
-
-} // namespace
-
-llvm::DenseMap<size_t, llvm::DenseMap<mlir::Value, std::string>>
-createAndPopulateWeightFolder(func::FuncOp funcOp, StringRef outputDirPath) {
-  ModuleOp moduleOp = funcOp->getParentOfType<ModuleOp>();
   auto coreWeightsDirPath = outputDirPath + "/weights";
   auto error = sys::fs::create_directory(coreWeightsDirPath);
   assert(!error && "Error creating weights directory");
   size_t indexFileName = 0;
 
   int64_t xbarSize = crossbarSize.getValue();
-  llvm::DenseMap<size_t, llvm::DenseMap<mlir::Value, std::string>> mapCoreWeightToFileName;
+  llvm::DenseMap<size_t, llvm::SmallVector<std::string, 8>> mapCoreWeightToFileName;
-  llvm::DenseMap<memref::GlobalOp, std::string> mapGlobalOpToFileName;
+  llvm::SmallVector<std::pair<ResolvedWeightView, std::string>, 16> materializedWeights;
 
-  SmallVector<Operation*> coreLikeOps = collectTopLevelCoreLikeOps(funcOp);
+  auto materializeWeight = [&](const ResolvedWeightView& weightView) -> std::string {
+    if (auto it = llvm::find_if(materializedWeights, [&](const auto& entry) { return entry.first == weightView; });
+        it != materializedWeights.end())
+      return it->second;
 
-  for (Operation* op : coreLikeOps) {
+    auto globalOp = weightView.globalOp;
-    auto processCore = [&](pim::PimCoreOp coreOp) {
+    auto denseAttr = mlir::dyn_cast<DenseElementsAttr>(*globalOp.getInitialValue());
-      size_t coreId = static_cast<size_t>(coreOp.getCoreId());
+    assert(denseAttr && "Weight global must have dense initial value");
-      for (unsigned index : getUsedWeightIndices(coreOp)) {
-        if (index >= coreOp.getWeights().size()) {
-          coreOp.emitWarning("Weight index " + std::to_string(index) + " is out of range");
-          assert(index < coreOp.getWeights().size() && "Weight index is out of range");
-        }
-        mlir::Value weight = coreOp.getWeights()[index];
 
-        auto weightView = resolveDenseWeightView(moduleOp, weight);
+    ArrayRef<int64_t> shape = weightView.shape;
-        if (failed(weightView)) {
-          coreOp.emitWarning("Weight is not from a memref.get_global at index " + std::to_string(index));
-          assert(succeeded(weightView) && "Weight is not from a dense memref.global view");
-        }
-
-        if (mapCoreWeightToFileName[coreId].contains(weight))
-          continue;
-
-        auto getGlobalOp = weight.getDefiningOp<memref::GetGlobalOp>();
-        auto globalOp = getGlobalOp ? lookupGlobalForGetGlobal(moduleOp, getGlobalOp) : memref::GlobalOp {};
-        if (globalOp && mapGlobalOpToFileName.contains(globalOp)) {
-          auto& fileName = mapGlobalOpToFileName[globalOp];
-          mapCoreWeightToFileName[coreId].insert({weight, fileName});
-          continue;
-        }
-
-        DenseElementsAttr denseAttr = weightView->denseAttr;
-        ArrayRef<int64_t> shape = weightView->shape;
     assert(isMatrixShape(shape) && "Weight matrix must be 2-dimensional");
     int64_t numRows = shape[0];
     int64_t numCols = shape[1];
     assert(numRows <= xbarSize && numCols <= xbarSize && "Weight dimensions must not exceed crossbar size");
 
-        size_t elementByteWidth = denseAttr.getElementType().getIntOrFloatBitWidth() / 8;
+    size_t elementByteWidth = getElementTypeSizeInBytes(denseAttr.getElementType());
 
     std::string newFileName = "crossbar_" + std::to_string(indexFileName++) + ".bin";
     auto weightFilePath = (coreWeightsDirPath + "/" + newFileName).str();
@@ -215,7 +59,7 @@ createAndPopulateWeightFolder(func::FuncOp funcOp, StringRef outputDirPath) {
     for (int64_t row = 0; row < xbarSize; row++) {
       for (int64_t col = 0; col < xbarSize; col++) {
         if (row < numRows && col < numCols) {
-              int64_t elementIndex = weightView->offset + row * weightView->strides[0] + col * weightView->strides[1];
+          int64_t elementIndex = weightView.offset + row * weightView.strides[0] + col * weightView.strides[1];
           APInt bits = denseAttr.getValues<APFloat>()[elementIndex].bitcastToAPInt();
           uint64_t word = bits.getZExtValue();
           weightFileStream.write(reinterpret_cast<const char*>(&word), elementByteWidth);
@@ -227,23 +71,17 @@ createAndPopulateWeightFolder(func::FuncOp funcOp, StringRef outputDirPath) {
     }
 
     weightFileStream.close();
-        if (globalOp)
+    materializedWeights.push_back({weightView, newFileName});
-          mapGlobalOpToFileName.insert({globalOp, newFileName});
+    return newFileName;
-        mapCoreWeightToFileName[coreId].insert({weight, newFileName});
-      }
-      return success();
   };
 
-    if (auto coreOp = dyn_cast<pim::PimCoreOp>(op)) {
+  for (const WeightFileRequest& request : requests) {
-      (void) processCore(coreOp);
+    auto& coreFiles = mapCoreWeightToFileName[request.coreId];
-      continue;
+    coreFiles.reserve(request.weights.size());
+    for (const ResolvedWeightView& weight : request.weights)
+      coreFiles.push_back(materializeWeight(weight));
   }
 
-    auto coreBatchOp = cast<pim::PimCoreBatchOp>(op);
-    for (unsigned lane = 0; lane < static_cast<unsigned>(coreBatchOp.getLaneCount()); ++lane)
-      if (failed(withScalarCoreFromBatchLane(coreBatchOp, lane, processCore)))
-        return mapCoreWeightToFileName;
-  }
   return mapCoreWeightToFileName;
 }
 

src/PIM/Compiler/PimWeightEmitter.hpp

@@ -1,16 +1,23 @@
 #pragma once
 
 #include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/IR/Value.h"
 
 #include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 
 #include <string>
 
+#include "src/Accelerators/PIM/Common/IR/WeightUtils.hpp"
+
 namespace onnx_mlir {
 
-llvm::DenseMap<size_t, llvm::DenseMap<mlir::Value, std::string>>
+struct WeightFileRequest {
-createAndPopulateWeightFolder(mlir::func::FuncOp funcOp, llvm::StringRef outputDirPath);
+  size_t coreId = 0;
+  llvm::SmallVector<ResolvedWeightView, 8> weights;
+};
+
+llvm::DenseMap<size_t, llvm::SmallVector<std::string, 8>>
+createAndPopulateWeightFolder(llvm::ArrayRef<WeightFileRequest> requests, llvm::StringRef outputDirPath);
 
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/CMakeLists.txt

@@ -3,11 +3,12 @@ mlir_tablegen(ONNXToSpatial.hpp.inc -gen-rewriters "-I${ONNX_MLIR_SRC_ROOT}")
 add_public_tablegen_target(ONNXToSpatialIncGen)
 
 add_pim_library(OMONNXToSpatial
-  ConversionPatterns.cpp
+  Patterns.cpp
-  HostFoldability.cpp
+  CompileTime.cpp
-  HostLegality.cpp
+  ONNXToSpatialVerifier.cpp
-  PrePatterns.cpp
+  Patterns/Pre.cpp
-  PostPatterns.cpp
+  Patterns/Post.cpp
+  Patterns/GeneratedConversion.cpp
   Patterns/Math/Conv.cpp
   Patterns/Math/Elementwise.cpp
   Patterns/Math/Gemm.cpp
@@ -22,6 +23,7 @@ add_pim_library(OMONNXToSpatial
   Patterns/Tensor/Resize.cpp
   Patterns/Tensor/Reshape.cpp
   Patterns/Tensor/Split.cpp
+  Patterns/Tensor/Transpose.cpp
   ONNXToSpatialPass.cpp
   Common/ComputeRegionBuilder.cpp
   Common/ShapeTilingUtils.cpp
@@ -33,6 +35,7 @@ add_pim_library(OMONNXToSpatial
   ONNXToSpatialIncGen
 
   LINK_LIBS PUBLIC
+  MLIRLinalgDialect
   MLIRSCFDialect
   MLIRTosaDialect
   OMCompilerOptions

src/PIM/Conversion/ONNXToSpatial/Common/ComputeRegionBuilder.hpp

@@ -18,13 +18,17 @@ namespace detail {
 
 inline mlir::ValueRange getBlockArgs(mlir::Block* block) { return mlir::ValueRange(block->getArguments()); }
 
+inline mlir::ValueRange getInputBlockArgs(mlir::Block* block, size_t weightCount) {
+  return mlir::ValueRange(block->getArguments()).drop_front(weightCount);
+}
+
 template <typename Fn, size_t... Is>
 decltype(auto) invokeWithBlockArgs(Fn&& fn, mlir::Block* block, std::index_sequence<Is...>) {
   return std::forward<Fn>(fn)(block->getArgument(Is)...);
 }
 
 template <typename Fn, size_t... Is>
-decltype(auto) invokeWithValues(Fn&& fn, mlir::ArrayRef<mlir::Value> values, std::index_sequence<Is...>) {
+decltype(auto) invokeWithValues(Fn&& fn, mlir::ValueRange values, std::index_sequence<Is...>) {
   return std::forward<Fn>(fn)(values[Is]...);
 }
 
@@ -85,6 +89,8 @@ auto createSpatCompute(RewriterT& rewriter,
   auto computeOp = spatial::SpatCompute::create(rewriter, loc, resultTypes, weights, inputs);
 
   auto* block = new mlir::Block();
+  for (mlir::Value weight : weights)
+    block->addArgument(weight.getType(), loc);
   for (mlir::Value input : inputs)
     block->addArgument(input.getType(), loc);
 
@@ -93,14 +99,17 @@ auto createSpatCompute(RewriterT& rewriter,
 
   using BodyResult = detail::InvokeWithBlockArgsResultT<std::decay_t<BodyFn>, std::make_index_sequence<NumInputs>>;
   if constexpr (std::is_same_v<BodyResult, void>) {
-    detail::invokeWithBlockArgs(std::forward<BodyFn>(body), block, std::make_index_sequence<NumInputs> {});
+    detail::invokeWithValues(std::forward<BodyFn>(body),
+                             detail::getInputBlockArgs(block, weights.size()),
+                             std::make_index_sequence<NumInputs> {});
 
     rewriter.setInsertionPointAfter(computeOp);
     return computeOp;
   }
   else {
-    auto bodyResult =
+    auto bodyResult = detail::invokeWithValues(std::forward<BodyFn>(body),
-      detail::invokeWithBlockArgs(std::forward<BodyFn>(body), block, std::make_index_sequence<NumInputs> {});
+                                               detail::getInputBlockArgs(block, weights.size()),
+                                               std::make_index_sequence<NumInputs> {});
     if (mlir::failed(bodyResult)) {
       rewriter.setInsertionPointAfter(computeOp);
       rewriter.eraseOp(computeOp);
@@ -123,6 +132,8 @@ auto createSpatCompute(RewriterT& rewriter,
   auto computeOp = spatial::SpatCompute::create(rewriter, loc, resultTypes, weights, inputs);
 
   auto* block = new mlir::Block();
+  for (mlir::Value weight : weights)
+    block->addArgument(weight.getType(), loc);
   for (mlir::Value input : inputs)
     block->addArgument(input.getType(), loc);
 
@@ -131,13 +142,13 @@ auto createSpatCompute(RewriterT& rewriter,
 
   using BodyResult = detail::InvokeWithValueRangeResultT<std::decay_t<BodyFn>>;
   if constexpr (std::is_same_v<BodyResult, void>) {
-    std::forward<BodyFn>(body)(detail::getBlockArgs(block));
+    std::forward<BodyFn>(body)(detail::getInputBlockArgs(block, weights.size()));
 
     rewriter.setInsertionPointAfter(computeOp);
     return computeOp;
   }
   else {
-    auto bodyResult = std::forward<BodyFn>(body)(detail::getBlockArgs(block));
+    auto bodyResult = std::forward<BodyFn>(body)(detail::getInputBlockArgs(block, weights.size()));
     if (mlir::failed(bodyResult)) {
       rewriter.setInsertionPointAfter(computeOp);
       rewriter.eraseOp(computeOp);

src/PIM/Conversion/ONNXToSpatial/Common/ShapeTilingUtils.cpp

@@ -1,17 +1,87 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/IR/Matchers.h"
 
 #include "llvm/ADT/SmallVector.h"
 
+#include <algorithm>
+
 #include "ShapeTilingUtils.hpp"
 #include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
 
 using namespace mlir;
 
 namespace onnx_mlir {
 
+static Value getIndexValue(OpFoldResult result, ConversionPatternRewriter& rewriter, Location loc) {
+  if (auto attr = dyn_cast<Attribute>(result))
+    return arith::ConstantIndexOp::create(rewriter, loc, cast<IntegerAttr>(attr).getInt()).getResult();
+  return cast<Value>(result);
+}
+
+static Value addIndexValues(Value lhs, Value rhs, ConversionPatternRewriter& rewriter, Location loc) {
+  APInt lhsConst;
+  if (matchPattern(lhs, m_ConstantInt(&lhsConst)) && lhsConst.isZero())
+    return rhs;
+
+  APInt rhsConst;
+  if (matchPattern(rhs, m_ConstantInt(&rhsConst)) && rhsConst.isZero())
+    return lhs;
+
+  return arith::AddIOp::create(rewriter, loc, lhs, rhs).getResult();
+}
+
+static Value multiplyIndexValue(Value value, OpFoldResult factor, ConversionPatternRewriter& rewriter, Location loc) {
+  APInt factorConst;
+  if (auto attr = dyn_cast<Attribute>(factor))
+    factorConst = cast<IntegerAttr>(attr).getValue();
+  else if (!matchPattern(cast<Value>(factor), m_ConstantInt(&factorConst)))
+    return arith::MulIOp::create(rewriter, loc, value, cast<Value>(factor)).getResult();
+
+  if (factorConst.isZero())
+    return arith::ConstantIndexOp::create(rewriter, loc, 0).getResult();
+  if (factorConst.isOne())
+    return value;
+
+  auto factorValue = arith::ConstantIndexOp::create(rewriter, loc, factorConst.getSExtValue()).getResult();
+  return arith::MulIOp::create(rewriter, loc, value, factorValue).getResult();
+}
+
+static bool isContiguousTensorSlice(Value source, RankedTensorType resultType, ArrayRef<OpFoldResult> strides) {
+  auto sourceType = dyn_cast<RankedTensorType>(source.getType());
+  if (!sourceType || !sourceType.hasStaticShape() || !resultType.hasStaticShape() || sourceType.getRank() != resultType.getRank())
+    return false;
+
+  for (OpFoldResult stride : strides) {
+    APInt strideValue;
+    if (auto attr = dyn_cast<Attribute>(stride)) {
+      if (cast<IntegerAttr>(attr).getInt() != 1)
+        return false;
+      continue;
+    }
+    if (!matchPattern(cast<Value>(stride), m_ConstantInt(&strideValue)) || !strideValue.isOne())
+      return false;
+  }
+
+  auto sizesAndShape = llvm::zip_equal(llvm::make_range(resultType.getShape().rbegin(), resultType.getShape().rend()),
+                                       llvm::make_range(sourceType.getShape().rbegin(), sourceType.getShape().rend()));
+  auto firstDifferentSize = std::find_if(sizesAndShape.begin(), sizesAndShape.end(), [&](auto sizeAndShape) -> bool {
+    auto [size, dimension] = sizeAndShape;
+    return size != dimension;
+  });
+  if (firstDifferentSize == sizesAndShape.end())
+    return true;
+
+  ++firstDifferentSize;
+  return std::all_of(firstDifferentSize, sizesAndShape.end(), [](auto sizeAndShape) {
+    auto [size, _dimension] = sizeAndShape;
+    return size == 1;
+  });
+}
+
 SmallVector<Value> sliceTensor(
   const Value& tensorToSlice, size_t axis, int64_t sliceSize, ConversionPatternRewriter& rewriter, Location loc) {
   ArrayRef<long> shape = getTensorShape(tensorToSlice);
@@ -44,7 +114,7 @@ SmallVector<Value> sliceTensor(
       RankedTensorType::get(sliceShape, cast<RankedTensorType>(tensorToSlice.getType()).getElementType());
 
     Value slice;
-    if (isHostFoldableValue(tensorToSlice)) {
+    if (isCompileTimeComputable(tensorToSlice)) {
       slice = tensor::ExtractSliceOp::create(rewriter, loc, tensorToSlice, offsets, sizes, strides);
     }
     else {
@@ -113,7 +183,7 @@ Value broadcastToVector(Value scalarToBroadcast, int64_t length, ConversionPatte
     return tensor::SplatOp::create(rewriter, loc, type, elementValue);
   };
 
-  if (isHostFoldableValue(scalarToBroadcast))
+  if (isCompileTimeComputable(scalarToBroadcast))
     return buildBroadcast(scalarToBroadcast);
 
   auto broadcastCompute =
@@ -123,4 +193,87 @@ Value broadcastToVector(Value scalarToBroadcast, int64_t length, ConversionPatte
   return broadcastCompute.getResult(0);
 }
 
+Value materializeContiguousTensorSlice(Value source,
+                                       RankedTensorType resultType,
+                                       ArrayRef<OpFoldResult> offsets,
+                                       ArrayRef<OpFoldResult> strides,
+                                       ConversionPatternRewriter& rewriter,
+                                       Location loc) {
+  assert(resultType.hasStaticShape() && "expected static result type");
+  size_t rank = static_cast<size_t>(resultType.getRank());
+  assert(offsets.size() == rank && "expected rank-matching offsets");
+  assert(strides.size() == rank && "expected rank-matching strides");
+
+  SmallVector<OpFoldResult> sizes;
+  sizes.reserve(resultType.getRank());
+  for (int64_t size : resultType.getShape())
+    sizes.push_back(rewriter.getIndexAttr(size));
+
+  if (isContiguousTensorSlice(source, resultType, strides))
+    return tensor::ExtractSliceOp::create(rewriter, loc, resultType, source, offsets, sizes, strides).getResult();
+
+  if (resultType.getRank() == 0)
+    return tensor::ExtractSliceOp::create(rewriter, loc, resultType, source, offsets, sizes, strides).getResult();
+
+  Value init = tensor::EmptyOp::create(rewriter, loc, resultType.getShape(), resultType.getElementType()).getResult();
+  SmallVector<Value> zeroIndices(resultType.getRank());
+  for (Value& zeroIndex : zeroIndices)
+    zeroIndex = arith::ConstantIndexOp::create(rewriter, loc, 0).getResult();
+
+  SmallVector<Value> resultIndices;
+  resultIndices.reserve(resultType.getRank());
+
+  auto buildLoopNest = [&](auto&& self, unsigned dim, Value accumulator) -> Value {
+    if (dim == resultType.getRank()) {
+      SmallVector<Value> sourceIndices;
+      sourceIndices.reserve(resultType.getRank());
+      for (unsigned idx = 0; idx < resultType.getRank(); ++idx) {
+        Value offsetValue = getIndexValue(offsets[idx], rewriter, loc);
+        Value scaledIndex = multiplyIndexValue(resultIndices[idx], strides[idx], rewriter, loc);
+        sourceIndices.push_back(addIndexValues(offsetValue, scaledIndex, rewriter, loc));
+      }
+
+      SmallVector<OpFoldResult> sourceOffsets;
+      SmallVector<OpFoldResult> destinationOffsets;
+      SmallVector<OpFoldResult> unitSizes;
+      SmallVector<OpFoldResult> unitStrides;
+      sourceOffsets.reserve(resultType.getRank());
+      destinationOffsets.reserve(resultType.getRank());
+      unitSizes.reserve(resultType.getRank());
+      unitStrides.reserve(resultType.getRank());
+      for (Value index : sourceIndices)
+        sourceOffsets.push_back(index);
+      for (Value index : resultIndices)
+        destinationOffsets.push_back(index);
+      for (int64_t idx = 0; idx < resultType.getRank(); ++idx) {
+        unitSizes.push_back(rewriter.getIndexAttr(1));
+        unitStrides.push_back(rewriter.getIndexAttr(1));
+      }
+
+      auto elementTensorType =
+        RankedTensorType::get(SmallVector<int64_t>(resultType.getRank(), 1), resultType.getElementType());
+      Value elementSlice =
+        tensor::ExtractSliceOp::create(rewriter, loc, elementTensorType, source, sourceOffsets, unitSizes, unitStrides)
+          .getResult();
+      return tensor::InsertSliceOp::create(
+               rewriter, loc, elementSlice, accumulator, destinationOffsets, unitSizes, unitStrides)
+        .getResult();
+    }
+
+    Value lower = zeroIndices[dim];
+    Value upper = arith::ConstantIndexOp::create(rewriter, loc, resultType.getDimSize(dim)).getResult();
+    Value step = arith::ConstantIndexOp::create(rewriter, loc, 1).getResult();
+    auto loop = scf::ForOp::create(rewriter, loc, lower, upper, step, ValueRange {accumulator});
+    rewriter.setInsertionPointToStart(loop.getBody());
+    resultIndices.push_back(loop.getInductionVar());
+    Value updated = self(self, dim + 1, loop.getRegionIterArgs().front());
+    resultIndices.pop_back();
+    scf::YieldOp::create(rewriter, loc, updated);
+    rewriter.setInsertionPointAfter(loop);
+    return loop.getResult(0);
+  };
+
+  return buildLoopNest(buildLoopNest, 0, init);
+}
+
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Common/ShapeTilingUtils.hpp

@@ -141,4 +141,11 @@ mlir::Value broadcastToVector(mlir::Value scalarToBroadcast,
                               mlir::ConversionPatternRewriter& rewriter,
                               mlir::Location loc);
 
+mlir::Value materializeContiguousTensorSlice(mlir::Value source,
+                                             mlir::RankedTensorType resultType,
+                                             llvm::ArrayRef<mlir::OpFoldResult> offsets,
+                                             llvm::ArrayRef<mlir::OpFoldResult> strides,
+                                             mlir::ConversionPatternRewriter& rewriter,
+                                             mlir::Location loc);
+
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Common/WeightMaterialization.cpp

@@ -1,4 +1,5 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/IRMapping.h"
@@ -43,8 +44,8 @@ bool isWeightLikeComputeOperand(Value value) {
       value = collapseShapeOp.getSrc();
       continue;
     }
-    if (auto transposeOp = dyn_cast<ONNXTransposeOp>(definingOp)) {
+    if (auto transposeOp = dyn_cast<linalg::TransposeOp>(definingOp)) {
-      value = transposeOp.getData();
+      value = transposeOp.getInput();
       continue;
     }
 
@@ -80,7 +81,7 @@ FailureOr<Value> materializeWeightLikeValueInBlock(Value value, IRRewriter& rewr
     return referencedValue.getResult();
   }
 
-  if (!isa<tensor::ExtractSliceOp, tensor::ExpandShapeOp, tensor::CollapseShapeOp, ONNXTransposeOp>(definingOp))
+  if (!isa<tensor::ExtractSliceOp, tensor::ExpandShapeOp, tensor::CollapseShapeOp, linalg::TransposeOp>(definingOp))
     return failure();
 
   IRMapping localMapper;

src/PIM/Conversion/ONNXToSpatial/CompileTime.cpp

@@ -1,4 +1,5 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/IR/BuiltinTypes.h"
@@ -7,8 +8,11 @@
 #include "llvm/ADT/SmallBitVector.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/ErrorHandling.h"
 
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include <utility>
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -145,7 +149,7 @@ static DenseElementsAttr getDirectDenseConstantAttr(Value value) {
   return nullptr;
 }
 
-static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm::SmallPtrSetImpl<Operation*>& visited) {
+static DenseElementsAttr getHostConstantDenseElementsAttrImpl(Value value, llvm::SmallPtrSetImpl<Operation*>& visited) {
   auto* definingOp = value.getDefiningOp();
   if (!definingOp || !visited.insert(definingOp).second)
     return nullptr;
@@ -156,7 +160,7 @@ static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm:
     return denseAttr;
 
   if (auto transposeOp = dyn_cast<ONNXTransposeOp>(definingOp)) {
-    auto inputAttr = getHostFoldableDenseElementsAttrImpl(transposeOp.getData(), visited);
+    auto inputAttr = getHostConstantDenseElementsAttrImpl(transposeOp.getData(), visited);
     if (!inputAttr)
       return nullptr;
 
@@ -168,8 +172,18 @@ static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm:
     return succeeded(transposedAttr) ? *transposedAttr : nullptr;
   }
 
+  if (auto transposeOp = dyn_cast<linalg::TransposeOp>(definingOp)) {
+    auto inputAttr = getHostConstantDenseElementsAttrImpl(transposeOp.getInput(), visited);
+    if (!inputAttr)
+      return nullptr;
+
+    SmallVector<int64_t> perm(transposeOp.getPermutation().begin(), transposeOp.getPermutation().end());
+    auto transposedAttr = transposeDenseElements(inputAttr, perm);
+    return succeeded(transposedAttr) ? *transposedAttr : nullptr;
+  }
+
   if (auto collapseShapeOp = dyn_cast<tensor::CollapseShapeOp>(definingOp)) {
-    auto inputAttr = getHostFoldableDenseElementsAttrImpl(collapseShapeOp.getSrc(), visited);
+    auto inputAttr = getHostConstantDenseElementsAttrImpl(collapseShapeOp.getSrc(), visited);
     if (!inputAttr)
       return nullptr;
     auto reshapedAttr = reshapeDenseElements(inputAttr, cast<RankedTensorType>(collapseShapeOp.getType()));
@@ -177,7 +191,7 @@ static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm:
   }
 
   if (auto expandShapeOp = dyn_cast<tensor::ExpandShapeOp>(definingOp)) {
-    auto inputAttr = getHostFoldableDenseElementsAttrImpl(expandShapeOp.getSrc(), visited);
+    auto inputAttr = getHostConstantDenseElementsAttrImpl(expandShapeOp.getSrc(), visited);
     if (!inputAttr)
       return nullptr;
     auto reshapedAttr = reshapeDenseElements(inputAttr, cast<RankedTensorType>(expandShapeOp.getType()));
@@ -185,7 +199,7 @@ static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm:
   }
 
   if (auto extractSliceOp = dyn_cast<tensor::ExtractSliceOp>(definingOp)) {
-    auto inputAttr = getHostFoldableDenseElementsAttrImpl(extractSliceOp.getSource(), visited);
+    auto inputAttr = getHostConstantDenseElementsAttrImpl(extractSliceOp.getSource(), visited);
     if (!inputAttr)
       return nullptr;
     auto slicedAttr = extractSliceDenseElements(inputAttr, extractSliceOp);
@@ -195,62 +209,98 @@ static DenseElementsAttr getHostFoldableDenseElementsAttrImpl(Value value, llvm:
   return nullptr;
 }
 
-static bool isHostFoldableOpImpl(Operation* op, llvm::SmallPtrSetImpl<Operation*>& visited) {
+static std::optional<CompileTimeSource>
-  if (!op || !visited.insert(op).second)
+getCompileTimeSourceImpl(Operation* op, llvm::SmallPtrSetImpl<Operation*>& visited, size_t chainLength = 0) {
-    return false;
+  if (!op)
+    return std::nullopt;
+
+  if (!visited.insert(op).second)
+    return {
+      {op, chainLength}
+    };
 
   if (isa<arith::ConstantOp, ONNXConstantOp, ONNXNoneOp>(op))
-    return true;
+    return {
+      {op, chainLength}
+    };
+
+  chainLength += 1;
 
   if (auto extractOp = dyn_cast<tensor::ExtractOp>(op))
-    return hasConstantIndices(extractOp) && isHostFoldableValue(extractOp.getTensor());
+    return hasConstantIndices(extractOp)
+           ? getCompileTimeSourceImpl(extractOp.getTensor().getDefiningOp(), visited, chainLength)
+           : std::nullopt;
 
   if (!isStaticTensorResult(op))
-    return false;
+    return std::nullopt;
 
   if (auto transposeOp = dyn_cast<ONNXTransposeOp>(op))
-    return isHostFoldableValue(transposeOp.getData());
+    return getCompileTimeSourceImpl(transposeOp.getData().getDefiningOp(), visited, chainLength);
+
+  if (auto transposeOp = dyn_cast<linalg::TransposeOp>(op))
+    return getCompileTimeSourceImpl(transposeOp.getInput().getDefiningOp(), visited, chainLength);
 
   if (auto collapseShapeOp = dyn_cast<tensor::CollapseShapeOp>(op))
-    return isHostFoldableValue(collapseShapeOp.getSrc());
+    return getCompileTimeSourceImpl(collapseShapeOp.getSrc().getDefiningOp(), visited, chainLength);
 
   if (auto expandShapeOp = dyn_cast<tensor::ExpandShapeOp>(op))
-    return isHostFoldableValue(expandShapeOp.getSrc());
+    return getCompileTimeSourceImpl(expandShapeOp.getSrc().getDefiningOp(), visited, chainLength);
 
   if (auto extractSliceOp = dyn_cast<tensor::ExtractSliceOp>(op))
-    return hasStaticUnitStrides(extractSliceOp) && isHostFoldableValue(extractSliceOp.getSource());
+    return hasStaticUnitStrides(extractSliceOp)
+           ? getCompileTimeSourceImpl(extractSliceOp.getSource().getDefiningOp(), visited, chainLength)
+           : std::nullopt;
 
   if (auto splatOp = dyn_cast<tensor::SplatOp>(op))
-    return isHostFoldableValue(splatOp.getInput());
+    return getCompileTimeSourceImpl(splatOp.getInput().getDefiningOp(), visited, chainLength);
 
   if (auto extractRowsOp = dyn_cast<spatial::SpatExtractRowsOp>(op))
-    return isHostFoldableValue(extractRowsOp.getInput());
+    return getCompileTimeSourceImpl(extractRowsOp.getInput().getDefiningOp(), visited, chainLength);
 
-  if (auto concatOp = dyn_cast<spatial::SpatConcatOp>(op))
+  if (auto concatOp = dyn_cast<spatial::SpatConcatOp>(op)) {
-    return llvm::all_of(concatOp.getInputs(), isHostFoldableValue);
+    std::optional<CompileTimeSource> res = {};
+    for (auto operandValue : concatOp.getOperands()) {
+      auto partialRes = getCompileTimeSourceImpl(operandValue.getDefiningOp(), visited, chainLength);
+      if (!partialRes)
+        return std::nullopt;
 
-  return false;
+      if (!res) {
+        res = partialRes;
+        continue;
+      }
+      if (res->chainLength < partialRes->chainLength)
+        res = partialRes;
+    }
+    return res;
+  }
+
+  return std::nullopt;
 }
 
 } // namespace
 
-bool isHostFoldableValue(Value value) {
+std::optional<CompileTimeSource> getCompileTimeSource(Operation* op) {
+  llvm::SmallPtrSet<Operation*, 8> visited;
+  return getCompileTimeSourceImpl(op, visited);
+}
+
+bool isCompileTimeComputable(Value value) {
   auto* definingOp = value.getDefiningOp();
   if (!definingOp)
     return false;
 
   llvm::SmallPtrSet<Operation*, 8> visited;
-  return isHostFoldableOpImpl(definingOp, visited);
+  return getCompileTimeSourceImpl(definingOp, visited).has_value();
 }
 
-bool isHostFoldableOp(Operation* op) {
+bool isCompileTimeOp(Operation* op) {
   llvm::SmallPtrSet<Operation*, 8> visited;
-  return isHostFoldableOpImpl(op, visited);
+  return getCompileTimeSourceImpl(op, visited).has_value();
 }
 
-DenseElementsAttr getHostFoldableDenseElementsAttr(Value value) {
+DenseElementsAttr getHostConstDenseElementsAttr(Value value) {
   llvm::SmallPtrSet<Operation*, 8> visited;
-  return getHostFoldableDenseElementsAttrImpl(value, visited);
+  return getHostConstantDenseElementsAttrImpl(value, visited);
 }
 
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/CompileTime.hpp

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/Operation.h"
+#include "mlir/IR/Value.h"
+
+namespace onnx_mlir {
+
+struct CompileTimeSource {
+  mlir::Operation* source;
+  size_t chainLength;
+};
+
+std::optional<CompileTimeSource> getCompileTimeSource(mlir::Operation* op);
+
+bool isCompileTimeComputable(mlir::Value value);
+
+bool isCompileTimeOp(mlir::Operation* op);
+
+mlir::DenseElementsAttr getHostConstDenseElementsAttr(mlir::Value value);
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp

@@ -1,15 +0,0 @@
-#pragma once
-
-#include "mlir/IR/BuiltinAttributes.h"
-#include "mlir/IR/Operation.h"
-#include "mlir/IR/Value.h"
-
-namespace onnx_mlir {
-
-bool isHostFoldableValue(mlir::Value value);
-
-bool isHostFoldableOp(mlir::Operation* op);
-
-mlir::DenseElementsAttr getHostFoldableDenseElementsAttr(mlir::Value value);
-
-} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/HostLegality.cpp

@@ -1,34 +0,0 @@
-#include "mlir/Dialect/Arith/IR/Arith.h"
-#include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/Dialect/Tensor/IR/Tensor.h"
-
-#include "src/Accelerators/PIM/Common/Support/Diagnostics.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostLegality.hpp"
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-
-using namespace mlir;
-
-namespace onnx_mlir {
-
-LogicalResult verifyONNXToSpatialHostLegality(func::FuncOp funcOp) {
-  pim::CappedDiagnosticReporter diagnostics;
-
-  for (Operation& op : funcOp.getFunctionBody().front()) {
-    if (isa<func::ReturnOp, spatial::SpatCompute, spatial::SpatComputeBatch>(&op))
-      continue;
-    if (isHostFoldableOp(&op))
-      continue;
-
-    diagnostics.report(&op, [](Operation* illegalOp) {
-      illegalOp->emitOpError("non-foldable top-level runtime op remains after ONNX-to-Spatial; lower it inside "
-                             "spat.compute");
-    });
-  }
-
-  diagnostics.emitSuppressedSummary(funcOp, "ONNX-to-Spatial host legality failures");
-
-  return success(!diagnostics.hasFailure());
-}
-
-} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/HostLegality.hpp

@@ -1,10 +0,0 @@
-#pragma once
-
-#include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/Support/LogicalResult.h"
-
-namespace onnx_mlir {
-
-mlir::LogicalResult verifyONNXToSpatialHostLegality(mlir::func::FuncOp funcOp);
-
-} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/ONNXToSpatialPass.cpp

@@ -1,5 +1,7 @@
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/IRMapping.h"
@@ -12,13 +14,10 @@
 
 #include "Common/Common.hpp"
 #include "Common/PimCommon.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostLegality.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatialVerifier.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/PostPatterns.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/PrePatterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-#include "src/Compiler/CompilerOptions.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
 using namespace mlir;
@@ -44,7 +43,8 @@ static void populateEmptyFunction(func::FuncOp funcOp) {
   IRRewriter rewriter(funcOp.getContext());
   IRMapping mapper;
   SmallVector<spatial::SpatCompute> computes(funcOp.getOps<spatial::SpatCompute>());
-  if (!computes.empty())
+  SmallVector<spatial::SpatComputeBatch> computeBatches(funcOp.getOps<spatial::SpatComputeBatch>());
+  if (!computes.empty() || !computeBatches.empty())
     return;
 
   auto returnOp = cast<func::ReturnOp>(funcOp.getFunctionBody().front().getTerminator());
@@ -85,30 +85,6 @@ static void populateEmptyFunction(func::FuncOp funcOp) {
     returnOp.setOperand(index, computeResult);
 }
 
-static void wrapTopLevelRuntimeTransposes(func::FuncOp funcOp) {
-  IRRewriter rewriter(funcOp.getContext());
-  Block& entryBlock = funcOp.getFunctionBody().front();
-
-  for (Operation& op : llvm::make_early_inc_range(entryBlock)) {
-    auto transposeOp = dyn_cast<ONNXTransposeOp>(&op);
-    if (!transposeOp || isHostFoldableOp(transposeOp))
-      continue;
-
-    // Transpose stays globally legal because constant/view-only cases are
-    // allowed on the host. Any residual runtime transpose must be sunk into
-    // spat.compute before the host legality check.
-    auto resultType = transposeOp.getResult().getType();
-    rewriter.setInsertionPoint(transposeOp);
-    auto computeOp = createSpatCompute<1>(
-      rewriter, transposeOp.getLoc(), TypeRange {resultType}, {}, ValueRange {transposeOp.getData()}, [&](Value input) {
-        Value transposed =
-          ONNXTransposeOp::create(rewriter, transposeOp.getLoc(), resultType, input, transposeOp.getPermAttr());
-        spatial::SpatYieldOp::create(rewriter, transposeOp.getLoc(), transposed);
-      });
-    rewriter.replaceOp(transposeOp, computeOp.getResult(0));
-  }
-}
-
 void ONNXToSpatialPass::runOnOperation() {
   ModuleOp moduleOp = getOperation();
   MLIRContext* ctx = &getContext();
@@ -116,7 +92,9 @@ void ONNXToSpatialPass::runOnOperation() {
   ConversionTarget preTarget(*ctx);
   preTarget.addLegalDialect<spatial::SpatialDialect,
                             ONNXDialect,
+                            linalg::LinalgDialect,
                             tensor::TensorDialect,
+                            affine::AffineDialect,
                             arith::ArithDialect,
                             scf::SCFDialect>();
   preTarget.addIllegalOp<ONNXConstantOp, ONNXFlattenOp>();
@@ -154,10 +132,13 @@ void ONNXToSpatialPass::runOnOperation() {
   ConversionTarget target(*ctx);
   target.addLegalDialect<spatial::SpatialDialect,
                          ONNXDialect,
+                         linalg::LinalgDialect,
                          tensor::TensorDialect,
+                         affine::AffineDialect,
                          arith::ArithDialect,
                          scf::SCFDialect>();
   target.addIllegalOp<ONNXMatMulOp>();
+  target.addIllegalOp<ONNXTransposeOp>();
   target.addIllegalOp<ONNXAddOp>();
   target.addIllegalOp<ONNXDivOp>();
   target.addIllegalOp<ONNXMulOp>();
@@ -184,22 +165,18 @@ void ONNXToSpatialPass::runOnOperation() {
     return;
   }
 
+  RewritePatternSet transposePatterns(ctx);
+  populateTransposePatterns(transposePatterns, ctx);
+  walkAndApplyPatterns(moduleOp, std::move(transposePatterns));
+
   ConversionTarget earlyPostTarget(*ctx);
   earlyPostTarget.addLegalDialect<spatial::SpatialDialect,
                                   ONNXDialect,
+                                  linalg::LinalgDialect,
                                   tensor::TensorDialect,
+                                  affine::AffineDialect,
                                   arith::ArithDialect,
                                   scf::SCFDialect>();
-  earlyPostTarget.addDynamicallyLegalOp<spatial::SpatComputeBatch>(
-    [](spatial::SpatComputeBatch batchOp) { return !requiresEarlyPostRewrite(batchOp); });
-
-  RewritePatternSet earlyPostPatterns(ctx);
-  populateEarlyPostPatterns(earlyPostPatterns, ctx);
-  if (failed(applyPartialConversion(*entryFunc, earlyPostTarget, std::move(earlyPostPatterns)))) {
-    moduleOp.emitError("failed to normalize single-lane spat.compute_batch ops before core assignment checks");
-    signalPassFailure();
-    return;
-  }
 
   PassManager cleanupPM(ctx);
   cleanupPM.addPass(createCanonicalizerPass());
@@ -211,7 +188,9 @@ void ONNXToSpatialPass::runOnOperation() {
   ConversionTarget postTarget(*ctx);
   postTarget.addLegalDialect<spatial::SpatialDialect,
                              ONNXDialect,
+                             linalg::LinalgDialect,
                              tensor::TensorDialect,
+                             affine::AffineDialect,
                              arith::ArithDialect,
                              scf::SCFDialect>();
   postTarget.addDynamicallyLegalOp<spatial::SpatCompute>(
@@ -227,9 +206,7 @@ void ONNXToSpatialPass::runOnOperation() {
     return;
   }
 
-  wrapTopLevelRuntimeTransposes(*entryFunc);
+  if (failed(verifyONNXToSpatial(*entryFunc))) {
-
-  if (failed(verifyONNXToSpatialHostLegality(*entryFunc))) {
     moduleOp.emitError("ONNX-to-Spatial host legality verification failed");
     signalPassFailure();
     return;

src/PIM/Conversion/ONNXToSpatial/ONNXToSpatialVerifier.cpp

@@ -0,0 +1,157 @@
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Support/LLVM.h"
+
+#include "Common/IR/WeightUtils.hpp"
+#include "src/Accelerators/PIM/Common/Support/Diagnostics.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatialVerifier.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+
+namespace {
+
+void checkWeightUseChains(func::FuncOp func, pim::CappedDiagnosticReporter& diagnostics) {
+  func.walk([&](Operation* op) {
+    if (!hasWeightAlways(op))
+      return;
+
+    for (Value result : op->getResults()) {
+      if (hasOnlySpatialMvmVmmWeightUses(result))
+        continue;
+
+      diagnostics.report(op, [&](Operation* illegalOp) {
+        illegalOp->emitOpError(
+          "weight-marked values may only flow through static view/slice helper chains into Spatial VMM weights");
+      });
+      return;
+    }
+  });
+}
+
+Region* getParentRegion(Value value) {
+  if (auto blockArg = dyn_cast<BlockArgument>(value))
+    return blockArg.getOwner()->getParent();
+  if (Operation* definingOp = value.getDefiningOp())
+    return definingOp->getParentRegion();
+  return nullptr;
+}
+
+bool isDefinedInsideRegion(Value value, Region& region) {
+  Region* parentRegion = getParentRegion(value);
+  return parentRegion && (&region == parentRegion || region.isAncestor(parentRegion));
+}
+
+bool isLegalHostBackedValue(Value value) {
+  Operation* definingOp = value.getDefiningOp();
+  if (!definingOp)
+    return isa<BlockArgument>(value);
+
+  if (isa<spatial::SpatChannelReceiveOp>(definingOp))
+    return false;
+
+  return definingOp->getDialect()->getNamespace() != "spat";
+}
+
+LogicalResult verifyComputeLikeInputs(Operation* computeLikeOp,
+                                      ValueRange inputs,
+                                      bool allowChannelReceiveInputs,
+                                      StringRef kind,
+                                      pim::CappedDiagnosticReporter& diagnostics) {
+  for (auto [inputIndex, input] : llvm::enumerate(inputs)) {
+    unsigned currentInputIndex = inputIndex;
+    Operation* definingOp = input.getDefiningOp();
+    if (allowChannelReceiveInputs && isa_and_nonnull<spatial::SpatChannelReceiveOp>(definingOp))
+      continue;
+    if (isLegalHostBackedValue(input))
+      continue;
+
+    diagnostics.report(computeLikeOp, [&](Operation* illegalOp) {
+      InFlightDiagnostic diagnostic = illegalOp->emitOpError() << kind << " input #" << currentInputIndex
+                                                               << (allowChannelReceiveInputs
+                                                                     ? " must come from the host or an explicit "
+                                                                       "spat.channel_receive"
+                                                                     : " must come from the host");
+      if (definingOp)
+        diagnostic.attachNote(definingOp->getLoc()) << "illegal Spatial producer is " << definingOp->getName();
+    });
+    return failure();
+  }
+  return success();
+}
+
+void verifyNoExternalTensorCaptures(Operation* ownerOp,
+                                    Region& region,
+                                    StringRef kind,
+                                    pim::CappedDiagnosticReporter& diagnostics) {
+  region.walk([&](Operation* op) {
+    for (OpOperand& operand : op->getOpOperands()) {
+      Value value = operand.get();
+      if (!isa<TensorType>(value.getType()))
+        continue;
+      if (isDefinedInsideRegion(value, region) || isa<BlockArgument>(value))
+        continue;
+
+      Operation* definingOp = value.getDefiningOp();
+      if (definingOp && definingOp->hasTrait<OpTrait::ConstantLike>())
+        continue;
+
+      diagnostics.report(ownerOp, [&](Operation* illegalOp) {
+        InFlightDiagnostic diagnostic = illegalOp->emitOpError() << kind << " body may not capture external tensor "
+                                                                 << "values";
+        diagnostic.attachNote(op->getLoc())
+          << "tensor operand #" << operand.getOperandNumber() << " is defined outside the compute body by "
+          << (definingOp ? definingOp->getName().getStringRef() : StringRef("<block argument>"));
+      });
+    }
+  });
+}
+
+} // namespace
+
+LogicalResult verifyONNXToSpatial(func::FuncOp funcOp) {
+  pim::CappedDiagnosticReporter diagnostics;
+
+  for (Operation& op : funcOp.getOps()) {
+    if (isa<func::ReturnOp, spatial::SpatCompute, spatial::SpatComputeBatch>(&op))
+      continue;
+    if (isCompileTimeOp(&op))
+      continue;
+
+    diagnostics.report(&op, [](Operation* illegalOp) {
+      illegalOp->emitOpError(
+        "non-foldable top-level runtime op remains after ONNX-to-Spatial; lower it inside spat.compute");
+    });
+  }
+  checkWeightUseChains(funcOp, diagnostics);
+  diagnostics.emitSuppressedSummary(funcOp, "ONNX-to-Spatial verification failed");
+
+  return success(!diagnostics.hasFailure());
+}
+
+LogicalResult verifySpatialCommunicationInvariants(func::FuncOp funcOp) {
+  pim::CappedDiagnosticReporter diagnostics;
+
+  for (auto computeOp : funcOp.getOps<spatial::SpatCompute>()) {
+    (void)verifyComputeLikeInputs(
+      computeOp.getOperation(), computeOp.getInputs(), /*allowChannelReceiveInputs=*/true, "spat.compute", diagnostics);
+    verifyNoExternalTensorCaptures(computeOp.getOperation(), computeOp.getBody(), "spat.compute", diagnostics);
+  }
+
+  for (auto computeBatchOp : funcOp.getOps<spatial::SpatComputeBatch>()) {
+    (void)verifyComputeLikeInputs(computeBatchOp.getOperation(),
+                                  computeBatchOp.getInputs(),
+                                  /*allowChannelReceiveInputs=*/false,
+                                  "spat.compute_batch",
+                                  diagnostics);
+    verifyNoExternalTensorCaptures(
+      computeBatchOp.getOperation(), computeBatchOp.getBody(), "spat.compute_batch", diagnostics);
+  }
+
+  diagnostics.emitSuppressedSummary(funcOp, "Spatial communication invariant verification failed");
+  return success(!diagnostics.hasFailure());
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/ONNXToSpatialVerifier.hpp

@@ -0,0 +1,11 @@
+#pragma once
+
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Support/LogicalResult.h"
+
+namespace onnx_mlir {
+
+mlir::LogicalResult verifyONNXToSpatial(mlir::func::FuncOp funcOp);
+mlir::LogicalResult verifySpatialCommunicationInvariants(mlir::func::FuncOp funcOp);
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns.cpp

@@ -1,19 +1,16 @@
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 
 using namespace mlir;
 
 namespace onnx_mlir {
 
-namespace {
+void populatePrePatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
-
+  populateGeneratedPrePatterns(patterns, ctx);
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatial.hpp.inc"
+}
-
-} // namespace
-
-void populateConversionPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx) {
-  patterns.add<removeLRN>(ctx);
 
+void populateConversionPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  populateGeneratedConversionPatterns(patterns, ctx);
   populateElementwisePatterns(patterns, ctx);
   populateGemmPatterns(patterns, ctx);
   populateConvPatterns(patterns, ctx);
@@ -27,6 +24,11 @@ void populateConversionPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRCon
   populateResizePatterns(patterns, ctx);
   populateReshapePatterns(patterns, ctx);
   populateSplitPatterns(patterns, ctx);
+  populateTransposePatterns(patterns, ctx);
+}
+
+void populatePostPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  populateWeightPromotionPatterns(patterns, ctx);
 }
 
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns.hpp

@@ -1,38 +1,39 @@
 #pragma once
 
+#include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/IR/MLIRContext.h"
 #include "mlir/Transforms/DialectConversion.h"
 
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+
 namespace onnx_mlir {
 
+void populatePrePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 void populateConversionPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+void populatePostPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+
+void populateGeneratedPrePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+void populateGeneratedConversionPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+void populateWeightPromotionPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
 void populateConvPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateElementwisePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateGemmPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateMatMulRewritePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populatePoolPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateReduceMeanPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateReluPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateSigmoidPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateSoftmaxPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateConcatPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateGatherPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateResizePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateReshapePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
 void populateSplitPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+void populateTransposePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+
+bool requiresPostRewrite(spatial::SpatCompute computeOp);
+bool requiresPostRewrite(spatial::SpatComputeBatch computeOp);
+void annotateWeightsConstants(mlir::func::FuncOp funcOp);
 
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns/GeneratedConversion.cpp

@@ -0,0 +1,18 @@
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+
+namespace {
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatial.hpp.inc"
+
+} // namespace
+
+void populateGeneratedConversionPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  patterns.add<removeLRN>(ctx);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/Conv.cpp

@@ -11,7 +11,7 @@
 #include "src/Accelerators/PIM/Common/Support/Diagnostics.hpp"
 #include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -111,6 +111,32 @@ static Value buildPackedWeight(DenseElementsAttr wDenseAttr,
   return arith::ConstantOp::create(rewriter, loc, packedWeightType, packedAttr);
 }
 
+static Value createConvWeightMatrix(Value w,
+                                    RankedTensorType wFlatType,
+                                    RankedTensorType wTransType,
+                                    ConversionPatternRewriter& rewriter,
+                                    Location loc) {
+  auto buildWeightMatrix = [&](Value weight) -> Value {
+    Value wFlat = tensor::CollapseShapeOp::create(rewriter,
+                                                  loc,
+                                                  wFlatType,
+                                                  weight,
+                                                  SmallVector<ReassociationIndices> {
+                                                    {0},
+                                                    {1, 2, 3}
+    });
+    return ONNXTransposeOp::create(rewriter, loc, wTransType, wFlat, rewriter.getI64ArrayAttr({1, 0})).getResult();
+  };
+
+  if (isCompileTimeComputable(w))
+    return buildWeightMatrix(w);
+
+  auto computeOp = createSpatCompute<1>(rewriter, loc, TypeRange {wTransType}, {}, ValueRange {w}, [&](Value weight) {
+    spatial::SpatYieldOp::create(rewriter, loc, buildWeightMatrix(weight));
+  });
+  return computeOp.getResult(0);
+}
+
 static Value buildPackedBias(bool hasBias,
                              Value gemmBias,
                              Value biasMatrix,
@@ -391,19 +417,11 @@ static Value lowerSingleConvGroup(Value x,
   const int64_t xbarSize = static_cast<int64_t>(crossbarSize.getValue());
   const int64_t wMaxDim = std::max(patchSize, numChannelsOut);
   const int64_t maxParallelPixels = std::max<int64_t>(1, xbarSize / wMaxDim);
-  auto wDenseAttr = getHostFoldableDenseElementsAttr(w);
+  auto wDenseAttr = getHostConstDenseElementsAttr(w);
 
   // Prepare weight matrix W for crossbar storage:
   // W: [numChannelsOut, numChannelsIn, wHeight, wWidth] -> [numChannelsOut, patchSize] -> [patchSize, numChannelsOut]
-  Value wFlat = tensor::CollapseShapeOp::create(rewriter,
+  Value wTrans = createConvWeightMatrix(w, wFlatType, wTransType, rewriter, loc);
-                                                loc,
-                                                wFlatType,
-                                                w,
-                                                SmallVector<ReassociationIndices> {
-                                                  {0},
-                                                  {1, 2, 3}
-  });
-  Value wTrans = ONNXTransposeOp::create(rewriter, loc, wTransType, wFlat, rewriter.getI64ArrayAttr({1, 0}));
 
   // Pass bias through directly; Gemm handles rank-1 C canonicalization.
   bool hasB = !isa<ONNXNoneOp>(b.getDefiningOp());
@@ -412,7 +430,7 @@ static Value lowerSingleConvGroup(Value x,
   DenseElementsAttr biasDenseAttr;
   if (hasB) {
     gemmBias = b;
-    biasDenseAttr = getHostFoldableDenseElementsAttr(b);
+    biasDenseAttr = getHostConstDenseElementsAttr(b);
     biasMatrix = expandBiasIfNeeded(b, rewriter, loc);
   }
   const bool canPackWeightsAsConstants = static_cast<bool>(wDenseAttr);
@@ -717,7 +735,7 @@ LogicalResult ConvToGemm::matchAndRewrite(ONNXConvOp convOp,
   }
 
   Value result;
-  if (llvm::all_of(groupResults, isHostFoldableValue)) {
+  if (llvm::all_of(groupResults, isCompileTimeComputable)) {
     result = createSpatConcat(rewriter, loc, /*axis=*/1, groupResults);
   }
   else {

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/Elementwise.cpp

@@ -7,7 +7,7 @@
 
 #include "src/Accelerators/PIM/Common/IR/ShapeUtils.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/MatMul.cpp

@@ -9,8 +9,8 @@
 #include <numeric>
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -38,23 +38,16 @@ static FailureOr<SmallVector<int64_t>> inferSupportedBatchShape(ArrayRef<int64_t
   return SmallVector<int64_t>(lhsBatchShape.begin(), lhsBatchShape.end());
 }
 
-static Value collapseBatchDims(Value value,
+static Value
-                               int64_t batchSize,
+collapseBatchDims(Value value, int64_t batchSize, int64_t rows, int64_t cols, PatternRewriter& rewriter, Location loc) {
-                               int64_t rows,
-                               int64_t cols,
-                               PatternRewriter& rewriter,
-                               Location loc) {
   auto type = cast<RankedTensorType>(value.getType());
   if (type.getRank() == 2 || type.getRank() == 3)
     return value;
 
-  auto collapsedType =
+  auto collapsedType = RankedTensorType::get({batchSize, rows, cols}, type.getElementType(), type.getEncoding());
-    RankedTensorType::get({batchSize, rows, cols}, type.getElementType(), type.getEncoding());
+  SmallVector<ReassociationIndices> reassociation = {ReassociationIndices {},
-  SmallVector<ReassociationIndices> reassociation = {
-    ReassociationIndices {},
                                                      ReassociationIndices {static_cast<int64_t>(type.getRank() - 2)},
-    ReassociationIndices {static_cast<int64_t>(type.getRank() - 1)}
+                                                     ReassociationIndices {static_cast<int64_t>(type.getRank() - 1)}};
-  };
   for (int64_t dim = 0; dim < type.getRank() - 2; ++dim)
     reassociation.front().push_back(dim);
 
@@ -62,7 +55,7 @@ static Value collapseBatchDims(Value value,
     return tensor::CollapseShapeOp::create(rewriter, loc, collapsedType, input, reassociation);
   };
 
-  if (isHostFoldableValue(value))
+  if (isCompileTimeComputable(value))
     return buildCollapsed(value);
 
   auto collapseCompute =
@@ -72,19 +65,14 @@ static Value collapseBatchDims(Value value,
   return collapseCompute.getResult(0);
 }
 
-static Value expandBatchDims(Value value,
+static Value
-                             RankedTensorType outputType,
+expandBatchDims(Value value, RankedTensorType outputType, size_t batchRank, PatternRewriter& rewriter, Location loc) {
-                             size_t batchRank,
-                             PatternRewriter& rewriter,
-                             Location loc) {
   if (cast<RankedTensorType>(value.getType()) == outputType)
     return value;
 
-  SmallVector<ReassociationIndices> reassociation = {
+  SmallVector<ReassociationIndices> reassociation = {ReassociationIndices {},
-    ReassociationIndices {},
                                                      ReassociationIndices {static_cast<int64_t>(batchRank)},
-    ReassociationIndices {static_cast<int64_t>(batchRank + 1)}
+                                                     ReassociationIndices {static_cast<int64_t>(batchRank + 1)}};
-  };
   for (size_t dim = 0; dim < batchRank; ++dim)
     reassociation.front().push_back(static_cast<int64_t>(dim));
 
@@ -126,7 +114,7 @@ static Value extractBatchMatrix(Value value,
     });
   };
 
-  if (isHostFoldableValue(value))
+  if (isCompileTimeComputable(value))
     return buildMatrix(value);
 
   auto batchMatrixCompute =
@@ -154,7 +142,7 @@ static Value transposeLastTwoDims(Value value, PatternRewriter& rewriter, Locati
     return ONNXTransposeOp::create(rewriter, loc, transposedType, input, rewriter.getI64ArrayAttr(perm));
   };
 
-  if (isHostFoldableValue(value))
+  if (isCompileTimeComputable(value))
     return buildTranspose(value);
 
   auto transposeCompute =
@@ -194,7 +182,7 @@ static Value concatValues(ValueRange inputs, int64_t axis, PatternRewriter& rewr
   outputShape[axis] = concatDimSize;
   auto resultType = RankedTensorType::get(outputShape, firstType.getElementType(), firstType.getEncoding());
 
-  if (llvm::all_of(inputs, isHostFoldableValue))
+  if (llvm::all_of(inputs, isCompileTimeComputable))
     return createSpatConcat(rewriter, loc, axis, inputs);
 
   auto concatCompute = createSpatCompute(rewriter, loc, TypeRange {resultType}, {}, inputs, [&](ValueRange args) {
@@ -247,7 +235,7 @@ struct MatMulToGemm : OpRewritePattern<ONNXMatMulOp> {
     }
 
     Location loc = matmulOp.getLoc();
-    bool useTransposedForm = isHostFoldableValue(matmulOp.getA()) && !isHostFoldableValue(matmulOp.getB());
+    bool useTransposedForm = isCompileTimeComputable(matmulOp.getA()) && !isCompileTimeComputable(matmulOp.getB());
 
     Value lhs = collapseBatchDims(matmulOp.getA(), lhsBatch, m, k, rewriter, loc);
     Value rhs = collapseBatchDims(matmulOp.getB(), rhsBatch, k, n, rewriter, loc);

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/ReduceMean.cpp

@@ -6,8 +6,8 @@
 #include <algorithm>
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -91,7 +91,7 @@ static Value concatValues(ValueRange inputs, int64_t axis, ConversionPatternRewr
   outputShape[axis] = concatDimSize;
   auto resultType = RankedTensorType::get(outputShape, firstType.getElementType(), firstType.getEncoding());
 
-  if (llvm::all_of(inputs, isHostFoldableValue))
+  if (llvm::all_of(inputs, isCompileTimeComputable))
     return createSpatConcat(rewriter, loc, axis, inputs);
 
   auto concatCompute = createSpatCompute(rewriter, loc, TypeRange {resultType}, {}, inputs, [&](ValueRange args) {
@@ -135,7 +135,7 @@ static Value squeezeReducedAxes(Value keepdimsValue,
   }
 
   auto reassociation = buildCollapseReassociation(reducedAxes);
-  if (isHostFoldableValue(keepdimsValue))
+  if (isCompileTimeComputable(keepdimsValue))
     return tensor::CollapseShapeOp::create(rewriter, loc, resultType, keepdimsValue, reassociation).getResult();
 
   auto squeezeCompute =

src/PIM/Conversion/ONNXToSpatial/Patterns/NN/Softmax.cpp

@@ -4,7 +4,7 @@
 #include "mlir/Transforms/DialectConversion.h"
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 

src/PIM/Conversion/ONNXToSpatial/Patterns/Post.cpp

@@ -1,5 +1,6 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/IRMapping.h"
 #include "mlir/IR/PatternMatch.h"
@@ -9,7 +10,7 @@
 
 #include "src/Accelerators/PIM/Common/IR/WeightUtils.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/WeightMaterialization.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/PostPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -20,73 +21,33 @@ namespace onnx_mlir {
 namespace {
 
 static bool isWeightMaterializationHelperUser(Operation* op) {
-  return isa<tensor::ExtractSliceOp, tensor::ExpandShapeOp, tensor::CollapseShapeOp, ONNXTransposeOp>(op);
+  return isa<tensor::ExtractSliceOp, tensor::ExpandShapeOp, tensor::CollapseShapeOp, linalg::TransposeOp>(op);
 }
 
 static bool canPromoteInputBlockArgument(BlockArgument arg) {
   return !arg.use_empty() && llvm::all_of(arg.getUsers(), isWeightMaterializationHelperUser);
 }
 
+static bool canPromoteInputBlockArgument(std::optional<BlockArgument> arg) {
+  return arg && canPromoteInputBlockArgument(*arg);
+}
+
 static bool isDirectConstantValue(Value value) {
   return isa_and_nonnull<arith::ConstantOp, ONNXConstantOp>(value.getDefiningOp());
 }
 
 template <typename ComputeOpTy>
 static bool hasPromotableWeightLikeInputs(ComputeOpTy compute) {
-  Block& block = compute.getBody().front();
   for (auto [inputIdx, input] : llvm::enumerate(compute.getInputs())) {
-    if (inputIdx >= block.getNumArguments())
-      continue;
     if (!isWeightLikeComputeOperand(input))
       continue;
-    if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(block.getArgument(inputIdx)))
+    if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(compute.getInputArgument(inputIdx)))
       continue;
     return true;
   }
   return false;
 }
 
-// Collapses one-lane batches so later phases do not carry batch-only structure unnecessarily.
-struct FoldSingleLaneComputeBatchPattern : OpRewritePattern<spatial::SpatComputeBatch> {
-  using OpRewritePattern<spatial::SpatComputeBatch>::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(spatial::SpatComputeBatch batchOp, PatternRewriter& rewriter) const override {
-    if (batchOp.getLaneCount() != 1)
-      return rewriter.notifyMatchFailure(batchOp, "requires a single lane");
-
-    auto loc = batchOp.getLoc();
-    rewriter.setInsertionPoint(batchOp);
-    auto computeOp =
-      spatial::SpatCompute::create(rewriter, loc, batchOp.getResultTypes(), batchOp.getWeights(), batchOp.getInputs());
-    computeOp.getProperties().setOperandSegmentSizes(
-      {static_cast<int>(batchOp.getWeights().size()), static_cast<int>(batchOp.getInputs().size())});
-
-    Block& templateBlock = batchOp.getBody().front();
-    SmallVector<Type> blockArgTypes;
-    SmallVector<Location> blockArgLocs;
-    blockArgTypes.reserve(templateBlock.getNumArguments());
-    blockArgLocs.reserve(templateBlock.getNumArguments());
-    for (BlockArgument arg : templateBlock.getArguments()) {
-      blockArgTypes.push_back(arg.getType());
-      blockArgLocs.push_back(loc);
-    }
-
-    auto* newBlock =
-      rewriter.createBlock(&computeOp.getBody(), computeOp.getBody().end(), TypeRange(blockArgTypes), blockArgLocs);
-    IRMapping mapper;
-    for (auto [oldArg, newArg] : llvm::zip(templateBlock.getArguments(), newBlock->getArguments()))
-      mapper.map(oldArg, newArg);
-
-    rewriter.setInsertionPointToEnd(newBlock);
-    for (Operation& op : templateBlock)
-      rewriter.clone(op, mapper);
-
-    batchOp->replaceAllUsesWith(computeOp->getResults());
-    rewriter.eraseOp(batchOp);
-    return success();
-  }
-};
-
 // Promotes foldable helper chains from runtime inputs to weights to avoid artificial compute inputs.
 struct PromoteWeightLikeComputeInputsPattern : OpRewritePattern<spatial::SpatCompute> {
   using OpRewritePattern<spatial::SpatCompute>::OpRewritePattern;
@@ -96,11 +57,9 @@ struct PromoteWeightLikeComputeInputsPattern : OpRewritePattern<spatial::SpatCom
     bool needsRewrite = false;
     Block& oldBlock = compute.getBody().front();
     for (auto [inputIdx, input] : llvm::enumerate(compute.getInputs())) {
-      if (inputIdx >= oldBlock.getNumArguments())
-        continue;
       if (!isWeightLikeComputeOperand(input))
         continue;
-      if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(oldBlock.getArgument(inputIdx)))
+      if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(compute.getInputArgument(inputIdx)))
         continue;
       promoteInput[inputIdx] = true;
       needsRewrite = true;
@@ -131,8 +90,16 @@ struct PromoteWeightLikeComputeInputsPattern : OpRewritePattern<spatial::SpatCom
 
     auto newCompute =
       spatial::SpatCompute::create(rewriter, compute.getLoc(), compute.getResultTypes(), newWeights, newInputs);
-    auto* newBlock =
+    SmallVector<Type> newBlockArgTypes;
-      rewriter.createBlock(&newCompute.getBody(), newCompute.getBody().end(), newInputTypes, newInputLocs);
+    SmallVector<Location> newBlockArgLocs;
+    for (Value weight : newWeights) {
+      newBlockArgTypes.push_back(weight.getType());
+      newBlockArgLocs.push_back(weight.getLoc());
+    }
+    llvm::append_range(newBlockArgTypes, newInputTypes);
+    llvm::append_range(newBlockArgLocs, newInputLocs);
+    auto* newBlock = rewriter.createBlock(
+      &newCompute.getBody(), newCompute.getBody().end(), TypeRange(newBlockArgTypes), newBlockArgLocs);
     newCompute.getProperties().setOperandSegmentSizes(
       {static_cast<int>(newWeights.size()), static_cast<int>(newInputs.size())});
     rewriter.setInsertionPointToStart(newBlock);
@@ -141,17 +108,30 @@ struct PromoteWeightLikeComputeInputsPattern : OpRewritePattern<spatial::SpatCom
     bodyRewriter.setInsertionPointToStart(newBlock);
 
     IRMapping mapper;
+    for (auto [weightIndex, weight] : llvm::enumerate(compute.getWeights())) {
+      auto oldWeightArg = compute.getWeightArgument(weightIndex);
+      auto newWeightArg = newCompute.getWeightArgument(weightIndex);
+      if (!oldWeightArg || !newWeightArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute weight block argument during rewrite");
+      mapper.map(*oldWeightArg, *newWeightArg);
+    }
     size_t newInputIdx = 0;
-    for (auto [oldInputIdx, oldArg] : llvm::enumerate(oldBlock.getArguments())) {
+    for (auto [oldInputIdx, input] : llvm::enumerate(compute.getInputs())) {
+      auto oldArg = compute.getInputArgument(oldInputIdx);
+      if (!oldArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute input block argument during rewrite");
       if (!promoteInput[oldInputIdx]) {
-        mapper.map(oldArg, newBlock->getArgument(newInputIdx++));
+        auto newInputArg = newCompute.getInputArgument(newInputIdx++);
+        if (!newInputArg)
+          return rewriter.notifyMatchFailure(compute, "missing rewritten compute input block argument");
+        mapper.map(*oldArg, *newInputArg);
         continue;
       }
 
-      auto clonedValue = materializeWeightLikeValueInBlock(compute.getInputs()[oldInputIdx], bodyRewriter, mapper);
+      auto clonedValue = materializeWeightLikeValueInBlock(input, bodyRewriter, mapper);
       if (failed(clonedValue))
         return rewriter.notifyMatchFailure(compute, "failed to materialize promoted weight-like operand");
-      mapper.map(oldArg, *clonedValue);
+      mapper.map(*oldArg, *clonedValue);
     }
 
     for (Operation& op : oldBlock.without_terminator())
@@ -180,11 +160,9 @@ struct PromoteWeightLikeComputeBatchInputsPattern : OpRewritePattern<spatial::Sp
     bool needsRewrite = false;
     Block& oldBlock = compute.getBody().front();
     for (auto [inputIdx, input] : llvm::enumerate(compute.getInputs())) {
-      if (inputIdx >= oldBlock.getNumArguments())
-        continue;
       if (!isWeightLikeComputeOperand(input))
         continue;
-      if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(oldBlock.getArgument(inputIdx)))
+      if (isDirectConstantValue(input) && !canPromoteInputBlockArgument(compute.getInputArgument(inputIdx)))
         continue;
       promoteInput[inputIdx] = true;
       needsRewrite = true;
@@ -220,8 +198,31 @@ struct PromoteWeightLikeComputeBatchInputsPattern : OpRewritePattern<spatial::Sp
                                         rewriter.getI32IntegerAttr(static_cast<int32_t>(compute.getLaneCount())),
                                         newWeights,
                                         newInputs);
-    auto* newBlock =
+    auto laneArg = compute.getLaneArgument();
-      rewriter.createBlock(&newCompute.getBody(), newCompute.getBody().end(), newInputTypes, newInputLocs);
+    if (!laneArg)
+      return rewriter.notifyMatchFailure(compute, "missing compute_batch lane block argument");
+    SmallVector<Type> newBlockArgTypes;
+    SmallVector<Location> newBlockArgLocs;
+    newBlockArgTypes.reserve(1 + newWeights.size() + newInputTypes.size() + compute.getNumResults());
+    newBlockArgLocs.reserve(1 + newWeights.size() + newInputLocs.size() + compute.getNumResults());
+    newBlockArgTypes.push_back(laneArg->getType());
+    newBlockArgLocs.push_back(laneArg->getLoc());
+    for (Value weight : newWeights) {
+      newBlockArgTypes.push_back(weight.getType());
+      newBlockArgLocs.push_back(weight.getLoc());
+    }
+    llvm::append_range(newBlockArgTypes, newInputTypes);
+    llvm::append_range(newBlockArgLocs, newInputLocs);
+    for (auto [resultIndex, resultType] : llvm::enumerate(compute.getResultTypes())) {
+      auto outputArg = compute.getOutputArgument(resultIndex);
+      if (!outputArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute_batch output block argument");
+      newBlockArgTypes.push_back(resultType);
+      newBlockArgLocs.push_back(outputArg->getLoc());
+    }
+
+    auto* newBlock = rewriter.createBlock(
+      &newCompute.getBody(), newCompute.getBody().end(), TypeRange(newBlockArgTypes), newBlockArgLocs);
     newCompute.getProperties().setOperandSegmentSizes(
       {static_cast<int>(newWeights.size()), static_cast<int>(newInputs.size())});
     rewriter.setInsertionPointToStart(newBlock);
@@ -230,31 +231,45 @@ struct PromoteWeightLikeComputeBatchInputsPattern : OpRewritePattern<spatial::Sp
     bodyRewriter.setInsertionPointToStart(newBlock);
 
     IRMapping mapper;
+    auto newLaneArg = newCompute.getLaneArgument();
+    if (!newLaneArg)
+      return rewriter.notifyMatchFailure(compute, "missing rewritten compute_batch lane block argument");
+    mapper.map(*laneArg, *newLaneArg);
+    for (auto [weightIndex, weight] : llvm::enumerate(compute.getWeights())) {
+      auto oldWeightArg = compute.getWeightArgument(weightIndex);
+      auto newWeightArg = newCompute.getWeightArgument(weightIndex);
+      if (!oldWeightArg || !newWeightArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute_batch weight block argument during rewrite");
+      mapper.map(*oldWeightArg, *newWeightArg);
+    }
     size_t newInputIdx = 0;
-    for (auto [oldInputIdx, oldArg] : llvm::enumerate(oldBlock.getArguments())) {
+    for (auto [oldInputIdx, input] : llvm::enumerate(compute.getInputs())) {
+      auto oldArg = compute.getInputArgument(oldInputIdx);
+      if (!oldArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute_batch input block argument during rewrite");
       if (!promoteInput[oldInputIdx]) {
-        mapper.map(oldArg, newBlock->getArgument(newInputIdx++));
+        auto newInputArg = newCompute.getInputArgument(newInputIdx++);
+        if (!newInputArg)
+          return rewriter.notifyMatchFailure(compute, "missing rewritten compute_batch input block argument");
+        mapper.map(*oldArg, *newInputArg);
         continue;
       }
 
-      auto clonedValue = materializeWeightLikeValueInBlock(compute.getInputs()[oldInputIdx], bodyRewriter, mapper);
+      auto clonedValue = materializeWeightLikeValueInBlock(input, bodyRewriter, mapper);
       if (failed(clonedValue))
         return rewriter.notifyMatchFailure(compute, "failed to materialize promoted batch weight-like operand");
-      mapper.map(oldArg, *clonedValue);
+      mapper.map(*oldArg, *clonedValue);
+    }
+    for (auto resultIndex : llvm::seq<size_t>(0, compute.getNumResults())) {
+      auto outputArg = compute.getOutputArgument(resultIndex);
+      if (!outputArg)
+        return rewriter.notifyMatchFailure(compute, "missing compute_batch output block argument during rewrite");
+      mapper.map(*outputArg, newBlock->getArgument(1 + newWeights.size() + newInputs.size() + resultIndex));
     }
 
-    for (Operation& op : oldBlock.without_terminator())
+    for (Operation& op : oldBlock)
       rewriter.clone(op, mapper);
 
-    auto oldYield = cast<spatial::SpatYieldOp>(oldBlock.getTerminator());
-    SmallVector<Value> newYieldOperands;
-    newYieldOperands.reserve(oldYield.getOutputs().size());
-    for (Value operand : oldYield.getOutputs()) {
-      auto mapped = mapper.lookupOrNull(operand);
-      newYieldOperands.push_back(mapped ? cast<Value>(mapped) : operand);
-    }
-    spatial::SpatYieldOp::create(rewriter, oldYield.getLoc(), newYieldOperands);
-
     rewriter.replaceOp(compute, newCompute.getResults());
     return success();
   }
@@ -262,11 +277,7 @@ struct PromoteWeightLikeComputeBatchInputsPattern : OpRewritePattern<spatial::Sp
 
 } // namespace
 
-void populateEarlyPostPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+void populateWeightPromotionPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
-  patterns.add<FoldSingleLaneComputeBatchPattern>(ctx);
-}
-
-void populatePostPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
   patterns.add<PromoteWeightLikeComputeInputsPattern, PromoteWeightLikeComputeBatchInputsPattern>(ctx);
 }
 
@@ -277,8 +288,6 @@ void annotateWeightsConstants(func::FuncOp funcOp) {
   });
 }
 
-bool requiresEarlyPostRewrite(spatial::SpatComputeBatch batchOp) { return batchOp.getLaneCount() == 1; }
-
 bool requiresPostRewrite(spatial::SpatCompute computeOp) { return hasPromotableWeightLikeInputs(computeOp); }
 
 bool requiresPostRewrite(spatial::SpatComputeBatch computeOp) { return hasPromotableWeightLikeInputs(computeOp); }

src/PIM/Conversion/ONNXToSpatial/Patterns/Pre.cpp

@@ -1,6 +1,5 @@
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/PrePatterns.hpp"
 
 using namespace mlir;
 
@@ -12,7 +11,7 @@ namespace {
 
 } // namespace
 
-void populatePrePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx) {
+void populateGeneratedPrePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx) {
   patterns.add<onnxToArithConstant>(ctx);
   patterns.add<convAddToConvWithBiasLeft>(ctx);
   patterns.add<convAddToConvWithBiasRight>(ctx);

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Concat.cpp

@@ -3,7 +3,7 @@
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/ComputeRegionBuilder.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -20,7 +20,7 @@ struct Concat : public OpConversionPattern<ONNXConcatOp> {
     auto inputs = adaptor.getInputs();
     int64_t axis = adaptor.getAxis();
 
-    if (llvm::all_of(inputs, isHostFoldableValue)) {
+    if (llvm::all_of(inputs, isCompileTimeComputable)) {
       rewriter.replaceOp(maxpoolOp, createSpatConcat(rewriter, maxpoolOp.getLoc(), axis, inputs));
       return success();
     }

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Gather.cpp

@@ -6,7 +6,7 @@
 #include "llvm/ADT/SmallVector.h"
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Reshape.cpp

@@ -4,8 +4,8 @@
 #include "llvm/ADT/SmallVector.h"
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -115,7 +115,7 @@ struct Reshape : OpConversionPattern<ONNXReshapeOp> {
     }
 
     auto replaceWithReshape = [&](auto buildReshape) -> LogicalResult {
-      if (isHostFoldableValue(adaptor.getData())) {
+      if (isCompileTimeComputable(adaptor.getData())) {
         rewriter.replaceOp(reshapeOp, buildReshape(adaptor.getData()));
         return success();
       }

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Resize.cpp

@@ -6,7 +6,7 @@
 #include "llvm/ADT/STLExtras.h"
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -58,24 +58,21 @@ static Value buildNearestResizeLoop(Value input,
 
   Value outputC = channelLoop.getInductionVar();
   Value outputChannelAcc = channelLoop.getRegionIterArgs().front();
-  Value inputC =
+  Value inputC = buildNearestAsymmetricIndex(outputC, inputType.getDimSize(1), resultType.getDimSize(1), rewriter, loc);
-    buildNearestAsymmetricIndex(outputC, inputType.getDimSize(1), resultType.getDimSize(1), rewriter, loc);
 
   auto heightLoop = scf::ForOp::create(rewriter, loc, c0, cOutputH, c1, ValueRange {outputChannelAcc});
   rewriter.setInsertionPointToStart(heightLoop.getBody());
 
   Value outputH = heightLoop.getInductionVar();
   Value outputHeightAcc = heightLoop.getRegionIterArgs().front();
-  Value inputH =
+  Value inputH = buildNearestAsymmetricIndex(outputH, inputType.getDimSize(2), resultType.getDimSize(2), rewriter, loc);
-    buildNearestAsymmetricIndex(outputH, inputType.getDimSize(2), resultType.getDimSize(2), rewriter, loc);
 
   auto widthLoop = scf::ForOp::create(rewriter, loc, c0, cOutputW, c1, ValueRange {outputHeightAcc});
   rewriter.setInsertionPointToStart(widthLoop.getBody());
 
   Value outputW = widthLoop.getInductionVar();
   Value outputWidthAcc = widthLoop.getRegionIterArgs().front();
-  Value inputW =
+  Value inputW = buildNearestAsymmetricIndex(outputW, inputType.getDimSize(3), resultType.getDimSize(3), rewriter, loc);
-    buildNearestAsymmetricIndex(outputW, inputType.getDimSize(3), resultType.getDimSize(3), rewriter, loc);
 
   SmallVector<OpFoldResult> inputOffsets = {inputN, inputC, inputH, inputW};
   Value inputSlice =
@@ -114,8 +111,8 @@ struct Resize : OpConversionPattern<ONNXResizeOp> {
 
     if (resizeOp.getMode() != "nearest" || resizeOp.getCoordinateTransformationMode() != "asymmetric"
         || resizeOp.getNearestMode() != "floor")
-      return rewriter.notifyMatchFailure(
+      return rewriter.notifyMatchFailure(resizeOp,
-        resizeOp, "resize lowering currently supports only nearest + asymmetric + floor.");
+                                         "resize lowering currently supports only nearest + asymmetric + floor.");
 
     if (llvm::any_of(inputType.getShape(), [](int64_t dim) { return dim <= 0; })
         || llvm::any_of(resultType.getShape(), [](int64_t dim) { return dim <= 0; }))

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Split.cpp

@@ -2,8 +2,8 @@
 #include "mlir/Transforms/DialectConversion.h"
 
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ConversionPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/CompileTime.hpp"
-#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/HostFoldability.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -61,7 +61,7 @@ struct Split : OpConversionPattern<ONNXSplitOp> {
       sliceSizes.push_back(resultType.getShape()[axis]);
     }
 
-    if (isHostFoldableValue(adaptor.getInput())) {
+    if (isCompileTimeComputable(adaptor.getInput())) {
       for (int64_t sliceSize : sliceSizes) {
         outputs.push_back(extractSliceAt(adaptor.getInput(), axis, offset, sliceSize, rewriter, splitOp.getLoc()));
         offset += sliceSize;

src/PIM/Conversion/ONNXToSpatial/Patterns/Tensor/Transpose.cpp

@@ -0,0 +1,75 @@
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Transforms/DialectConversion.h"
+
+#include "llvm/ADT/SmallVector.h"
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
+#include "src/Dialect/ONNX/ONNXOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace {
+
+static Value createTransposeInit(Value input,
+                                 RankedTensorType resultType,
+                                 ArrayRef<int64_t> permutation,
+                                 ConversionPatternRewriter& rewriter,
+                                 Location loc) {
+  SmallVector<OpFoldResult> sizes;
+  sizes.reserve(resultType.getRank());
+  for (auto [resultDim, sourceDim] : llvm::zip_equal(resultType.getShape(), permutation)) {
+    if (!ShapedType::isDynamic(resultDim)) {
+      sizes.push_back(rewriter.getIndexAttr(resultDim));
+      continue;
+    }
+    sizes.push_back(tensor::DimOp::create(rewriter, loc, input, sourceDim).getResult());
+  }
+  return tensor::EmptyOp::create(rewriter, loc, sizes, resultType.getElementType()).getResult();
+}
+
+static SmallVector<int64_t> getTransposePermutation(ONNXTransposeOp transposeOp) {
+  auto inputType = cast<RankedTensorType>(transposeOp.getData().getType());
+  SmallVector<int64_t> permutation;
+  if (auto permAttr = transposeOp.getPermAttr()) {
+    permutation.reserve(permAttr.size());
+    for (IntegerAttr attr : permAttr.getAsRange<IntegerAttr>())
+      permutation.push_back(attr.getInt());
+    return permutation;
+  }
+
+  permutation.reserve(inputType.getRank());
+  for (int64_t dim = inputType.getRank() - 1; dim >= 0; --dim)
+    permutation.push_back(dim);
+  return permutation;
+}
+
+struct TransposeToLinalgTranspose : OpConversionPattern<ONNXTransposeOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult matchAndRewrite(ONNXTransposeOp transposeOp,
+                                ONNXTransposeOpAdaptor adaptor,
+                                ConversionPatternRewriter& rewriter) const override {
+    auto inputType = dyn_cast<RankedTensorType>(adaptor.getData().getType());
+    auto resultType = dyn_cast<RankedTensorType>(transposeOp.getResult().getType());
+    if (!inputType || !resultType)
+      return failure();
+
+    SmallVector<int64_t> permutation = getTransposePermutation(transposeOp);
+    Value init = createTransposeInit(adaptor.getData(), resultType, permutation, rewriter, transposeOp.getLoc());
+    Value transposed =
+      linalg::TransposeOp::create(rewriter, transposeOp.getLoc(), adaptor.getData(), init, permutation)
+        .getResult()[0];
+    rewriter.replaceOp(transposeOp, transposed);
+    return success();
+  }
+};
+
+} // namespace
+
+void populateTransposePatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  patterns.add<TransposeToLinalgTranspose>(ctx);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/PostPatterns.hpp

@@ -1,22 +0,0 @@
-#pragma once
-
-#include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/IR/MLIRContext.h"
-
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-
-namespace onnx_mlir {
-
-bool requiresEarlyPostRewrite(spatial::SpatComputeBatch batchOp);
-
-bool requiresPostRewrite(spatial::SpatCompute computeOp);
-
-bool requiresPostRewrite(spatial::SpatComputeBatch computeOp);
-
-void populateEarlyPostPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
-void populatePostPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
-void annotateWeightsConstants(mlir::func::FuncOp funcOp);
-
-} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/PrePatterns.hpp

@@ -1,10 +0,0 @@
-#pragma once
-
-#include "mlir/IR/MLIRContext.h"
-#include "mlir/Transforms/DialectConversion.h"
-
-namespace onnx_mlir {
-
-void populatePrePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/BatchCoreLoweringPatterns.cpp

@@ -1,11 +1,14 @@
+#include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/IRMapping.h"
+#include "mlir/IR/Matchers.h"
 
 #include "Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "Conversion/SpatialToPim/SpatialToPimPass.hpp"
 #include "src/Accelerators/PIM/Common/PimCommon.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/BatchCoreLoweringPatterns.hpp"
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
@@ -15,7 +18,17 @@ using namespace onnx_mlir::pim;
 namespace onnx_mlir {
 namespace {
 
-static int32_t translateSpatialCoreIdToPimCoreId(size_t spatialCoreId) { return static_cast<int32_t>(spatialCoreId); }
+static bool isExplicitHostOperand(Operation* op, unsigned operandIndex) {
+  if (isa<pim::PimMemCopyDevToHostOp>(op))
+    return operandIndex == 2;
+  return false;
+}
+
+static bool isUsedOnlyAsExplicitHostOperand(Value value) {
+  return !value.use_empty() && llvm::all_of(value.getUses(), [](OpOperand& use) {
+    return isExplicitHostOperand(use.getOwner(), use.getOperandNumber());
+  });
+}
 
 static SmallVector<int32_t> getPimCoreIdsForBatchOp(spatial::SpatComputeBatch computeBatchOp, size_t& fallbackCoreId) {
   if (auto coreIdsAttr = computeBatchOp->getAttrOfType<DenseI32ArrayAttr>(onnx_mlir::kCoreIdsAttrName))
@@ -28,54 +41,75 @@ static SmallVector<int32_t> getPimCoreIdsForBatchOp(spatial::SpatComputeBatch co
   return coreIds;
 }
 
-static void lowerChannelSendTensorBatch(spatial::SpatChannelSendTensorBatchOp sendTensorBatchOp,
+static FailureOr<unsigned> getDirectReturnOperandIndex(OpResult result) {
-                                        IRMapping& mapper,
+  if (!result.hasOneUse())
-                                        IRRewriter& rewriter) {
+    return failure();
-  SmallVector<int32_t> targetCoreIds;
-  targetCoreIds.reserve(sendTensorBatchOp.getTargetCoreIds().size());
-  for (int32_t targetCoreId : sendTensorBatchOp.getTargetCoreIds())
-    targetCoreIds.push_back(translateSpatialCoreIdToPimCoreId(targetCoreId));
 
-  pim::PimSendTensorBatchOp::create(rewriter,
+  auto returnOp = dyn_cast<func::ReturnOp>(*result.getUsers().begin());
-                                    sendTensorBatchOp.getLoc(),
+  if (!returnOp)
-                                    mapper.lookup(sendTensorBatchOp.getInput()),
+    return failure();
-                                    rewriter.getDenseI32ArrayAttr(targetCoreIds));
+  return result.getUses().begin()->getOperandNumber();
 }
 
-static void lowerChannelReceiveTensorBatch(spatial::SpatChannelReceiveTensorBatchOp receiveTensorBatchOp,
+static Value createScaledIndexValue(IRRewriter& rewriter, Location loc, Value base, int64_t scale) {
-                                           IRMapping& mapper,
+  if (scale == 1)
-                                           IRRewriter& rewriter) {
+    return base;
-  SmallVector<int32_t> sourceCoreIds;
-  sourceCoreIds.reserve(receiveTensorBatchOp.getSourceCoreIds().size());
-  for (int32_t sourceCoreId : receiveTensorBatchOp.getSourceCoreIds())
-    sourceCoreIds.push_back(translateSpatialCoreIdToPimCoreId(sourceCoreId));
 
-  auto outputType = cast<ShapedType>(receiveTensorBatchOp.getOutput().getType());
+  auto scaleValue = arith::ConstantIndexOp::create(rewriter, loc, scale).getResult();
-  auto outputBuffer = createEmptyTensorFromShaped(rewriter, receiveTensorBatchOp.getLoc(), outputType);
+  return arith::MulIOp::create(rewriter, loc, base, scaleValue).getResult();
-  Value received = pim::PimReceiveTensorBatchOp::create(rewriter,
+}
-                                                        receiveTensorBatchOp.getLoc(),
+
-                                                        outputBuffer.getType(),
+static Value createHostTargetOffset(IRRewriter& rewriter,
-                                                        outputBuffer,
+                                    tensor::ParallelInsertSliceOp insertSlice,
-                                                        rewriter.getDenseI32ArrayAttr(sourceCoreIds))
+                                    ShapedType destinationType,
-                     .getOutput();
+                                    IRMapping& mapper) {
-  mapper.map(receiveTensorBatchOp.getOutput(), received);
+  int64_t elementBytes = static_cast<int64_t>(getElementTypeSizeInBytes(destinationType.getElementType()));
+  SmallVector<int64_t> strides(destinationType.getRank(), 1);
+  ArrayRef<int64_t> shape = destinationType.getShape();
+  for (int64_t dim = destinationType.getRank() - 2; dim >= 0; --dim)
+    strides[dim] = strides[dim + 1] * shape[dim + 1];
+
+  Value totalOffset;
+  Location loc = insertSlice.getLoc();
+  for (auto [dim, offset] : llvm::enumerate(insertSlice.getMixedOffsets())) {
+    int64_t scale = strides[dim] * elementBytes;
+    Value scaledOffset;
+    if (auto attr = dyn_cast<Attribute>(offset)) {
+      auto intAttr = dyn_cast<IntegerAttr>(attr);
+      assert(intAttr && "expected integer offset attribute");
+      scaledOffset = arith::ConstantIndexOp::create(rewriter, loc, intAttr.getInt() * scale).getResult();
+    }
+    else {
+      scaledOffset = createScaledIndexValue(rewriter, loc, mapper.lookupOrDefault(cast<Value>(offset)), scale);
+    }
+
+    totalOffset =
+      totalOffset ? arith::AddIOp::create(rewriter, loc, totalOffset, scaledOffset).getResult() : scaledOffset;
+  }
+
+  if (!totalOffset)
+    totalOffset = arith::ConstantIndexOp::create(rewriter, loc, 0).getResult();
+  return totalOffset;
 }
 
 } // namespace
 
-LogicalResult
+LogicalResult raptor::SpatialToPimPass::lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp,
-lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState& state, IRRewriter& rewriter) {
+                                                            IRRewriter& rewriter) {
-  if (computeBatchOp.getNumResults() != 0)
-    return computeBatchOp.emitOpError(
-      "batched Spatial-to-PIM lowering currently requires channelized compute_batch with no results");
-
   Location loc = computeBatchOp.getLoc();
   Block& oldBlock = computeBatchOp.getBody().front();
-  auto oldYield = cast<spatial::SpatYieldOp>(oldBlock.getTerminator());
+  auto oldYield = dyn_cast<spatial::SpatYieldOp>(oldBlock.getTerminator());
-  if (oldYield.getNumOperands() != 0)
+  auto inParallelOp = dyn_cast<spatial::SpatInParallelOp>(oldBlock.getTerminator());
-    return computeBatchOp.emitOpError("batched Spatial-to-PIM lowering currently requires empty spat.yield");
+  if (computeBatchOp.getNumResults() == 0) {
+    if (!oldYield || oldYield.getNumOperands() != 0)
+      return computeBatchOp.emitOpError("resultless compute_batch lowering requires empty spat.yield");
+  }
+  else if (!inParallelOp) {
+    return computeBatchOp.emitOpError(
+      "resultful compute_batch lowering currently requires a spat.in_parallel terminator");
+  }
 
-  SmallVector<int32_t> coreIds = getPimCoreIdsForBatchOp(computeBatchOp, state.nextCoreId);
+  SmallVector<int32_t> coreIds = getPimCoreIdsForBatchOp(computeBatchOp, coreId);
   SmallVector<Value> batchWeights(computeBatchOp.getWeights().begin(), computeBatchOp.getWeights().end());
   SmallVector<Value> batchInputs;
   if (!computeBatchOp.getInputs().empty())
@@ -91,9 +125,22 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
     {static_cast<int>(batchWeights.size()), static_cast<int>(batchInputs.size())});
   coreBatchOp->setAttr(onnx_mlir::kCoreIdsAttrName, rewriter.getDenseI32ArrayAttr(coreIds));
 
+  SmallVector<unsigned> returnOperandIndices;
+  if (computeBatchOp.getNumResults() != 0) {
+    returnOperandIndices.resize(computeBatchOp.getNumResults());
+    for (auto [resultIndex, result] : llvm::enumerate(computeBatchOp.getResults())) {
+      FailureOr<unsigned> returnOperandIndex = getDirectReturnOperandIndex(cast<OpResult>(result));
+      if (failed(returnOperandIndex))
+        return computeBatchOp.emitOpError(
+          "resultful compute_batch lowering currently requires each result to be used directly by func.return");
+      returnOperandIndices[resultIndex] = *returnOperandIndex;
+    }
+  }
+
   SmallVector<Type> blockArgTypes;
   SmallVector<Location> blockArgLocs;
-  for (BlockArgument arg : oldBlock.getArguments()) {
+  unsigned inputArgLimit = 1 + computeBatchOp.getWeights().size() + computeBatchOp.getInputs().size();
+  for (BlockArgument arg : oldBlock.getArguments().take_front(inputArgLimit)) {
     blockArgTypes.push_back(arg.getType());
     blockArgLocs.push_back(arg.getLoc());
   }
@@ -102,7 +149,21 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
 
   IRMapping mapper;
   rewriter.setInsertionPointToStart(newBlock);
-  for (auto [oldArg, newArg] : llvm::zip(oldBlock.getArguments(), newBlock->getArguments())) {
+  auto oldLaneArg = computeBatchOp.getLaneArgument();
+  if (!oldLaneArg)
+    return computeBatchOp.emitOpError("expected compute_batch lane block argument before lowering");
+  mapper.map(*oldLaneArg, coreBatchOp.getLaneArgument());
+  for (unsigned weightIndex = 0; weightIndex < computeBatchOp.getWeights().size(); ++weightIndex) {
+    auto oldWeightArg = computeBatchOp.getWeightArgument(weightIndex);
+    if (!oldWeightArg)
+      return computeBatchOp.emitOpError("expected compute_batch weight block arguments before lowering");
+    mapper.map(*oldWeightArg, coreBatchOp.getWeightArgument(weightIndex));
+  }
+  for (unsigned inputIndex = 0; inputIndex < computeBatchOp.getInputs().size(); ++inputIndex) {
+    auto oldArg = computeBatchOp.getInputArgument(inputIndex);
+    if (!oldArg)
+      return computeBatchOp.emitOpError("expected compute_batch input block arguments before lowering");
+    BlockArgument newArg = coreBatchOp.getInputArgument(inputIndex);
     auto newArgType = cast<ShapedType>(newArg.getType());
     auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, newArgType);
     auto copied = pim::PimMemCopyHostToDevBatchOp::create(rewriter,
@@ -114,26 +175,17 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
                                                           rewriter.getI32IntegerAttr(0),
                                                           getTensorSizeInBytesAttr(rewriter, newArg))
                     .getOutput();
-    mapper.map(oldArg, copied);
+    mapper.map(*oldArg, copied);
   }
 
-  auto materializeCapturedTensor = [&](Value capturedTensor) -> Value {
+  SmallVector<Value> hostOutputTensors(returnOperandIndices.size());
-    if (auto mapped = mapper.lookupOrNull(capturedTensor))
+  auto getOrCreateHostOutputTensor = [&](unsigned resultIndex, Location resultLoc) -> Value {
-      return mapped;
+    Value& hostOutputTensor = hostOutputTensors[resultIndex];
+    if (hostOutputTensor)
+      return hostOutputTensor;
 
-    auto capturedType = cast<ShapedType>(capturedTensor.getType());
+    hostOutputTensor = outputTensors[returnOperandIndices[resultIndex]](rewriter, resultLoc);
-    auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, capturedType);
+    return hostOutputTensor;
-    auto copied = pim::PimMemCopyHostToDevBatchOp::create(rewriter,
-                                                          loc,
-                                                          outputBuffer.getType(),
-                                                          outputBuffer,
-                                                          capturedTensor,
-                                                          rewriter.getI32IntegerAttr(0),
-                                                          rewriter.getI32IntegerAttr(0),
-                                                          getTensorSizeInBytesAttr(rewriter, capturedTensor))
-                    .getOutput();
-    mapper.map(capturedTensor, copied);
-    return copied;
   };
 
   rewriter.setInsertionPointToEnd(newBlock);
@@ -141,36 +193,37 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
     if (isa<spatial::SpatYieldOp>(op))
       continue;
 
-    if (auto sendBatchOp = dyn_cast<spatial::SpatChannelSendBatchOp>(op)) {
+    if (auto parallelOp = dyn_cast<spatial::SpatInParallelOp>(op)) {
-      pim::PimSendBatchOp::create(rewriter,
+      auto firstOutputArg = computeBatchOp.getOutputArgument(0);
-                                  loc,
+      if (!firstOutputArg)
-                                  mapper.lookup(sendBatchOp.getInput()),
+        return computeBatchOp.emitOpError("expected compute_batch output block arguments before lowering");
-                                  getTensorSizeInBytesAttr(rewriter, mapper.lookup(sendBatchOp.getInput())),
+      for (Operation& nestedOp : parallelOp.getRegion().front()) {
-                                  sendBatchOp.getTargetCoreIdsAttr());
+        auto insertSlice = dyn_cast<tensor::ParallelInsertSliceOp>(&nestedOp);
-      continue;
+        if (!insertSlice)
-    }
+          return parallelOp.emitOpError("expected only tensor.parallel_insert_slice in spat.in_parallel");
 
-    if (auto sendTensorBatchOp = dyn_cast<spatial::SpatChannelSendTensorBatchOp>(op)) {
+        auto outputArg = dyn_cast<BlockArgument>(insertSlice.getDest());
-      lowerChannelSendTensorBatch(sendTensorBatchOp, mapper, rewriter);
+        if (!outputArg || outputArg.getOwner() != &oldBlock)
-      continue;
+          return insertSlice.emitOpError("expected compute_batch output block argument destination");
-    }
 
-    if (auto receiveBatchOp = dyn_cast<spatial::SpatChannelReceiveBatchOp>(op)) {
+        unsigned resultIndex = outputArg.getArgNumber() - firstOutputArg->getArgNumber();
-      auto outputType = cast<ShapedType>(receiveBatchOp.getOutput().getType());
+        if (resultIndex >= returnOperandIndices.size())
-      auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, outputType);
+          return insertSlice.emitOpError("result index out of range while lowering host batch output");
-      auto received = pim::PimReceiveBatchOp::create(rewriter,
-                                                     loc,
-                                                     outputBuffer.getType(),
-                                                     outputBuffer,
-                                                     getTensorSizeInBytesAttr(rewriter, receiveBatchOp.getOutput()),
-                                                     receiveBatchOp.getSourceCoreIdsAttr())
-                        .getOutput();
-      mapper.map(receiveBatchOp.getOutput(), received);
-      continue;
-    }
 
-    if (auto receiveTensorBatchOp = dyn_cast<spatial::SpatChannelReceiveTensorBatchOp>(op)) {
+        Value mappedSource = mapper.lookup(insertSlice.getSource());
-      lowerChannelReceiveTensorBatch(receiveTensorBatchOp, mapper, rewriter);
+        Value hostTarget = getOrCreateHostOutputTensor(resultIndex, insertSlice.getLoc());
+        auto hostTargetType = cast<ShapedType>(hostTarget.getType());
+        Value hostTargetOffset = createHostTargetOffset(rewriter, insertSlice, hostTargetType, mapper);
+        Value zeroOffset = arith::ConstantIndexOp::create(rewriter, insertSlice.getLoc(), 0).getResult();
+        pim::PimMemCopyDevToHostOp::create(rewriter,
+                                           insertSlice.getLoc(),
+                                           hostTarget.getType(),
+                                           hostTargetOffset,
+                                           zeroOffset,
+                                           hostTarget,
+                                           mappedSource,
+                                           getTensorSizeInBytesAttr(rewriter, mappedSource));
+      }
       continue;
     }
 
@@ -178,6 +231,10 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
       if (isa_and_present<memref::GetGlobalOp>(toTensorOp.getBuffer().getDefiningOp())) {
         Operation* cloned = rewriter.clone(op, mapper);
         auto clonedTensor = cloned->getResult(0);
+        if (isUsedOnlyAsExplicitHostOperand(toTensorOp.getResult())) {
+          mapper.map(toTensorOp.getResult(), clonedTensor);
+          continue;
+        }
         auto clonedType = cast<ShapedType>(clonedTensor.getType());
         auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, clonedType);
         auto copied = pim::PimMemCopyHostToDevBatchOp::create(rewriter,
@@ -194,15 +251,18 @@ lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState&
       }
     }
 
-    for (Value operand : op.getOperands()) {
+    for (auto [operandIndex, operand] : llvm::enumerate(op.getOperands())) {
       if (!isa<TensorType>(operand.getType()) || mapper.contains(operand))
         continue;
+      if (isExplicitHostOperand(&op, operandIndex))
+        continue;
 
       Operation* definingOp = operand.getDefiningOp();
       if (definingOp && definingOp->getBlock() == &oldBlock)
         continue;
 
-      materializeCapturedTensor(operand);
+      return computeBatchOp.emitOpError(
+        "expected external tensor communication to be materialized in Spatial before batch lowering");
     }
 
     Operation* cloned = rewriter.clone(op, mapper);

src/PIM/Conversion/SpatialToPim/BatchCoreLoweringPatterns.hpp

@@ -1,10 +0,0 @@
-#pragma once
-
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/CoreLoweringPatterns.hpp"
-
-namespace onnx_mlir {
-
-mlir::LogicalResult
-lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, CoreLoweringState& state, mlir::IRRewriter& rewriter);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/CMakeLists.txt

@@ -3,17 +3,17 @@ mlir_tablegen(SpatialToPim.hpp.inc -gen-rewriters "-I${ONNX_MLIR_SRC_ROOT}")
 add_public_tablegen_target(SpatialToPimIncGen)
 
 add_pim_library(OMSpatialToPim
+  Patterns.cpp
   SpatialToPimPass.cpp
   BatchCoreLoweringPatterns.cpp
-  ChannelLoweringPatterns.cpp
-  Cleanup.cpp
   Common.cpp
   ComputeLikeRegionUtils.cpp
   CoreLoweringPatterns.cpp
-  GlobalTensorMaterialization.cpp
-  PhaseVerification.cpp
   ReturnPathNormalization.cpp
-  TensorPackingPatterns.cpp
+  Patterns/ChannelLowering.cpp
+  Patterns/GlobalTensorMaterialization.cpp
+  Patterns/TensorPacking.cpp
+  Patterns/Transpose.cpp
 
   EXCLUDE_FROM_OM_LIBS
 
@@ -21,7 +21,10 @@ add_pim_library(OMSpatialToPim
   SpatialToPimIncGen
 
   LINK_LIBS PUBLIC
+  MLIRLinalgDialect
   MLIRSCFDialect
+  MLIRSCFUtils
+  MLIRTransformUtils
   MLIRTosaDialect
   OMCompilerOptions
   OMPimCommon

src/PIM/Conversion/SpatialToPim/ChannelLoweringPatterns.hpp

@@ -1,9 +0,0 @@
-#pragma once
-
-#include "mlir/IR/PatternMatch.h"
-
-namespace onnx_mlir {
-
-void populateChannelLoweringPatterns(mlir::RewritePatternSet& patterns);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Cleanup.cpp

@@ -1,42 +0,0 @@
-#include "llvm/ADT/STLExtras.h"
-
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/Cleanup.hpp"
-
-using namespace mlir;
-
-namespace onnx_mlir {
-
-LogicalResult erasePendingOps(SmallVectorImpl<Operation*>& pendingOps, IRRewriter& rewriter) {
-  while (!pendingOps.empty()) {
-    bool erasedAnyOp = false;
-    for (auto it = pendingOps.begin(); it != pendingOps.end();) {
-      Operation* opToRemove = *it;
-      if (!opToRemove->use_empty()) {
-        ++it;
-        continue;
-      }
-
-      rewriter.eraseOp(opToRemove);
-      it = pendingOps.erase(it);
-      erasedAnyOp = true;
-    }
-
-    if (erasedAnyOp)
-      continue;
-
-    for (Operation* opToRemove : pendingOps) {
-      InFlightDiagnostic diag = opToRemove->emitError("pending Spatial-to-PIM cleanup could not erase operation");
-      diag << "; op has " << llvm::range_size(opToRemove->getUsers()) << " remaining user(s)";
-      for (Operation* user : opToRemove->getUsers()) {
-        bool userPendingRemoval = llvm::is_contained(pendingOps, user);
-        opToRemove->emitRemark() << "remaining user `" << user->getName() << "`"
-                                 << (userPendingRemoval ? " is also pending removal" : " is not pending removal");
-      }
-    }
-    return failure();
-  }
-
-  return success();
-}
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Cleanup.hpp

@@ -1,11 +0,0 @@
-#pragma once
-
-#include "mlir/IR/Operation.h"
-#include "mlir/IR/PatternMatch.h"
-#include "mlir/Support/LLVM.h"
-
-namespace onnx_mlir {
-
-mlir::LogicalResult erasePendingOps(llvm::SmallVectorImpl<mlir::Operation*>& pendingOps, mlir::IRRewriter& rewriter);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Common.cpp

@@ -55,10 +55,6 @@ size_t getSliceActualOffset(tensor::ExtractSliceOp& sliceOp, ShapedType& inputSh
   return returnValue;
 }
 
-size_t getShapedTypeSizeInBytes(ShapedType shapedType) {
-  return shapedType.getNumElements() * shapedType.getElementTypeBitWidth() / 8;
-}
-
 IntegerAttr getTensorSizeInBytesAttr(Builder& builder, mlir::Value value) {
   return builder.getI32IntegerAttr(static_cast<int32_t>(getShapedTypeSizeInBytes(cast<ShapedType>(value.getType()))));
 }

src/PIM/Conversion/SpatialToPim/Common.hpp

@@ -20,8 +20,6 @@ namespace onnx_mlir {
  */
 size_t getSliceActualOffset(mlir::tensor::ExtractSliceOp& sliceOp, mlir::ShapedType& inputShape);
 
-size_t getShapedTypeSizeInBytes(mlir::ShapedType shapedType);
-
 mlir::IntegerAttr getTensorSizeInBytesAttr(mlir::Builder& builder, mlir::Value value);
 
 template <class T>

src/PIM/Conversion/SpatialToPim/ComputeLikeRegionUtils.cpp

@@ -1,3 +1,5 @@
+#include <cassert>
+
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/ComputeLikeRegionUtils.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 
@@ -29,7 +31,18 @@ void replaceAndEraseDirectComputeLikeInput(PatternRewriter& rewriter,
                                            unsigned inputIndex,
                                            Value replacement) {
   Block& body = owner->getRegion(0).front();
-  BlockArgument bodyArgument = body.getArgument(inputIndex);
+  BlockArgument bodyArgument;
+  if (auto compute = dyn_cast<spatial::SpatCompute>(owner)) {
+    auto computeArg = compute.getInputArgument(inputIndex);
+    assert(computeArg && "expected compute input block argument");
+    bodyArgument = *computeArg;
+  }
+  else {
+    auto batchArg = cast<spatial::SpatComputeBatch>(owner).getInputArgument(inputIndex);
+    assert(batchArg && "expected compute_batch input block argument");
+    bodyArgument = *batchArg;
+  }
+  unsigned bodyArgIndex = bodyArgument.getArgNumber();
 
   rewriter.startOpModification(owner);
   bodyArgument.replaceAllUsesWith(replacement);
@@ -37,7 +50,7 @@ void replaceAndEraseDirectComputeLikeInput(PatternRewriter& rewriter,
     compute.getInputsMutable().erase(inputIndex);
   else
     cast<spatial::SpatComputeBatch>(owner).getInputsMutable().erase(inputIndex);
-  body.eraseArgument(inputIndex);
+  body.eraseArgument(bodyArgIndex);
   rewriter.finalizeOpModification(owner);
 }
 

src/PIM/Conversion/SpatialToPim/CoreLoweringPatterns.cpp

@@ -1,13 +1,15 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Tosa/IR/TosaOps.h"
 #include "mlir/IR/IRMapping.h"
+#include "mlir/IR/Matchers.h"
 
 #include "Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "Conversion/SpatialToPim/SpatialToPimPass.hpp"
 #include "src/Accelerators/PIM/Common/PimCommon.hpp"
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/CoreLoweringPatterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -23,11 +25,12 @@ static bool isChannelUseChainOp(Operation* op) {
              tensor::ExpandShapeOp,
              tensor::CastOp,
              tosa::ReshapeOp,
-             ONNXTransposeOp,
+             linalg::TransposeOp,
              pim::PimTransposeOp>(op);
 }
 
-static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewriter& rewriter) {
+static void
+cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewriter& rewriter, OperationFolder& constantFolder) {
   for (Value operand : op->getOperands()) {
     if (mapping.lookupOrNull(operand))
       continue;
@@ -36,7 +39,12 @@ static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewri
     if (!definingOp)
       continue;
 
-    if (!isa<tensor::EmptyOp, arith::ConstantOp>(definingOp))
+    if (auto constantOp = dyn_cast<arith::ConstantOp>(definingOp)) {
+      mapping.map(operand, getOrCreateHostConstantLike(constantOp, constantFolder));
+      continue;
+    }
+
+    if (!isa<tensor::EmptyOp>(definingOp))
       continue;
 
     Operation* clonedOp = rewriter.clone(*definingOp, mapping);
@@ -46,8 +54,6 @@ static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewri
   }
 }
 
-static int32_t translateSpatialCoreIdToPimCoreId(size_t spatialCoreId) { return static_cast<int32_t>(spatialCoreId); }
-
 static int32_t getPimCoreIdForComputeOp(spatial::SpatCompute computeOp, size_t& fallbackCoreId) {
   if (auto spatialCoreIdAttr = computeOp->getAttrOfType<IntegerAttr>(onnx_mlir::kCoreIdAttrName))
     return static_cast<int32_t>(spatialCoreIdAttr.getInt());
@@ -92,7 +98,9 @@ static LogicalResult collectHelperComputeChain(spatial::SpatCompute computeOp,
   return success();
 }
 
-static bool inlineInputlessHelperComputeForWeightLikeUsers(spatial::SpatCompute computeOp, IRRewriter& rewriter) {
+static bool inlineInputlessHelperComputeForWeightLikeUsers(spatial::SpatCompute computeOp,
+                                                           IRRewriter& rewriter,
+                                                           OperationFolder& constantFolder) {
   if (!computeOp.getInputs().empty() || computeOp.getNumResults() != 1)
     return false;
   if (!llvm::all_of(computeOp.getResult(0).getUsers(), [](Operation* user) {
@@ -101,7 +109,7 @@ static bool inlineInputlessHelperComputeForWeightLikeUsers(spatial::SpatCompute
     return false;
 
   Block& block = computeOp.getBody().front();
-  if (block.getNumArguments() != 0)
+  if (block.getNumArguments() != computeOp.getWeights().size())
     return false;
 
   auto yieldOp = dyn_cast<spatial::SpatYieldOp>(block.getTerminator());
@@ -110,8 +118,14 @@ static bool inlineInputlessHelperComputeForWeightLikeUsers(spatial::SpatCompute
 
   rewriter.setInsertionPoint(computeOp);
   IRMapping mapping;
+  for (auto [weightIndex, weight] : llvm::enumerate(computeOp.getWeights())) {
+    auto weightArg = computeOp.getWeightArgument(weightIndex);
+    if (!weightArg)
+      return false;
+    mapping.map(*weightArg, weight);
+  }
   for (Operation& op : block.without_terminator()) {
-    cloneMappedHelperOperands(&op, mapping, rewriter);
+    cloneMappedHelperOperands(&op, mapping, rewriter, constantFolder);
     Operation* clonedOp = rewriter.clone(op, mapping);
     for (auto [originalResult, newResult] : llvm::zip(op.getResults(), clonedOp->getResults()))
       mapping.map(originalResult, newResult);
@@ -125,15 +139,12 @@ static bool inlineInputlessHelperComputeForWeightLikeUsers(spatial::SpatCompute
 
 } // namespace
 
-void markOpToRemove(CoreLoweringState& state, Operation* op) {
+LogicalResult raptor::SpatialToPimPass::lowerComputeOp(spatial::SpatCompute computeOp,
-  if (!llvm::is_contained(state.operationsToRemove, op))
+                                                       IRRewriter& rewriter,
-    state.operationsToRemove.push_back(op);
+                                                       OperationFolder& constantFolder) {
-}
-
-LogicalResult lowerComputeOp(spatial::SpatCompute computeOp, CoreLoweringState& state, IRRewriter& rewriter) {
   Location loc = computeOp->getLoc();
 
-  if (inlineInputlessHelperComputeForWeightLikeUsers(computeOp, rewriter))
+  if (inlineInputlessHelperComputeForWeightLikeUsers(computeOp, rewriter, constantFolder))
     return success();
 
   SmallVector<Operation*> helperChain;
@@ -143,21 +154,24 @@ LogicalResult lowerComputeOp(spatial::SpatCompute computeOp, CoreLoweringState&
   auto& block = computeOp.getRegion().front();
   auto yieldOp = cast<spatial::SpatYieldOp>(block.getTerminator());
 
-  for (auto [argIndex, blockArg] : llvm::enumerate(block.getArguments())) {
+  for (auto [inputIndex, input] : llvm::enumerate(computeOp.getInputs())) {
-    auto receiveOp = dyn_cast_or_null<spatial::SpatChannelReceiveOp>(computeOp.getInputs()[argIndex].getDefiningOp());
+    auto blockArg = computeOp.getInputArgument(inputIndex);
-    if (!receiveOp || blockArg.use_empty())
+    if (!blockArg)
-      continue;
+      return computeOp.emitOpError("expected compute input block arguments during lowering");
-
+    auto receiveOp = dyn_cast_or_null<spatial::SpatChannelReceiveOp>(input.getDefiningOp());
-    rewriter.setInsertionPoint(getEarliestUserWithinBlock(blockArg));
+    if (receiveOp && !blockArg->use_empty()) {
-    auto outputType = cast<ShapedType>(blockArg.getType());
+      rewriter.setInsertionPoint(getEarliestUserWithinBlock(*blockArg));
+      auto outputType = cast<ShapedType>(blockArg->getType());
       auto outputBuffer = createEmptyTensorFromShaped(rewriter, receiveOp.getLoc(), outputType);
-    auto sizeAttr = getTensorSizeInBytesAttr(rewriter, blockArg);
+      auto sizeAttr = getTensorSizeInBytesAttr(rewriter, *blockArg);
-    auto sourceCoreIdAttr = rewriter.getI32IntegerAttr(translateSpatialCoreIdToPimCoreId(receiveOp.getSourceCoreId()));
+      Value received =
-    Value received = PimReceiveOp::create(
+        PimReceiveOp::create(
-                       rewriter, receiveOp.getLoc(), outputBuffer.getType(), outputBuffer, sizeAttr, sourceCoreIdAttr)
+          rewriter, receiveOp.getLoc(), outputBuffer.getType(), outputBuffer, sizeAttr, receiveOp.getSourceCoreId())
           .getOutput();
-    blockArg.replaceAllUsesWith(received);
+      blockArg->replaceAllUsesWith(received);
-    markOpToRemove(state, receiveOp);
+      markOpToRemove(receiveOp);
+      continue;
+    }
   }
 
   if (computeOp.getNumResults() != yieldOp.getNumOperands())
@@ -167,9 +181,8 @@ LogicalResult lowerComputeOp(spatial::SpatCompute computeOp, CoreLoweringState&
     if (result.use_empty())
       continue;
 
-    ReturnPathState returnPathState {state.outputTensors, state.operationsToRemove};
     ReturnPathLoweringResult returnPathResult =
-      lowerComputeResultReturnPath(computeOp, cast<OpResult>(result), yieldValue, returnPathState, rewriter);
+      lowerComputeResultReturnPath(computeOp, cast<OpResult>(result), yieldValue, rewriter);
     if (returnPathResult == ReturnPathLoweringResult::Failure)
       return failure();
     if (returnPathResult == ReturnPathLoweringResult::Handled)
@@ -193,15 +206,40 @@ LogicalResult lowerComputeOp(spatial::SpatCompute computeOp, CoreLoweringState&
   if (!computeOp.getWeights().empty())
     computeWeights.append(computeOp.getWeights().begin(), computeOp.getWeights().end());
   rewriter.setInsertionPointAfter(computeOp);
-  auto coreOp = PimCoreOp::create(rewriter,
+  auto coreOp = PimCoreOp::create(
-                                  loc,
+    rewriter, loc, ValueRange(computeWeights), rewriter.getI32IntegerAttr(getPimCoreIdForComputeOp(computeOp, coreId)));
-                                  ValueRange(computeWeights),
+  rewriter.setInsertionPointToStart(&block);
-                                  rewriter.getI32IntegerAttr(getPimCoreIdForComputeOp(computeOp, state.nextCoreId)));
   auto& coreOpBlocks = coreOp.getBody().getBlocks();
-  for (auto [argIndex, blockArg] : llvm::enumerate(block.getArguments()))
+  for (auto [inputIndex, input] : llvm::enumerate(computeOp.getInputs())) {
-    if (!blockArg.use_empty())
+    auto blockArg = computeOp.getInputArgument(inputIndex);
-      blockArg.replaceAllUsesWith(computeOp.getInputs()[argIndex]);
+    if (!blockArg)
-  block.eraseArguments(0, block.getNumArguments());
+      return computeOp.emitOpError("expected compute input block arguments during input materialization");
+    if (blockArg->use_empty())
+      continue;
+
+    if (auto constantOp = input.getDefiningOp<arith::ConstantOp>()) {
+      blockArg->replaceAllUsesWith(getOrCreateHostConstantLike(constantOp, constantFolder));
+      continue;
+    }
+
+    auto inputType = dyn_cast<ShapedType>(input.getType());
+    if (!inputType)
+      return computeOp.emitOpError("expected shaped compute input during pim.core lowering");
+    auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, inputType);
+    auto copied =
+      PimMemCopyHostToDevOp::create(rewriter,
+                                    loc,
+                                    outputBuffer.getType(),
+                                    getOrCreateHostIndexConstant(outputBuffer.getOperation(), 0, constantFolder),
+                                    getOrCreateHostIndexConstant(outputBuffer.getOperation(), 0, constantFolder),
+                                    outputBuffer,
+                                    input,
+                                    getTensorSizeInBytesAttr(rewriter, input))
+        .getOutput();
+    blockArg->replaceAllUsesWith(copied);
+  }
+  if (!computeOp.getInputs().empty())
+    block.eraseArguments(computeOp.getWeights().size(), computeOp.getInputs().size());
   coreOpBlocks.splice(coreOpBlocks.begin(), computeOp.getBody().getBlocks());
   Block* tempComputeBlock = new Block();
   computeOp.getBody().push_back(tempComputeBlock);

src/PIM/Conversion/SpatialToPim/CoreLoweringPatterns.hpp

@@ -1,21 +0,0 @@
-#pragma once
-
-#include "mlir/IR/PatternMatch.h"
-
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/ReturnPathNormalization.hpp"
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-
-namespace onnx_mlir {
-
-struct CoreLoweringState {
-  size_t& nextCoreId;
-  llvm::SmallVectorImpl<OutputTensorFactory>& outputTensors;
-  llvm::SmallVectorImpl<mlir::Operation*>& operationsToRemove;
-};
-
-void markOpToRemove(CoreLoweringState& state, mlir::Operation* op);
-
-mlir::LogicalResult
-lowerComputeOp(spatial::SpatCompute computeOp, CoreLoweringState& state, mlir::IRRewriter& rewriter);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/GlobalTensorMaterialization.hpp

@@ -1,9 +0,0 @@
-#pragma once
-
-#include "mlir/IR/PatternMatch.h"
-
-namespace onnx_mlir {
-
-void populateGlobalTensorMaterializationPatterns(mlir::RewritePatternSet& patterns);
-
-}

src/PIM/Conversion/SpatialToPim/Patterns.cpp

@@ -0,0 +1,40 @@
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Patterns.hpp"
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace raptor {
+
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/SpatialToPim.hpp.inc"
+
+} // namespace raptor
+
+void populateInitialPatterns(RewritePatternSet& patterns) {
+  raptor::populateWithGenerated(patterns);
+  populateTransposeLoweringPatterns(patterns);
+}
+
+void populateGlobalTensorMaterializationPatternPhase(RewritePatternSet& patterns) {
+  populateGlobalTensorMaterializationPatterns(patterns);
+}
+
+void populateInitialTensorPackingPatterns(RewritePatternSet& patterns) {
+  populateTensorPackingPatterns(patterns);
+}
+
+void populateCoreBodyPatterns(RewritePatternSet& patterns) {
+  raptor::populateWithGenerated(patterns);
+  populateTransposeLoweringPatterns(patterns);
+}
+
+void populateFinalTensorPackingPatterns(RewritePatternSet& patterns) {
+  populateTensorPackingPatterns(patterns);
+}
+
+void populateCommunicationPatterns(RewritePatternSet& patterns) {
+  populateChannelLoweringPatterns(patterns);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Patterns.hpp

@@ -8,6 +8,18 @@
 
 namespace onnx_mlir {
 
+void populateInitialPatterns(mlir::RewritePatternSet& patterns);
+void populateGlobalTensorMaterializationPatternPhase(mlir::RewritePatternSet& patterns);
+void populateInitialTensorPackingPatterns(mlir::RewritePatternSet& patterns);
+void populateCoreBodyPatterns(mlir::RewritePatternSet& patterns);
+void populateFinalTensorPackingPatterns(mlir::RewritePatternSet& patterns);
+void populateCommunicationPatterns(mlir::RewritePatternSet& patterns);
+
+void populateTransposeLoweringPatterns(mlir::RewritePatternSet& patterns);
+void populateChannelLoweringPatterns(mlir::RewritePatternSet& patterns);
+void populateGlobalTensorMaterializationPatterns(mlir::RewritePatternSet& patterns);
+void populateTensorPackingPatterns(mlir::RewritePatternSet& patterns);
+
 mlir::RankedTensorType getPackedTensorType(mlir::RankedTensorType elementType, int64_t count);
 mlir::Value extractPackedChunk(mlir::Value packedValue,
                                mlir::RankedTensorType chunkType,
@@ -20,7 +32,6 @@ mlir::Value createPackedExtractRowsSlice(spatial::SpatExtractRowsOp extractRowsO
                                          mlir::OpBuilder& builder,
                                          mlir::Location loc);
 mlir::Value createPackedExtractSliceTensor(mlir::ValueRange values, mlir::OpBuilder& builder, mlir::Location loc);
-void populateTensorPackingPatterns(mlir::RewritePatternSet& patterns);
 void eraseUnusedTensorPackingOps(mlir::func::FuncOp funcOp, mlir::IRRewriter& rewriter);
 
 } // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Patterns/ChannelLowering.cpp

@@ -1,6 +1,6 @@
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/ChannelLoweringPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Patterns.hpp"
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
@@ -10,17 +10,12 @@ using namespace mlir;
 namespace onnx_mlir {
 namespace {
 
-static int32_t toPimCoreId(int32_t spatialCoreId) { return spatialCoreId; }
-
 struct ChannelSendLowering : OpRewritePattern<spatial::SpatChannelSendOp> {
   using OpRewritePattern::OpRewritePattern;
 
   LogicalResult matchAndRewrite(spatial::SpatChannelSendOp op, PatternRewriter& rewriter) const override {
-    pim::PimSendOp::create(rewriter,
+    pim::PimSendOp::create(
-                           op.getLoc(),
+      rewriter, op.getLoc(), op.getInput(), getTensorSizeInBytesAttr(rewriter, op.getInput()), op.getTargetCoreId());
-                           op.getInput(),
-                           getTensorSizeInBytesAttr(rewriter, op.getInput()),
-                           rewriter.getI32IntegerAttr(toPimCoreId(op.getTargetCoreId())));
     rewriter.eraseOp(op);
     return success();
   }
@@ -42,41 +37,7 @@ struct ChannelReceiveLowering : OpRewritePattern<spatial::SpatChannelReceiveOp>
                                                op.getResult().getType(),
                                                outputBuffer,
                                                getTensorSizeInBytesAttr(rewriter, op.getResult()),
-                                               rewriter.getI32IntegerAttr(toPimCoreId(op.getSourceCoreId())))
+                                               op.getSourceCoreId())
-                       .getOutput();
-    rewriter.replaceOp(op, received);
-    return success();
-  }
-};
-
-struct ChannelSendTensorLowering : OpRewritePattern<spatial::SpatChannelSendTensorOp> {
-  using OpRewritePattern::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(spatial::SpatChannelSendTensorOp op, PatternRewriter& rewriter) const override {
-    SmallVector<int32_t> targetCoreIds;
-    targetCoreIds.reserve(op.getTargetCoreIds().size());
-    for (int32_t targetCoreId : op.getTargetCoreIds())
-      targetCoreIds.push_back(toPimCoreId(targetCoreId));
-    pim::PimSendTensorOp::create(rewriter, op.getLoc(), op.getInput(), rewriter.getDenseI32ArrayAttr(targetCoreIds));
-    rewriter.eraseOp(op);
-    return success();
-  }
-};
-
-struct ChannelReceiveTensorLowering : OpRewritePattern<spatial::SpatChannelReceiveTensorOp> {
-  using OpRewritePattern::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(spatial::SpatChannelReceiveTensorOp op, PatternRewriter& rewriter) const override {
-    SmallVector<int32_t> sourceCoreIds;
-    sourceCoreIds.reserve(op.getSourceCoreIds().size());
-    for (int32_t sourceCoreId : op.getSourceCoreIds())
-      sourceCoreIds.push_back(toPimCoreId(sourceCoreId));
-    auto outputType = cast<ShapedType>(op.getOutput().getType());
-    Value outputBuffer =
-      tensor::EmptyOp::create(rewriter, op.getLoc(), outputType.getShape(), outputType.getElementType()).getResult();
-    Value received =
-      pim::PimReceiveTensorOp::create(
-        rewriter, op.getLoc(), op.getOutput().getType(), outputBuffer, rewriter.getDenseI32ArrayAttr(sourceCoreIds))
                        .getOutput();
     rewriter.replaceOp(op, received);
     return success();
@@ -125,12 +86,7 @@ struct ConcatLowering : OpRewritePattern<spatial::SpatConcatOp> {
 } // namespace
 
 void populateChannelLoweringPatterns(RewritePatternSet& patterns) {
-  patterns.add<ChannelSendLowering,
+  patterns.add<ChannelSendLowering, ChannelReceiveLowering, ExtractRowsLowering, ConcatLowering>(patterns.getContext());
-               ChannelReceiveLowering,
-               ChannelSendTensorLowering,
-               ChannelReceiveTensorLowering,
-               ExtractRowsLowering,
-               ConcatLowering>(patterns.getContext());
 }
 
 } // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Patterns/GlobalTensorMaterialization.cpp

@@ -16,7 +16,7 @@
 
 #include "Common/PimCommon.hpp"
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/ComputeLikeRegionUtils.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/GlobalTensorMaterialization.hpp"
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 
 using namespace mlir;
@@ -76,10 +76,11 @@ struct MoveExtractSliceIntoCompute final : OpRewritePattern<mlir::tensor::Extrac
         auto inputIndex = getDirectComputeLikeInputIndex(spatCompute, uses.getOperandNumber());
         if (!inputIndex)
           return failure();
-        auto BBArgIndex = *inputIndex;
+        auto BBArgValue = spatCompute.getInputArgument(*inputIndex);
-        auto BBArgValue = spatCompute.getBody().front().getArgument(BBArgIndex);
+        if (!BBArgValue)
+          return failure();
 
-        if (BBArgValue.use_empty())
+        if (BBArgValue->use_empty())
           continue;
 
         rewriter.setInsertionPoint(&spatCompute.getBody().front().front());
@@ -89,16 +90,17 @@ struct MoveExtractSliceIntoCompute final : OpRewritePattern<mlir::tensor::Extrac
         }
 
         replaceAndEraseDirectComputeLikeInput(
-          rewriter, spatCompute.getOperation(), BBArgIndex, mapSpatToExtract[spatCompute.getOperation()]);
+          rewriter, spatCompute.getOperation(), *inputIndex, mapSpatToExtract[spatCompute.getOperation()]);
       }
       else if (auto spatComputeBatch = dyn_cast<spatial::SpatComputeBatch>(uses.getOwner())) {
         auto inputIndex = getDirectComputeLikeInputIndex(spatComputeBatch, uses.getOperandNumber());
         if (!inputIndex)
           return failure();
-        auto BBArgIndex = *inputIndex;
+        auto BBArgValue = spatComputeBatch.getInputArgument(*inputIndex);
-        auto BBArgValue = spatComputeBatch.getBody().front().getArgument(BBArgIndex);
+        if (!BBArgValue)
+          return failure();
 
-        if (BBArgValue.use_empty())
+        if (BBArgValue->use_empty())
           continue;
 
         rewriter.setInsertionPoint(&spatComputeBatch.getBody().front().front());
@@ -108,7 +110,7 @@ struct MoveExtractSliceIntoCompute final : OpRewritePattern<mlir::tensor::Extrac
         }
 
         replaceAndEraseDirectComputeLikeInput(
-          rewriter, spatComputeBatch.getOperation(), BBArgIndex, mapSpatToExtract[spatComputeBatch.getOperation()]);
+          rewriter, spatComputeBatch.getOperation(), *inputIndex, mapSpatToExtract[spatComputeBatch.getOperation()]);
       }
       else {
         {
@@ -143,170 +145,6 @@ struct MoveExtractSliceIntoCompute final : OpRewritePattern<mlir::tensor::Extrac
   }
 };
 
-// Turns runtime constants consumed by compute regions into private globals and local loads.
-struct ArithConstToGlobalMemoryPattern final : OpRewritePattern<mlir::arith::ConstantOp> {
-  using OpRewritePattern::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(mlir::arith::ConstantOp constantOp, PatternRewriter& rewriter) const override {
-    Location loc = constantOp.getLoc();
-
-    if (hasWeightAlways(constantOp))
-      return failure();
-
-    if (!isa<func::FuncOp>(constantOp->getParentOp()))
-      return failure();
-
-    if (llvm::all_of(constantOp->getUsers(), [](Operation* op) {
-          if (isa<spatial::SpatCompute>(op))
-            return false;
-          if (isa<func::FuncOp>(op->getParentOp()))
-            return true;
-          return false;
-        }))
-      return failure();
-
-    rewriter.setInsertionPoint(constantOp->getParentOfType<func::FuncOp>());
-
-    auto constRankedTensorType = llvm::dyn_cast<mlir::RankedTensorType>(constantOp.getType());
-
-    if (constRankedTensorType) {
-      mlir::MemRefType memRefType =
-        mlir::MemRefType::get(constRankedTensorType.getShape(), constRankedTensorType.getElementType());
-      auto globalOp = createPrivateMemrefGlobalWithUniqueName(rewriter,
-                                                              loc,
-                                                              constantOp->getParentOfType<ModuleOp>(),
-                                                              "const",
-                                                              memRefType,
-                                                              constantOp.getValueAttr(),
-                                                              rewriter.getUnitAttr());
-      std::string argName = globalOp.getSymName().str();
-
-      llvm::DenseMap<Operation*, Value> mapSpatComputeToConst;
-
-      for (auto& constUses : llvm::make_early_inc_range(constantOp->getUses())) {
-        auto constUsers = constUses.getOwner();
-
-        if (auto spatCompute = llvm::dyn_cast<spatial::SpatCompute>(constUsers)) {
-          auto inputIndex = getDirectComputeLikeInputIndex(spatCompute, constUses.getOperandNumber());
-          if (!inputIndex)
-            return failure();
-          auto BBArgIndex = *inputIndex;
-          rewriter.setInsertionPoint(&spatCompute.getBody().front().front());
-          if (!mapSpatComputeToConst.contains(spatCompute.getOperation())) {
-            auto getGlobalOp = memref::GetGlobalOp::create(rewriter, loc, memRefType, argName);
-            auto toTensor = bufferization::ToTensorOp::create(
-              rewriter, loc, constRankedTensorType, getGlobalOp, rewriter.getUnitAttr(), rewriter.getUnitAttr());
-            mapSpatComputeToConst.insert({spatCompute.getOperation(), toTensor.getResult()});
-          }
-
-          replaceAndEraseDirectComputeLikeInput(
-            rewriter, spatCompute.getOperation(), BBArgIndex, mapSpatComputeToConst[spatCompute.getOperation()]);
-        }
-        else if (auto spatComputeBatch = llvm::dyn_cast<spatial::SpatComputeBatch>(constUsers)) {
-          auto inputIndex = getDirectComputeLikeInputIndex(spatComputeBatch, constUses.getOperandNumber());
-          if (!inputIndex)
-            return failure();
-          auto BBArgIndex = *inputIndex;
-          rewriter.setInsertionPoint(&spatComputeBatch.getBody().front().front());
-          if (!mapSpatComputeToConst.contains(spatComputeBatch.getOperation())) {
-            auto getGlobalOp = memref::GetGlobalOp::create(rewriter, loc, memRefType, argName);
-            auto toTensor = bufferization::ToTensorOp::create(
-              rewriter, loc, constRankedTensorType, getGlobalOp, rewriter.getUnitAttr(), rewriter.getUnitAttr());
-            mapSpatComputeToConst.insert({spatComputeBatch.getOperation(), toTensor.getResult()});
-          }
-
-          replaceAndEraseDirectComputeLikeInput(rewriter,
-                                                spatComputeBatch.getOperation(),
-                                                BBArgIndex,
-                                                mapSpatComputeToConst[spatComputeBatch.getOperation()]);
-        }
-        else {
-          {
-
-            if (auto spatCompute = constUses.getOwner()->getParentOfType<spatial::SpatCompute>()) {
-              rewriter.setInsertionPoint(&spatCompute.getBody().front().front());
-              if (!mapSpatComputeToConst.contains(spatCompute.getOperation())) {
-                auto getGlobalOp = memref::GetGlobalOp::create(rewriter, loc, memRefType, argName);
-                auto toTensor = bufferization::ToTensorOp::create(
-                  rewriter, loc, constRankedTensorType, getGlobalOp, rewriter.getUnitAttr(), rewriter.getUnitAttr());
-                mapSpatComputeToConst.insert({spatCompute.getOperation(), toTensor.getResult()});
-              }
-
-              rewriter.startOpModification(spatCompute.getOperation());
-              constUses.set(mapSpatComputeToConst[spatCompute.getOperation()]);
-              rewriter.finalizeOpModification(spatCompute.getOperation());
-            }
-            else if (auto spatComputeBatch = constUses.getOwner()->getParentOfType<spatial::SpatComputeBatch>()) {
-              rewriter.setInsertionPoint(&spatComputeBatch.getBody().front().front());
-              if (!mapSpatComputeToConst.contains(spatComputeBatch.getOperation())) {
-                auto getGlobalOp = memref::GetGlobalOp::create(rewriter, loc, memRefType, argName);
-                auto toTensor = bufferization::ToTensorOp::create(
-                  rewriter, loc, constRankedTensorType, getGlobalOp, rewriter.getUnitAttr(), rewriter.getUnitAttr());
-                mapSpatComputeToConst.insert({spatComputeBatch.getOperation(), toTensor.getResult()});
-              }
-
-              rewriter.startOpModification(spatComputeBatch.getOperation());
-              constUses.set(mapSpatComputeToConst[spatComputeBatch.getOperation()]);
-              rewriter.finalizeOpModification(spatComputeBatch.getOperation());
-            }
-          }
-        }
-      }
-    }
-    else if (constantOp.getType().isIntOrIndexOrFloat()) {
-      llvm::DenseMap<Operation*, Value> mapSpatComputeToConst;
-
-      for (auto& constUses : llvm::make_early_inc_range(constantOp->getUses())) {
-        auto constUsers = constUses.getOwner();
-
-        if (auto spatCompute = llvm::dyn_cast<spatial::SpatCompute>(constUsers)) {
-          auto inputIndex = getDirectComputeLikeInputIndex(spatCompute, constUses.getOperandNumber());
-          if (!inputIndex)
-            return failure();
-          auto BBArgIndex = *inputIndex;
-          rewriter.setInsertionPoint(&spatCompute.getBody().front().front());
-          auto newConst = rewriter.clone(*constantOp);
-
-          replaceAndEraseDirectComputeLikeInput(
-            rewriter, spatCompute.getOperation(), BBArgIndex, newConst->getResult(0));
-        }
-        else if (auto spatComputeBatch = llvm::dyn_cast<spatial::SpatComputeBatch>(constUsers)) {
-          auto inputIndex = getDirectComputeLikeInputIndex(spatComputeBatch, constUses.getOperandNumber());
-          if (!inputIndex)
-            return failure();
-          auto BBArgIndex = *inputIndex;
-          rewriter.setInsertionPoint(&spatComputeBatch.getBody().front().front());
-          auto newConst = rewriter.clone(*constantOp);
-
-          replaceAndEraseDirectComputeLikeInput(
-            rewriter, spatComputeBatch.getOperation(), BBArgIndex, newConst->getResult(0));
-        }
-        else if (auto parent = constUsers->getParentOfType<spatial::SpatCompute>()) {
-          if (!mapSpatComputeToConst.contains(parent)) {
-            rewriter.setInsertionPoint(&parent.getBody().front().front());
-            auto newConst = rewriter.clone(*constantOp);
-            mapSpatComputeToConst.insert({parent.getOperation(), newConst->getResult(0)});
-          }
-          constUses.set(mapSpatComputeToConst[parent.getOperation()]);
-        }
-        else {
-          auto batchParent = constUsers->getParentOfType<spatial::SpatComputeBatch>();
-          assert(batchParent && "Global Constant used direcly not within a compute");
-          if (!mapSpatComputeToConst.contains(batchParent.getOperation())) {
-            rewriter.setInsertionPoint(&batchParent.getBody().front().front());
-            auto newConst = rewriter.clone(*constantOp);
-            mapSpatComputeToConst.insert({batchParent.getOperation(), newConst->getResult(0)});
-          }
-          constUses.set(mapSpatComputeToConst[batchParent.getOperation()]);
-        }
-      }
-    }
-    if (constantOp->use_empty())
-      rewriter.eraseOp(constantOp);
-    return success();
-  }
-};
-
 // Materializes public function tensor inputs as globals so compute bodies can load them uniformly.
 struct FuncOpArgToGlobalMemoryPattern final : OpRewritePattern<mlir::func::FuncOp> {
   using OpRewritePattern::OpRewritePattern;
@@ -383,8 +221,7 @@ struct FuncOpArgToGlobalMemoryPattern final : OpRewritePattern<mlir::func::FuncO
 
 } // namespace
 void populateGlobalTensorMaterializationPatterns(RewritePatternSet& patterns) {
-  patterns.add<MoveExtractSliceIntoCompute, FuncOpArgToGlobalMemoryPattern, ArithConstToGlobalMemoryPattern>(
+  patterns.add<MoveExtractSliceIntoCompute, FuncOpArgToGlobalMemoryPattern>(patterns.getContext());
-    patterns.getContext());
 }
 
 } // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/Patterns/TensorPacking.cpp

@@ -1,4 +1,4 @@
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/TensorPackingPatterns.hpp"
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Patterns.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
 using namespace mlir;

src/PIM/Conversion/SpatialToPim/Patterns/Transpose.cpp

@@ -0,0 +1,38 @@
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+
+#include "src/Accelerators/PIM/Conversion/SpatialToPim/Patterns.hpp"
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace {
+
+struct LinalgTransposeToPim final : OpRewritePattern<linalg::TransposeOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(linalg::TransposeOp transposeOp, PatternRewriter& rewriter) const override {
+    SmallVector<Attribute> permutationAttrs;
+    permutationAttrs.reserve(transposeOp.getPermutation().size());
+    for (int64_t dim : transposeOp.getPermutation())
+      permutationAttrs.push_back(rewriter.getI64IntegerAttr(dim));
+
+    auto permutation = rewriter.getArrayAttr(permutationAttrs);
+    auto pimTranspose = pim::PimTransposeOp::create(rewriter,
+                                                    transposeOp.getLoc(),
+                                                    TypeRange {transposeOp->getResult(0).getType()},
+                                                    transposeOp.getInput(),
+                                                    permutation,
+                                                    transposeOp.getInit());
+    rewriter.replaceOp(transposeOp, pimTranspose.getOutput());
+    return success();
+  }
+};
+
+} // namespace
+
+void populateTransposeLoweringPatterns(RewritePatternSet& patterns) {
+  patterns.add<LinalgTransposeToPim>(patterns.getContext());
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/PhaseVerification.cpp

@@ -1,20 +0,0 @@
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/PhaseVerification.hpp"
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-
-using namespace mlir;
-
-namespace onnx_mlir {
-
-LogicalResult verifySpatialToPimBoundary(ModuleOp moduleOp) {
-  bool hasFailure = false;
-  moduleOp.walk([&](Operation* op) {
-    if (op->getDialect()->getNamespace() != "spat")
-      return;
-
-    op->emitError("illegal Spatial operation remains after Spatial-to-PIM lowering");
-    hasFailure = true;
-  });
-  return success(!hasFailure);
-}
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/PhaseVerification.hpp

@@ -1,9 +0,0 @@
-#pragma once
-
-#include "mlir/IR/BuiltinOps.h"
-
-namespace onnx_mlir {
-
-mlir::LogicalResult verifySpatialToPimBoundary(mlir::ModuleOp moduleOp);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/ReturnPathNormalization.cpp

@@ -1,15 +1,18 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Tosa/IR/TosaOps.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/IRMapping.h"
 #include "mlir/IR/SymbolTable.h"
+#include "mlir/Transforms/FoldUtils.h"
 
 #include "Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "Conversion/SpatialToPim/SpatialToPimPass.hpp"
+#include "src/Accelerators/PIM/Common/PimCommon.hpp"
 #include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/ReturnPathNormalization.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
@@ -38,15 +41,10 @@ static bool isReturnHelperChainOp(Operation* op) {
              tensor::ExpandShapeOp,
              tensor::CastOp,
              tosa::ReshapeOp,
-             ONNXTransposeOp,
+             linalg::TransposeOp,
              pim::PimTransposeOp>(op);
 }
 
-static void markOpToRemove(ReturnPathState& state, Operation* op) {
-  if (!llvm::is_contained(state.operationsToRemove, op))
-    state.operationsToRemove.push_back(op);
-}
-
 static std::string makeUniqueSymbolName(Operation* symbolTableOp, StringRef baseName) {
   std::string name = baseName.str();
   unsigned suffix = 0;
@@ -279,11 +277,10 @@ static LogicalResult mapIndicesThroughHelperChain(ArrayRef<int64_t> sourceIndice
       continue;
     }
 
-    if (auto transposeOp = dyn_cast<ONNXTransposeOp>(op)) {
+    if (auto transposeOp = dyn_cast<linalg::TransposeOp>(op)) {
       SmallVector<int64_t> nextIndices(currentIndices.size());
       SmallVector<int64_t> nextShape(currentShape.size());
-      for (auto [destIndex, attr] : llvm::enumerate(transposeOp.getPermAttr().getAsRange<IntegerAttr>())) {
+      for (auto [destIndex, sourceIndex] : llvm::enumerate(transposeOp.getPermutation())) {
-        int64_t sourceIndex = attr.getInt();
         nextIndices[destIndex] = currentIndices[sourceIndex];
         nextShape[destIndex] = currentShape[sourceIndex];
       }
@@ -318,7 +315,8 @@ static LogicalResult mapIndicesThroughHelperChain(ArrayRef<int64_t> sourceIndice
   return success();
 }
 
-static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewriter& rewriter) {
+static void
+cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewriter& rewriter, OperationFolder& constantFolder) {
   for (Value operand : op->getOperands()) {
     if (mapping.lookupOrNull(operand))
       continue;
@@ -327,7 +325,12 @@ static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewri
     if (!definingOp)
       continue;
 
-    if (!isa<tensor::EmptyOp, arith::ConstantOp>(definingOp))
+    if (auto constantOp = dyn_cast<arith::ConstantOp>(definingOp)) {
+      mapping.map(operand, getOrCreateHostConstantLike(constantOp, constantFolder));
+      continue;
+    }
+
+    if (!isa<tensor::EmptyOp>(definingOp))
       continue;
 
     Operation* clonedOp = rewriter.clone(*definingOp, mapping);
@@ -337,15 +340,18 @@ static void cloneMappedHelperOperands(Operation* op, IRMapping& mapping, IRRewri
   }
 }
 
-static void
+static void cloneHelperChain(Value sourceValue,
-cloneHelperChain(Value sourceValue, ArrayRef<Operation*> helperChain, IRRewriter& rewriter, Value& clonedValue) {
+                             ArrayRef<Operation*> helperChain,
+                             IRRewriter& rewriter,
+                             OperationFolder& constantFolder,
+                             Value& clonedValue) {
   IRMapping mapping;
   mapping.map(sourceValue, sourceValue);
   clonedValue = sourceValue;
 
   rewriter.setInsertionPointAfterValue(sourceValue);
   for (Operation* op : helperChain) {
-    cloneMappedHelperOperands(op, mapping, rewriter);
+    cloneMappedHelperOperands(op, mapping, rewriter, constantFolder);
     Operation* clonedOp = rewriter.clone(*op, mapping);
     for (auto [originalResult, newResult] : llvm::zip(op->getResults(), clonedOp->getResults()))
       mapping.map(originalResult, newResult);
@@ -360,23 +366,26 @@ static Value emitHostCopy(IRRewriter& rewriter,
                           Value sourceValue,
                           int32_t hostTargetOffset,
                           int32_t deviceSourceOffset,
-                          int32_t sizeInBytes) {
+                          int32_t sizeInBytes,
+                          OperationFolder& constantFolder) {
+  Operation* anchorOp = sourceValue.getDefiningOp() ? sourceValue.getDefiningOp() : outputTensor.getDefiningOp();
+  assert(anchorOp && "expected a concrete op anchor for return-path host copy constants");
+  Value hostTargetOffsetValue = getOrCreateHostIndexConstant(anchorOp, hostTargetOffset, constantFolder);
+  Value deviceSourceOffsetValue = getOrCreateHostIndexConstant(anchorOp, deviceSourceOffset, constantFolder);
   return PimMemCopyDevToHostOp::create(rewriter,
                                        loc,
                                        outputTensor.getType(),
+                                       hostTargetOffsetValue,
+                                       deviceSourceOffsetValue,
                                        outputTensor,
                                        sourceValue,
-                                       rewriter.getI32IntegerAttr(hostTargetOffset),
-                                       rewriter.getI32IntegerAttr(deviceSourceOffset),
                                        rewriter.getI32IntegerAttr(sizeInBytes))
     .getOutput();
 }
 
 } // namespace
 
-void addReturnOutputBuffers(func::ReturnOp returnOp,
+void raptor::SpatialToPimPass::addReturnOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewriter) {
-                            IRRewriter& rewriter,
-                            SmallVectorImpl<OutputTensorFactory>& outputTensors) {
   outputTensors.reserve(returnOp->getNumOperands());
   for (auto [index, returnValue] : llvm::enumerate(returnOp->getOperands())) {
     Value currentReturnValue = returnValue;
@@ -411,70 +420,85 @@ void addReturnOutputBuffers(func::ReturnOp returnOp,
   }
 }
 
-ReturnPathLoweringResult lowerComputeResultReturnPath(
+raptor::SpatialToPimPass::ReturnPathLoweringResult raptor::SpatialToPimPass::lowerProducedValueReturnPath(
-  spatial::SpatCompute computeOp, OpResult result, Value yieldValue, ReturnPathState& state, IRRewriter& rewriter) {
+  Operation* producerOp, Value producedValue, Value storedValue, IRRewriter& rewriter) {
-  Location loc = computeOp->getLoc();
+  Location loc = producerOp->getLoc();
-  auto yieldType = cast<TensorType>(yieldValue.getType());
+  OperationFolder constantFolder(producerOp->getContext());
+  auto storedTensorType = cast<TensorType>(storedValue.getType());
 
-  if (auto returnUse = analyzeReturnUse(result)) {
+  if (auto returnUse = analyzeReturnUse(producedValue)) {
-    Value storedValue = yieldValue;
+    Value currentStoredValue = storedValue;
-    cloneHelperChain(yieldValue, returnUse->helperChain, rewriter, storedValue);
+    cloneHelperChain(storedValue, returnUse->helperChain, rewriter, constantFolder, currentStoredValue);
     for (Operation* op : returnUse->helperChain)
-      markOpToRemove(state, op);
+      markOpToRemove(op);
 
-    auto storedType = cast<ShapedType>(storedValue.getType());
+    auto storedType = cast<ShapedType>(currentStoredValue.getType());
-    size_t elementSize = storedType.getElementTypeBitWidth() / 8;
+    size_t elementSize = getElementTypeSizeInBytes(storedType.getElementType());
-    if (auto storedOp = storedValue.getDefiningOp())
+    if (auto storedOp = currentStoredValue.getDefiningOp())
       rewriter.setInsertionPointAfter(storedOp);
-    Value outputTensor = state.outputTensors[returnUse->returnIndex](rewriter, loc);
+    Value outputTensor = outputTensors[returnUse->returnIndex](rewriter, loc);
-    emitHostCopy(
+    emitHostCopy(rewriter,
-      rewriter, loc, outputTensor, storedValue, 0, 0, static_cast<int32_t>(storedType.getNumElements() * elementSize));
+                 loc,
+                 outputTensor,
+                 currentStoredValue,
+                 0,
+                 0,
+                 static_cast<int32_t>(storedType.getNumElements() * elementSize),
+                 constantFolder);
     return ReturnPathLoweringResult::Handled;
   }
 
-  auto resultUses = result.getUses();
+  auto resultUses = producedValue.getUses();
   if (rangeLength(resultUses) == 1) {
     OpOperand& resultUse = *resultUses.begin();
     Operation* resultUser = resultUse.getOwner();
 
     if (isa<func::ReturnOp>(resultUser)) {
       size_t resultIndexInReturn = resultUse.getOperandNumber();
-      size_t elementSize = yieldType.getElementType().getIntOrFloatBitWidth() / 8;
+      size_t elementSize = getElementTypeSizeInBytes(storedTensorType.getElementType());
-      rewriter.setInsertionPointAfterValue(yieldValue);
+      rewriter.setInsertionPointAfterValue(storedValue);
-      Value outputTensor = state.outputTensors[resultIndexInReturn](rewriter, loc);
+      Value outputTensor = outputTensors[resultIndexInReturn](rewriter, loc);
-      emitHostCopy(
+      emitHostCopy(rewriter,
-        rewriter, loc, outputTensor, yieldValue, 0, 0, static_cast<int32_t>(yieldType.getNumElements() * elementSize));
+                   loc,
+                   outputTensor,
+                   storedValue,
+                   0,
+                   0,
+                   static_cast<int32_t>(storedTensorType.getNumElements() * elementSize),
+                   constantFolder);
       return ReturnPathLoweringResult::Handled;
     }
   }
 
-  if (auto concatReturnUse = analyzeConcatReturnUse(result)) {
+  if (auto concatReturnUse = analyzeConcatReturnUse(producedValue)) {
-    size_t elementSize = yieldType.getElementTypeBitWidth() / 8;
+    size_t elementSize = getElementTypeSizeInBytes(storedTensorType.getElementType());
     for (Operation* concatOp : concatReturnUse->concatChain)
-      markOpToRemove(state, concatOp);
+      markOpToRemove(concatOp);
 
     if (concatReturnUse->helperChain.empty()) {
-      rewriter.setInsertionPointAfterValue(yieldValue);
+      rewriter.setInsertionPointAfterValue(storedValue);
-      Value outputTensor = state.outputTensors[concatReturnUse->returnIndex](rewriter, loc);
+      Value outputTensor = outputTensors[concatReturnUse->returnIndex](rewriter, loc);
       auto outputType = cast<ShapedType>(outputTensor.getType());
       int64_t flatOffset = computeFlatElementIndex(concatReturnUse->sliceOffsets, outputType.getShape());
       emitHostCopy(rewriter,
                    loc,
                    outputTensor,
-                   yieldValue,
+                   storedValue,
                    static_cast<int32_t>(flatOffset * elementSize),
                    0,
-                   static_cast<int32_t>(yieldType.getNumElements() * elementSize));
+                   static_cast<int32_t>(storedTensorType.getNumElements() * elementSize),
+                   constantFolder);
       return ReturnPathLoweringResult::Handled;
     }
 
-    auto storedType = dyn_cast<RankedTensorType>(yieldValue.getType());
+    auto storedType = dyn_cast<RankedTensorType>(storedValue.getType());
     if (!storedType) {
-      computeOp.emitOpError("has an unsupported non-ranked concat-return helper yield during Spatial-to-PIM lowering");
+      producerOp->emitOpError(
+        "has an unsupported non-ranked concat-return helper yield during Spatial-to-PIM lowering");
       return ReturnPathLoweringResult::Failure;
     }
-    rewriter.setInsertionPointAfterValue(yieldValue);
+    rewriter.setInsertionPointAfterValue(storedValue);
-    Value outputTensor = state.outputTensors[concatReturnUse->returnIndex](rewriter, loc);
+    Value outputTensor = outputTensors[concatReturnUse->returnIndex](rewriter, loc);
     auto outputType = cast<ShapedType>(outputTensor.getType());
     for (int64_t linearIndex = 0; linearIndex < storedType.getNumElements(); ++linearIndex) {
       SmallVector<int64_t> sourceIndices = expandFlatElementIndex(linearIndex, storedType.getShape());
@@ -484,7 +508,7 @@ ReturnPathLoweringResult lowerComputeResultReturnPath(
       SmallVector<int64_t> destinationIndices;
       if (failed(mapIndicesThroughHelperChain(
             sourceIndices, concatReturnUse->concatShape, concatReturnUse->helperChain, destinationIndices))) {
-        computeOp.emitOpError("has an unsupported concat-return helper chain during Spatial-to-PIM lowering");
+        producerOp->emitOpError("has an unsupported concat-return helper chain during Spatial-to-PIM lowering");
         return ReturnPathLoweringResult::Failure;
       }
 
@@ -503,7 +527,7 @@ ReturnPathLoweringResult lowerComputeResultReturnPath(
       auto scalarTensorType =
         RankedTensorType::get(SmallVector<int64_t>(storedType.getRank(), 1), storedType.getElementType());
       auto elementSlice = tensor::ExtractSliceOp::create(
-        rewriter, loc, scalarTensorType, yieldValue, extractOffsets, extractSizes, extractStrides);
+        rewriter, loc, scalarTensorType, storedValue, extractOffsets, extractSizes, extractStrides);
       rewriter.setInsertionPointAfter(elementSlice);
 
       int64_t destinationFlatOffset = computeFlatElementIndex(destinationIndices, outputType.getShape());
@@ -513,7 +537,8 @@ ReturnPathLoweringResult lowerComputeResultReturnPath(
                                   elementSlice.getResult(),
                                   static_cast<int32_t>(destinationFlatOffset * elementSize),
                                   0,
-                                  static_cast<int32_t>(elementSize));
+                                  static_cast<int32_t>(elementSize),
+                                  constantFolder);
     }
     return ReturnPathLoweringResult::Handled;
   }
@@ -521,7 +546,12 @@ ReturnPathLoweringResult lowerComputeResultReturnPath(
   return ReturnPathLoweringResult::NotReturnPath;
 }
 
-void replaceReturnWithOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewriter, ReturnPathState& state) {
+raptor::SpatialToPimPass::ReturnPathLoweringResult raptor::SpatialToPimPass::lowerComputeResultReturnPath(
+  spatial::SpatCompute computeOp, OpResult result, Value yieldValue, IRRewriter& rewriter) {
+  return lowerProducedValueReturnPath(computeOp.getOperation(), result, yieldValue, rewriter);
+}
+
+void raptor::SpatialToPimPass::replaceReturnWithOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewriter) {
   auto markOwnedReturnChain = [&](Operation* op, auto&& markOwnedReturnChain) -> void {
     if (!op)
       return;
@@ -538,13 +568,13 @@ void replaceReturnWithOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewrite
 
     if (isReturnHelperChainOp(op)) {
       Value source = op->getOperand(0);
-      markOpToRemove(state, op);
+      markOpToRemove(op);
       markOwnedReturnChain(source.getDefiningOp(), markOwnedReturnChain);
       return;
     }
 
     if (auto computeOp = dyn_cast<spatial::SpatCompute>(op)) {
-      markOpToRemove(state, computeOp);
+      markOpToRemove(computeOp);
       if (!computeOp.getInputs().empty())
         for (Value input : computeOp.getInputs())
           markOwnedReturnChain(input.getDefiningOp(), markOwnedReturnChain);
@@ -552,23 +582,29 @@ void replaceReturnWithOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewrite
     }
 
     if (auto concatOp = dyn_cast<tensor::ConcatOp>(op)) {
-      markOpToRemove(state, concatOp);
+      markOpToRemove(concatOp);
       for (Value operand : concatOp.getOperands())
         markOwnedReturnChain(operand.getDefiningOp(), markOwnedReturnChain);
       return;
     }
 
     if (auto concatOp = dyn_cast<spatial::SpatConcatOp>(op)) {
-      markOpToRemove(state, concatOp);
+      markOpToRemove(concatOp);
       for (Value operand : concatOp.getInputs())
         markOwnedReturnChain(operand.getDefiningOp(), markOwnedReturnChain);
       return;
     }
 
     if (auto concatOp = dyn_cast<pim::PimConcatOp>(op)) {
-      markOpToRemove(state, concatOp);
+      markOpToRemove(concatOp);
       for (Value operand : concatOp.getInputs())
         markOwnedReturnChain(operand.getDefiningOp(), markOwnedReturnChain);
+      return;
+    }
+
+    if (auto receiveOp = dyn_cast<spatial::SpatChannelReceiveOp>(op)) {
+      markOpToRemove(receiveOp);
+      return;
     }
   };
 
@@ -578,7 +614,7 @@ void replaceReturnWithOutputBuffers(func::ReturnOp returnOp, IRRewriter& rewrite
     size_t orderWithinReturn = it.index();
     Operation* returnOperand = it.value().getDefiningOp();
     rewriter.setInsertionPoint(returnOp);
-    Value outputTensor = state.outputTensors[orderWithinReturn](rewriter, loc);
+    Value outputTensor = outputTensors[orderWithinReturn](rewriter, loc);
     rewriter.modifyOpInPlace(returnOp, [&] { returnOp.setOperand(orderWithinReturn, outputTensor); });
     markOwnedReturnChain(returnOperand, markOwnedReturnChain);
   }

src/PIM/Conversion/SpatialToPim/ReturnPathNormalization.hpp

@@ -1,37 +0,0 @@
-#pragma once
-
-#include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/IR/PatternMatch.h"
-
-#include <functional>
-
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-
-namespace onnx_mlir {
-
-using OutputTensorFactory = std::function<mlir::Value(mlir::IRRewriter& rewriter, mlir::Location loc)>;
-
-struct ReturnPathState {
-  llvm::SmallVectorImpl<OutputTensorFactory>& outputTensors;
-  llvm::SmallVectorImpl<mlir::Operation*>& operationsToRemove;
-};
-
-enum class ReturnPathLoweringResult {
-  Handled,
-  NotReturnPath,
-  Failure
-};
-
-void addReturnOutputBuffers(mlir::func::ReturnOp returnOp,
-                            mlir::IRRewriter& rewriter,
-                            llvm::SmallVectorImpl<OutputTensorFactory>& outputTensors);
-
-ReturnPathLoweringResult lowerComputeResultReturnPath(spatial::SpatCompute computeOp,
-                                                      mlir::OpResult result,
-                                                      mlir::Value yieldValue,
-                                                      ReturnPathState& state,
-                                                      mlir::IRRewriter& rewriter);
-
-void replaceReturnWithOutputBuffers(mlir::func::ReturnOp returnOp, mlir::IRRewriter& rewriter, ReturnPathState& state);
-
-} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/SpatialToPim.td

@@ -9,15 +9,15 @@ include "src/Accelerators/PIM/Dialect/Spatial/Spatial.td"
 include "src/Accelerators/PIM/Dialect/Pim/Pim.td"
 #endif // OP_BASE
 
-def onnxToPimTranspose : Pat<
+def spatToPimVMM : Pat<
-  (ONNXTransposeOp:$srcOpRes $data, $perms),
+  (SpatVMMOp:$srcOpRes $weight, $vector),
-  (PimTransposeOp $data, $perms,
+  (PimVMMOp $weight, $vector,
     (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
 >;
 
-def spatToPimVMM : Pat<
+def spatToPimVVDMul : Pat<
-  (SpatVMMOp:$srcOpRes $weightIndex, $vector),
+  (SpatVVDMulOp:$srcOpRes $a, $b),
-  (PimVMMOp $weightIndex, $vector,
+  (PimVVDMulOp $a, $b,
     (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
 >;
 

src/PIM/Conversion/SpatialToPim/SpatialToPimPass.cpp

@@ -1,8 +1,11 @@
+#include "mlir/Conversion/AffineToStandard/AffineToStandard.h"
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
+#include "mlir/Dialect/SCF/Utils/Utils.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/BuiltinDialect.h"
 #include "mlir/IR/BuiltinOps.h"
@@ -12,6 +15,7 @@
 #include "mlir/IR/SymbolTable.h"
 #include "mlir/IR/Value.h"
 #include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/FoldUtils.h"
 #include "mlir/Transforms/WalkPatternRewriteDriver.h"
 
 #include "llvm/ADT/StringRef.h"
@@ -21,20 +25,14 @@
 #include <cassert>
 #include <utility>
 
+#include "Common/PimCommon.hpp"
 #include "Conversion/ONNXToSpatial/Common/Common.hpp"
-#include "src/Accelerators/PIM/Common/PimCommon.hpp"
+#include "Conversion/SpatialToPim/Common.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/BatchCoreLoweringPatterns.hpp"
+#include "Conversion/SpatialToPim/Patterns.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/ChannelLoweringPatterns.hpp"
+#include "Dialect/Pim/PimOps.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/Cleanup.hpp"
+#include "Dialect/Spatial/SpatialOps.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/Common.hpp"
+#include "Pass/PIMPasses.h"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/CoreLoweringPatterns.hpp"
+#include "SpatialToPimPass.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/GlobalTensorMaterialization.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/PhaseVerification.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/ReturnPathNormalization.hpp"
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/TensorPackingPatterns.hpp"
-#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
-#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
-#include "src/Accelerators/PIM/Pass/PIMPasses.h"
 
 using namespace mlir;
 using namespace onnx_mlir;
@@ -42,34 +40,6 @@ using namespace pim;
 
 namespace onnx_mlir {
 
-namespace {
-
-#include "src/Accelerators/PIM/Conversion/SpatialToPim/SpatialToPim.hpp.inc"
-
-struct SpatialToPimPass : PassWrapper<SpatialToPimPass, OperationPass<ModuleOp>> {
-  MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(SpatialToPimPass)
-  StringRef getArgument() const override { return "convert-spatial-to-pim"; }
-  StringRef getDescription() const override { return "Lower Spatial ops to PIM-ready format"; }
-
-  SpatialToPimPass() = default;
-  SpatialToPimPass(const SpatialToPimPass& pass) {}
-
-  void runOnOperation() final;
-
-private:
-  SmallVector<OutputTensorFactory> outputTensors;
-  size_t coreId = 0;
-  SmallVector<Operation*> operationsToRemove;
-
-  LogicalResult allocateAndInitializeCoreLocalVariables(func::FuncOp funcOp, IRRewriter& rewriter);
-
-  void markOpToRemove(Operation* op);
-
-  void enlargeVMMOutTensorsToCrossbarSize(func::FuncOp funcOp, IRRewriter& rewriter);
-};
-
-} // namespace
-
 static memref::GlobalOp getOrCreateZeroGlobal(IRRewriter& rewriter, Location loc, RankedTensorType tensorType) {
   auto moduleOp = rewriter.getBlock()->getParentOp()->getParentOfType<ModuleOp>();
   auto memRefType = MemRefType::get(tensorType.getShape(), tensorType.getElementType());
@@ -104,23 +74,34 @@ static memref::GlobalOp getOrCreateZeroGlobal(IRRewriter& rewriter, Location loc
                                   IntegerAttr {});
 }
 
-static Value createZeroedDeviceHVector(IRRewriter& rewriter, Location loc, RankedTensorType tensorType) {
+static Value createZeroedDeviceHVector(IRRewriter& rewriter,
+                                       Location loc,
+                                       RankedTensorType tensorType,
+                                       OperationFolder& constantFolder) {
   auto outputBuffer = createEmptyTensorFromShaped(rewriter, loc, tensorType);
   auto zeroGlobal = getOrCreateZeroGlobal(rewriter, loc, tensorType);
   auto zeroValue = memref::GetGlobalOp::create(rewriter, loc, zeroGlobal.getType(), zeroGlobal.getName());
-  auto zeroAttr = rewriter.getI32IntegerAttr(0);
+  auto zeroIndex = getOrCreateHostIndexConstant(outputBuffer.getOperation(), 0, constantFolder);
   auto sizeAttr = rewriter.getI32IntegerAttr(static_cast<int32_t>(getShapedTypeSizeInBytes(tensorType)));
 
   if (outputBuffer->getParentOfType<PimCoreBatchOp>())
-    return PimMemCopyHostToDevBatchOp::create(
+    return PimMemCopyHostToDevBatchOp::create(rewriter,
-             rewriter, loc, tensorType, outputBuffer, zeroValue, zeroAttr, zeroAttr, sizeAttr)
+                                              loc,
+                                              tensorType,
+                                              outputBuffer,
+                                              zeroValue,
+                                              rewriter.getI32IntegerAttr(0),
+                                              rewriter.getI32IntegerAttr(0),
+                                              sizeAttr)
       .getOutput();
 
-  return PimMemCopyHostToDevOp::create(rewriter, loc, tensorType, outputBuffer, zeroValue, zeroAttr, zeroAttr, sizeAttr)
+  return PimMemCopyHostToDevOp::create(
+           rewriter, loc, tensorType, zeroIndex, zeroIndex, outputBuffer, zeroValue, sizeAttr)
     .getOutput();
 }
 
-static Value padHVectorInputToCrossbarSize(IRRewriter& rewriter, Location loc, Value vector) {
+static Value
+padHVectorInputToCrossbarSize(IRRewriter& rewriter, Location loc, Value vector, OperationFolder& constantFolder) {
   auto vectorType = cast<RankedTensorType>(vector.getType());
   ArrayRef<int64_t> shape = vectorType.getShape();
   assert(isHVectorShape(shape) && "expected a horizontal vector");
@@ -131,14 +112,16 @@ static Value padHVectorInputToCrossbarSize(IRRewriter& rewriter, Location loc, V
 
   auto paddedType = RankedTensorType::get(
     {shape[0], static_cast<int64_t>(crossbarSize)}, vectorType.getElementType(), vectorType.getEncoding());
-  Value zeroed = createZeroedDeviceHVector(rewriter, loc, paddedType);
+  Value zeroed = createZeroedDeviceHVector(rewriter, loc, paddedType, constantFolder);
   auto zeroAttr = rewriter.getI32IntegerAttr(0);
   auto sizeAttr = rewriter.getI32IntegerAttr(static_cast<int32_t>(getShapedTypeSizeInBytes(vectorType)));
   return PimMemCopyOp::create(rewriter, loc, paddedType, zeroed, vector, zeroAttr, zeroAttr, sizeAttr).getOutput();
 }
 
-void SpatialToPimPass::runOnOperation() {
+void onnx_mlir::raptor::SpatialToPimPass::runOnOperation() {
   coreId = 0;
+  outputTensors.clear();
+  operationsToRemove.clear();
   ModuleOp moduleOp = getOperation();
   MLIRContext* ctx = moduleOp.getContext();
 
@@ -151,9 +134,11 @@ void SpatialToPimPass::runOnOperation() {
   func::FuncOp funcOp = *entryFunc;
 
   IRRewriter rewriter(&getContext());
+  OperationFolder constantFolder(&getContext());
 
   ConversionTarget target(*ctx);
-  target.addLegalDialect<PimDialect,
+  target.addLegalDialect<affine::AffineDialect,
+                         PimDialect,
                          tensor::TensorDialect,
                          arith::ArithDialect,
                          bufferization::BufferizationDialect,
@@ -163,15 +148,11 @@ void SpatialToPimPass::runOnOperation() {
                          BuiltinDialect>();
   target.addLegalOp<spatial::SpatConcatOp,
                     spatial::SpatChannelReceiveOp,
-                    spatial::SpatChannelReceiveTensorOp,
-                    spatial::SpatChannelReceiveTensorBatchOp,
                     spatial::SpatChannelSendOp,
-                    spatial::SpatChannelSendTensorOp,
-                    spatial::SpatChannelSendTensorBatchOp,
                     spatial::SpatExtractRowsOp>();
 
   RewritePatternSet initialPatterns(ctx);
-  populateWithGenerated(initialPatterns);
+  populateInitialPatterns(initialPatterns);
   if (failed(applyPartialConversion(moduleOp, target, std::move(initialPatterns)))) {
     moduleOp.emitError("failed to lower required Spatial ops to the initial PIM form");
     signalPassFailure();
@@ -179,21 +160,20 @@ void SpatialToPimPass::runOnOperation() {
   }
 
   RewritePatternSet globalTensorPatterns(ctx);
-  populateGlobalTensorMaterializationPatterns(globalTensorPatterns);
+  populateGlobalTensorMaterializationPatternPhase(globalTensorPatterns);
   walkAndApplyPatterns(moduleOp, std::move(globalTensorPatterns));
-  auto returnOp = cast<func::ReturnOp>(funcOp.front().getTerminator());
 
-  addReturnOutputBuffers(returnOp, rewriter, outputTensors);
+  auto returnOp = cast<func::ReturnOp>(funcOp.front().getTerminator());
+  addReturnOutputBuffers(returnOp, rewriter);
   if (failed(allocateAndInitializeCoreLocalVariables(funcOp, rewriter))) {
     funcOp.emitOpError("failed to allocate or initialize core-local tensors during Spatial-to-PIM lowering");
     signalPassFailure();
     return;
   }
 
-  CoreLoweringState coreLoweringState {coreId, outputTensors, operationsToRemove};
   for (auto computeOp : funcOp.getOps<spatial::SpatCompute>()) {
     markOpToRemove(computeOp);
-    if (failed(lowerComputeOp(computeOp, coreLoweringState, rewriter))) {
+    if (failed(lowerComputeOp(computeOp, rewriter, constantFolder))) {
       computeOp.emitOpError("failed to lower spat.compute to pim.core");
       signalPassFailure();
       return;
@@ -202,7 +182,7 @@ void SpatialToPimPass::runOnOperation() {
 
   for (auto computeBatchOp : funcOp.getOps<spatial::SpatComputeBatch>()) {
     markOpToRemove(computeBatchOp);
-    if (failed(lowerComputeBatchOp(computeBatchOp, coreLoweringState, rewriter))) {
+    if (failed(lowerComputeBatchOp(computeBatchOp, rewriter))) {
       computeBatchOp.emitOpError("failed to lower spat.compute_batch to pim.core_batch");
       signalPassFailure();
       return;
@@ -210,7 +190,7 @@ void SpatialToPimPass::runOnOperation() {
   }
 
   RewritePatternSet initialTensorPackingPatterns(ctx);
-  populateTensorPackingPatterns(initialTensorPackingPatterns);
+  populateInitialTensorPackingPatterns(initialTensorPackingPatterns);
   walkAndApplyPatterns(funcOp, std::move(initialTensorPackingPatterns));
   eraseUnusedTensorPackingOps(funcOp, rewriter);
 
@@ -227,13 +207,28 @@ void SpatialToPimPass::runOnOperation() {
   }
 
   RewritePatternSet coreBodyPatterns(ctx);
-  populateWithGenerated(coreBodyPatterns);
+  populateCoreBodyPatterns(coreBodyPatterns);
+  populateAffineToStdConversionPatterns(coreBodyPatterns);
   FrozenRewritePatternSet frozenCoreBodyPatterns(std::move(coreBodyPatterns));
 
+  ConversionTarget coreBodyTarget(*ctx);
+  coreBodyTarget.addLegalDialect<PimDialect,
+                                 tensor::TensorDialect,
+                                 arith::ArithDialect,
+                                 bufferization::BufferizationDialect,
+                                 func::FuncDialect,
+                                 memref::MemRefDialect,
+                                 scf::SCFDialect,
+                                 BuiltinDialect>();
+  coreBodyTarget.addLegalOp<spatial::SpatConcatOp,
+                            spatial::SpatChannelReceiveOp,
+                            spatial::SpatChannelSendOp,
+                            spatial::SpatExtractRowsOp>();
+
   SmallVector<pim::PimCoreOp> coreOps;
   funcOp.walk([&](pim::PimCoreOp coreOp) { coreOps.push_back(coreOp); });
   for (auto coreOp : coreOps) {
-    if (failed(applyFullConversion(coreOp.getOperation(), target, frozenCoreBodyPatterns))) {
+    if (failed(applyFullConversion(coreOp.getOperation(), coreBodyTarget, frozenCoreBodyPatterns))) {
       coreOp.emitOpError("failed to convert nested Spatial ops inside pim.core");
       signalPassFailure();
       return;
@@ -243,7 +238,7 @@ void SpatialToPimPass::runOnOperation() {
   SmallVector<pim::PimCoreBatchOp> coreBatchOps;
   funcOp.walk([&](pim::PimCoreBatchOp coreBatchOp) { coreBatchOps.push_back(coreBatchOp); });
   for (auto coreBatchOp : coreBatchOps) {
-    if (failed(applyFullConversion(coreBatchOp.getOperation(), target, frozenCoreBodyPatterns))) {
+    if (failed(applyFullConversion(coreBatchOp.getOperation(), coreBodyTarget, frozenCoreBodyPatterns))) {
       coreBatchOp.emitOpError("failed to convert nested Spatial ops inside pim.core_batch");
       signalPassFailure();
       return;
@@ -251,18 +246,11 @@ void SpatialToPimPass::runOnOperation() {
   }
 
   enlargeVMMOutTensorsToCrossbarSize(funcOp, rewriter);
-  ReturnPathState returnPathState {outputTensors, operationsToRemove};
+  replaceReturnWithOutputBuffers(returnOp, rewriter);
-  replaceReturnWithOutputBuffers(returnOp, rewriter, returnPathState);
+  eraseOpsToRemove();
-
-  SmallVector<Operation*> pendingRemovals(operationsToRemove.begin(), operationsToRemove.end());
-  if (failed(erasePendingOps(pendingRemovals, rewriter))) {
-    funcOp.emitOpError("failed to erase obsolete Spatial ops after lowering to PIM");
-    signalPassFailure();
-    return;
-  }
 
   RewritePatternSet finalTensorPackingPatterns(ctx);
-  populateTensorPackingPatterns(finalTensorPackingPatterns);
+  populateFinalTensorPackingPatterns(finalTensorPackingPatterns);
   walkAndApplyPatterns(funcOp, std::move(finalTensorPackingPatterns));
   eraseUnusedTensorPackingOps(funcOp, rewriter);
 
@@ -278,30 +266,22 @@ void SpatialToPimPass::runOnOperation() {
   communicationTarget.addLegalOp<ModuleOp>();
   communicationTarget.addIllegalOp<spatial::SpatConcatOp,
                                    spatial::SpatChannelReceiveOp,
-                                   spatial::SpatChannelReceiveTensorOp,
                                    spatial::SpatChannelSendOp,
-                                   spatial::SpatChannelSendTensorOp,
                                    spatial::SpatExtractRowsOp>();
 
   RewritePatternSet communicationPatterns(ctx);
-  populateChannelLoweringPatterns(communicationPatterns);
+  populateCommunicationPatterns(communicationPatterns);
   if (failed(applyFullConversion(funcOp, communicationTarget, std::move(communicationPatterns)))) {
     funcOp.emitOpError("failed to lower Spatial communication ops to PIM communication ops");
     signalPassFailure();
     return;
   }
-
-  if (failed(verifySpatialToPimBoundary(moduleOp))) {
-    moduleOp.emitError("Spatial-to-PIM boundary verification failed");
-    signalPassFailure();
-    return;
-  }
-
   // Dump to file for debug
   dumpModule(moduleOp, "pim0");
 }
 
-void SpatialToPimPass::enlargeVMMOutTensorsToCrossbarSize(func::FuncOp funcOp, IRRewriter& rewriter) {
+void raptor::SpatialToPimPass::enlargeVMMOutTensorsToCrossbarSize(func::FuncOp funcOp, IRRewriter& rewriter) {
+  OperationFolder constantFolder(funcOp.getContext());
   funcOp.walk([&](PimVMMOp vmmOp) {
     auto outputType = cast<RankedTensorType>(vmmOp.getOutput().getType());
     ArrayRef<int64_t> outputShape = outputType.getShape();
@@ -309,7 +289,7 @@ void SpatialToPimPass::enlargeVMMOutTensorsToCrossbarSize(func::FuncOp funcOp, I
     assert(outputShape[1] <= static_cast<int64_t>(crossbarSize) && "output width must fit in one crossbar");
 
     rewriter.setInsertionPoint(vmmOp);
-    Value paddedInput = padHVectorInputToCrossbarSize(rewriter, vmmOp.getLoc(), vmmOp.getInput());
+    Value paddedInput = padHVectorInputToCrossbarSize(rewriter, vmmOp.getLoc(), vmmOp.getInput(), constantFolder);
     auto paddedOutputType = RankedTensorType::get(
       {outputShape[0], static_cast<int64_t>(crossbarSize)}, outputType.getElementType(), outputType.getEncoding());
     Value paddedOutputBuffer = outputShape[1] == static_cast<int64_t>(crossbarSize)
@@ -334,13 +314,17 @@ void SpatialToPimPass::enlargeVMMOutTensorsToCrossbarSize(func::FuncOp funcOp, I
   });
 }
 
-LogicalResult SpatialToPimPass::allocateAndInitializeCoreLocalVariables(func::FuncOp funcOp, IRRewriter& rewriter) {
+LogicalResult raptor::SpatialToPimPass::allocateAndInitializeCoreLocalVariables(func::FuncOp funcOp,
+                                                                                IRRewriter& rewriter) {
   Location loc = funcOp.getLoc();
+  OperationFolder constantFolder(funcOp.getContext());
 
   auto insertMemCopyHostToDev = [&](Value inputTensor, int64_t elementsOffset) {
     auto tensorType = cast<ShapedType>(inputTensor.getType());
     Type elementType = tensorType.getElementType();
-    size_t elementByteSize = elementType.getIntOrFloatBitWidth() / 8;
+    if (!hasByteSizedElementType(elementType))
+      return;
+    size_t elementByteSize = getElementTypeSizeInBytes(elementType);
     rewriter.setInsertionPointAfter(inputTensor.getDefiningOp());
 
     auto deviceTensor = tensor::EmptyOp::create(rewriter, loc, tensorType.getShape(), elementType);
@@ -349,10 +333,11 @@ LogicalResult SpatialToPimPass::allocateAndInitializeCoreLocalVariables(func::Fu
       rewriter,
       loc,
       tensorType,
+      getOrCreateHostIndexConstant(deviceTensor.getOperation(), 0, constantFolder),
+      getOrCreateHostIndexConstant(
+        deviceTensor.getOperation(), static_cast<int64_t>(elementsOffset * elementByteSize), constantFolder),
       deviceTensor,
       inputTensor,
-      rewriter.getI32IntegerAttr(0),
-      rewriter.getI32IntegerAttr(static_cast<int32_t>(elementsOffset * elementByteSize)),
       rewriter.getI32IntegerAttr(static_cast<int32_t>(tensorType.getNumElements() * elementByteSize)));
 
     rewriter.replaceAllUsesExcept(inputTensor, memCopyHostToDevOp.getResult(), {memCopyHostToDevOp});
@@ -374,11 +359,18 @@ LogicalResult SpatialToPimPass::allocateAndInitializeCoreLocalVariables(func::Fu
   return success();
 }
 
-void SpatialToPimPass::markOpToRemove(Operation* op) {
+void raptor::SpatialToPimPass::markOpToRemove(Operation* op) {
   if (!llvm::is_contained(operationsToRemove, op))
     operationsToRemove.push_back(op);
 }
 
-std::unique_ptr<Pass> createSpatialToPimPass() { return std::make_unique<SpatialToPimPass>(); }
+void raptor::SpatialToPimPass::eraseOpsToRemove() {
+  for (Operation* op : operationsToRemove) {
+    op->dropAllUses();
+    op->erase();
+  }
+}
+
+std::unique_ptr<Pass> createSpatialToPimPass() { return std::make_unique<raptor::SpatialToPimPass>(); }
 
 } // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/SpatialToPimPass.hpp

@@ -0,0 +1,72 @@
+#pragma once
+
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Bufferization/IR/Bufferization.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/SCF/Utils/Utils.h"
+#include "mlir/IR/BuiltinOps.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/Value.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/FoldUtils.h"
+
+#include "llvm/ADT/StringRef.h"
+
+#include <functional>
+
+#include "Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "Conversion/SpatialToPim/Common.hpp"
+#include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+
+namespace onnx_mlir {
+namespace raptor {
+
+struct SpatialToPimPass : mlir::PassWrapper<SpatialToPimPass, mlir::OperationPass<mlir::ModuleOp>> {
+  MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(SpatialToPimPass)
+  llvm::StringRef getArgument() const override { return "convert-spatial-to-pim"; }
+  llvm::StringRef getDescription() const override { return "Lower Spatial ops to PIM-ready format"; }
+
+  SpatialToPimPass() = default;
+  SpatialToPimPass(const SpatialToPimPass& pass) {}
+
+  void runOnOperation() final;
+
+private:
+  using OutputTensorFactory = std::function<mlir::Value(mlir::IRRewriter& rewriter, mlir::Location loc)>;
+
+  llvm::SmallVector<OutputTensorFactory> outputTensors;
+  size_t coreId = 0;
+  llvm::SmallVector<mlir::Operation*> operationsToRemove;
+
+  mlir::LogicalResult allocateAndInitializeCoreLocalVariables(mlir::func::FuncOp funcOp, mlir::IRRewriter& rewriter);
+  mlir::LogicalResult
+  lowerComputeOp(spatial::SpatCompute computeOp, mlir::IRRewriter& rewriter, mlir::OperationFolder& constantFolder);
+  mlir::LogicalResult lowerComputeBatchOp(spatial::SpatComputeBatch computeBatchOp, mlir::IRRewriter& rewriter);
+
+  enum class ReturnPathLoweringResult {
+    Handled,
+    NotReturnPath,
+    Failure
+  };
+
+  void addReturnOutputBuffers(mlir::func::ReturnOp returnOp, mlir::IRRewriter& rewriter);
+  ReturnPathLoweringResult lowerComputeResultReturnPath(spatial::SpatCompute computeOp,
+                                                        mlir::OpResult result,
+                                                        mlir::Value yieldValue,
+                                                        mlir::IRRewriter& rewriter);
+  ReturnPathLoweringResult lowerProducedValueReturnPath(mlir::Operation* producerOp,
+                                                        mlir::Value producedValue,
+                                                        mlir::Value storedValue,
+                                                        mlir::IRRewriter& rewriter);
+  void replaceReturnWithOutputBuffers(mlir::func::ReturnOp returnOp, mlir::IRRewriter& rewriter);
+
+  void markOpToRemove(mlir::Operation* op);
+  void eraseOpsToRemove();
+
+  void enlargeVMMOutTensorsToCrossbarSize(mlir::func::FuncOp funcOp, mlir::IRRewriter& rewriter);
+};
+
+} // namespace raptor
+
+} // namespace onnx_mlir

src/PIM/Dialect/Pim/Pim.td

@@ -2,6 +2,7 @@
 #define PIM_DIALECT_H
 
 include "mlir/IR/OpBase.td"
+include "mlir/IR/OpAsmInterface.td"
 include "mlir/IR/AttrTypeBase.td"
 include "mlir/Dialect/MemRef/IR/MemRefBase.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
@@ -24,7 +25,8 @@ def PimTensor :
 // Execution
 //===----------------------------------------------------------------------===//
 
-def PimCoreOp : PimOp<"core", [SingleBlock, IsolatedFromAbove]> {
+def PimCoreOp : PimOp<"core", [SingleBlock,
+    DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmBlockArgumentNames"]>]> {
   let summary = "Execute a block on a PIM core";
 
   let regions = (region SizedRegion<1>:$body);
@@ -34,12 +36,16 @@ def PimCoreOp : PimOp<"core", [SingleBlock, IsolatedFromAbove]> {
     I32Attr:$coreId
   );
 
-  let assemblyFormat = [{
+  let extraClassDeclaration = [{
-    `(` $weights `)` attr-dict regions `:` type($weights) `->` `(` `)`
+    ::mlir::BlockArgument getWeightArgument(unsigned idx);
   }];
+
+  let hasVerifier = 1;
+  let hasCustomAssemblyFormat = 1;
 }
 
-def PimCoreBatchOp : PimOp<"core_batch", [SingleBlock, IsolatedFromAbove, AttrSizedOperandSegments]> {
+def PimCoreBatchOp : PimOp<"core_batch", [SingleBlock, AttrSizedOperandSegments,
+    DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmBlockArgumentNames"]>]> {
   let summary = "Execute equivalent batched core bodies";
 
   let regions = (region SizedRegion<1>:$body);
@@ -50,6 +56,13 @@ def PimCoreBatchOp : PimOp<"core_batch", [SingleBlock, IsolatedFromAbove, AttrSi
     Variadic<PimTensor>:$inputs
   );
 
+  let extraClassDeclaration = [{
+    ::mlir::BlockArgument getLaneArgument();
+    ::mlir::BlockArgument getWeightArgument(unsigned idx);
+    ::mlir::BlockArgument getInputArgument(unsigned idx);
+  }];
+
+  let hasVerifier = 1;
   let hasCustomAssemblyFormat = 1;
 }
 
@@ -61,16 +74,6 @@ def PimHaltOp : PimOp<"halt", [Terminator]> {
   }];
 }
 
-def PimYieldOp : PimOp<"yield", [Terminator]> {
-  let summary = "Yield results from a Pim region";
-
-  let arguments = (ins
-    Variadic<PimTensor>:$outputs
-  );
-
-  let hasCustomAssemblyFormat = 1;
-}
-
 //===----------------------------------------------------------------------===//
 // Communication
 //===----------------------------------------------------------------------===//
@@ -81,57 +84,21 @@ def PimSendOp : PimOp<"send", []> {
   let arguments = (ins
     PimTensor:$input,
     I32Attr:$size,
-    I32Attr:$targetCoreId
+    Index:$targetCoreId
   );
 
   let assemblyFormat = [{
-    `(` $input `)` attr-dict `:` type($input) `->` `(` `)`
+    `(` $input `,` $targetCoreId `)` attr-dict `:` type($input) `->` `(` `)`
   }];
 }
 
-def PimSendTensorOp : PimOp<"send_tensor", []> {
-  let summary = "Send equal contiguous chunks of one tensor to target cores";
-
-  let arguments = (ins
-    PimTensor:$input,
-    DenseI32ArrayAttr:$targetCoreIds
-  );
-
-  let hasVerifier = 1;
-  let hasCustomAssemblyFormat = 1;
-}
-
-def PimSendBatchOp : PimOp<"send_batch", []> {
-  let summary = "Send a per-lane tensor to target cores from a batched core";
-
-  let arguments = (ins
-    PimTensor:$input,
-    I32Attr:$size,
-    DenseI32ArrayAttr:$targetCoreIds
-  );
-
-  let hasCustomAssemblyFormat = 1;
-}
-
-def PimSendTensorBatchOp : PimOp<"send_tensor_batch", []> {
-  let summary = "Send equal contiguous chunks of one per-lane tensor from a batched core";
-
-  let arguments = (ins
-    PimTensor:$input,
-    DenseI32ArrayAttr:$targetCoreIds
-  );
-
-  let hasVerifier = 1;
-  let hasCustomAssemblyFormat = 1;
-}
-
 def PimReceiveOp : PimOp<"receive", [DestinationStyleOpInterface]> {
   let summary = "Receive a tensor from another core";
 
   let arguments = (ins
     PimTensor:$outputBuffer,
     I32Attr:$size,
-    I32Attr:$sourceCoreId
+    Index:$sourceCoreId
   );
 
   let results = (outs
@@ -145,84 +112,18 @@ def PimReceiveOp : PimOp<"receive", [DestinationStyleOpInterface]> {
   }];
 
   let assemblyFormat = [{
-    `(` $outputBuffer `)` attr-dict `:` type($outputBuffer) `->` type($output)
+    `(` $outputBuffer `,` $sourceCoreId `)` attr-dict `:` type($outputBuffer) `->` type($output)
   }];
 }
 
-def PimReceiveTensorOp : PimOp<"receive_tensor", [DestinationStyleOpInterface]> {
-  let summary = "Receive equal contiguous chunks from source cores into one tensor";
-
-  let arguments = (ins
-    PimTensor:$outputBuffer,
-    DenseI32ArrayAttr:$sourceCoreIds
-  );
-
-  let results = (outs
-    PimTensor:$output
-  );
-
-  let extraClassDeclaration = [{
-    mlir::MutableOperandRange getDpsInitsMutable() {
-      return getOutputBufferMutable();
-    }
-  }];
-
-  let hasVerifier = 1;
-  let hasCustomAssemblyFormat = 1;
-}
-
-def PimReceiveBatchOp : PimOp<"receive_batch", [DestinationStyleOpInterface]> {
-  let summary = "Receive per-lane tensors from source cores into a batched core";
-
-  let arguments = (ins
-    PimTensor:$outputBuffer,
-    I32Attr:$size,
-    DenseI32ArrayAttr:$sourceCoreIds
-  );
-
-  let results = (outs
-    PimTensor:$output
-  );
-
-  let extraClassDeclaration = [{
-    mlir::MutableOperandRange getDpsInitsMutable() {
-      return getOutputBufferMutable();
-    }
-  }];
-
-  let hasCustomAssemblyFormat = 1;
-}
-
-def PimReceiveTensorBatchOp : PimOp<"receive_tensor_batch", [DestinationStyleOpInterface]> {
-  let summary = "Receive equal contiguous chunks into one per-lane tensor inside a batched core";
-
-  let arguments = (ins
-    PimTensor:$outputBuffer,
-    DenseI32ArrayAttr:$sourceCoreIds
-  );
-
-  let results = (outs
-    PimTensor:$output
-  );
-
-  let extraClassDeclaration = [{
-    mlir::MutableOperandRange getDpsInitsMutable() {
-      return getOutputBufferMutable();
-    }
-  }];
-
-  let hasVerifier = 1;
-  let hasCustomAssemblyFormat = 1;
-}
-
 def PimMemCopyHostToDevOp : PimOp<"memcp_hd", [DestinationStyleOpInterface]> {
   let summary = "Copy a memory region from host memory into device memory";
 
   let arguments = (ins
+    Index:$deviceTargetOffset,
+    Index:$hostSourceOffset,
     PimTensor:$deviceTarget,
     PimTensor:$hostSource,
-    I32Attr:$deviceTargetOffset,
-    I32Attr:$hostSourceOffset,
     I32Attr:$size
   );
 
@@ -237,7 +138,9 @@ def PimMemCopyHostToDevOp : PimOp<"memcp_hd", [DestinationStyleOpInterface]> {
   }];
 
   let assemblyFormat = [{
-    `(` $deviceTarget `,` $hostSource `)` attr-dict `:` `(` type($deviceTarget) `,` type($hostSource) `)` `->` type($output)
+    `[` $deviceTargetOffset `,` $hostSourceOffset `]`
+    `(` $deviceTarget `,` $hostSource `)` attr-dict
+    `:` type($deviceTarget) `,` type($hostSource) `->` type($output)
   }];
 }
 
@@ -271,10 +174,10 @@ def PimMemCopyDevToHostOp : PimOp<"memcp_dh", [DestinationStyleOpInterface]> {
   let summary = "Copy a memory region from device memory into host memory";
 
   let arguments = (ins
+    Index:$hostTargetOffset,
+    Index:$deviceSourceOffset,
     PimTensor:$hostTarget,
     PimTensor:$deviceSource,
-    I32Attr:$hostTargetOffset,
-    I32Attr:$deviceSourceOffset,
     I32Attr:$size
   );
 
@@ -289,7 +192,9 @@ def PimMemCopyDevToHostOp : PimOp<"memcp_dh", [DestinationStyleOpInterface]> {
   }];
 
   let assemblyFormat = [{
-    `(` $hostTarget `,` $deviceSource `)` attr-dict `:` `(` type($hostTarget) `,` type($deviceSource) `)` `->` type($output)
+    `[` $hostTargetOffset `,` $deviceSourceOffset `]`
+    `(` $hostTarget `,` $deviceSource `)` attr-dict
+    `:` type($hostTarget) `,` type($deviceSource) `->` type($output)
   }];
 }
 
@@ -374,7 +279,7 @@ def PimVMMOp : PimOp<"vmm", [DestinationStyleOpInterface]> {
   let summary = "Vector-matrix multiplication: c = a * b";
 
   let arguments = (ins
-    I32Attr:$weightIndex,
+    PimTensor:$weight,
     PimTensor:$input,
     PimTensor:$outputBuffer
   );
@@ -391,7 +296,8 @@ def PimVMMOp : PimOp<"vmm", [DestinationStyleOpInterface]> {
 
   let hasVerifier = 1;
   let assemblyFormat = [{
-    `(` $input `,` $outputBuffer `)` attr-dict `:` `(` type($input) `,` type($outputBuffer) `)` `->` type($output)
+    `[` $weight `]` `(` $input `,` $outputBuffer `)` attr-dict `:` `(` type($weight) `,` type($input) `,`
+      type($outputBuffer) `)` `->` type($output)
   }];
 }
 

src/PIM/Dialect/Pim/PimOps.cpp

@@ -1,8 +1,41 @@
+#include <string>
+
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
+using namespace mlir;
+
 namespace onnx_mlir {
 namespace pim {
 
+BlockArgument PimCoreOp::getWeightArgument(unsigned idx) { return getBody().front().getArgument(idx); }
+
+void PimCoreOp::getAsmBlockArgumentNames(Region& region, OpAsmSetValueNameFn setNameFn) {
+  if (region.empty())
+    return;
+
+  for (unsigned index = 0; index < getWeights().size(); ++index)
+    setNameFn(getWeightArgument(index), ("w" + std::to_string(index)).c_str());
+}
+
+BlockArgument PimCoreBatchOp::getLaneArgument() { return getBody().front().getArgument(0); }
+
+BlockArgument PimCoreBatchOp::getWeightArgument(unsigned idx) { return getBody().front().getArgument(1 + idx); }
+
+BlockArgument PimCoreBatchOp::getInputArgument(unsigned idx) {
+  return getBody().front().getArgument(1 + getWeights().size() + idx);
+}
+
+void PimCoreBatchOp::getAsmBlockArgumentNames(Region& region, OpAsmSetValueNameFn setNameFn) {
+  if (region.empty())
+    return;
+
+  setNameFn(getLaneArgument(), "lane");
+  for (unsigned index = 0; index < getWeights().size(); ++index)
+    setNameFn(getWeightArgument(index), ("w" + std::to_string(index)).c_str());
+  for (unsigned index = 0; index < getInputs().size(); ++index)
+    setNameFn(getInputArgument(index), ("in" + std::to_string(index)).c_str());
+}
+
 void PimDialect::initialize() {
   addOperations<
 #define GET_OP_LIST

src/PIM/Dialect/Pim/PimOpsAsm.cpp

@@ -20,28 +20,128 @@ static DenseI32ArrayAttr getDenseI32ArrayAttr(OpAsmParser& parser, ArrayRef<int3
   return parser.getBuilder().getDenseI32ArrayAttr(values);
 }
 
+static IntegerAttr getI32Attr(OpAsmParser& parser, int32_t value) {
+  return parser.getBuilder().getI32IntegerAttr(value);
+}
+
+static bool parseOptionalKeywordAlias(OpAsmParser& parser, StringRef preferred, StringRef legacy) {
+  return succeeded(parser.parseOptionalKeyword(preferred)) || succeeded(parser.parseOptionalKeyword(legacy));
+}
+
+static void
+printBoundValueList(OpAsmPrinter& printer, ValueRange arguments, ValueRange operands, ListDelimiter delimiter) {
+  printCompressedValueList(printer, arguments, delimiter);
+  printer << " = ";
+  printCompressedValueList(printer, operands, delimiter);
+}
+
+static ParseResult parseBoundValueList(OpAsmParser& parser,
+                                       ListDelimiter delimiter,
+                                       SmallVectorImpl<OpAsmParser::Argument>& arguments,
+                                       SmallVectorImpl<OpAsmParser::UnresolvedOperand>& operands) {
+  if (parseOpenDelimiter(parser, delimiter))
+    return failure();
+  if (succeeded(parseOptionalCloseDelimiter(parser, delimiter))) {
+    if (parser.parseEqual() || parseCompressedOperandList(parser, delimiter, operands))
+      return failure();
+    return success();
+  }
+
+  if (parseOneCompressedArgumentEntry(parser, arguments))
+    return failure();
+  while (succeeded(parser.parseOptionalComma()))
+    if (parseOneCompressedArgumentEntry(parser, arguments))
+      return failure();
+
+  auto parseCloseDelimiter = [&](ListDelimiter currentDelimiter) -> ParseResult {
+    switch (currentDelimiter) {
+    case ListDelimiter::Paren:  return parser.parseRParen();
+    case ListDelimiter::Square: return parser.parseRSquare();
+    }
+    llvm_unreachable("unsupported delimiter");
+  };
+  if (parseCloseDelimiter(delimiter) || parser.parseEqual() || parseCompressedOperandList(parser, delimiter, operands))
+    return failure();
+  return success();
+}
+
 static void printCoreIdList(OpAsmPrinter& printer, StringRef keyword, ArrayRef<int32_t> coreIds) {
   printer << " " << keyword << " ";
   printCompressedIntegerList(printer, coreIds);
 }
 
-static ParseResult parseOptionalCoreIdList(OpAsmParser& parser, StringRef keyword, SmallVectorImpl<int32_t>& coreIds) {
-  if (failed(parser.parseOptionalKeyword(keyword)))
-    return success();
-  return parseCompressedIntegerList(parser, coreIds);
-}
-
 } // namespace
 
-void PimCoreBatchOp::print(OpAsmPrinter& printer) {
+void PimCoreOp::print(OpAsmPrinter& printer) {
-  printer << " lanes " << getLaneCount() << " ";
+  SmallVector<Value> weightArgs;
-  size_t weightsPerLane = getLaneCount() > 0 ? getWeights().size() / static_cast<size_t>(getLaneCount()) : 0;
+  weightArgs.reserve(getWeights().size());
-  if (getLaneCount() > 1 && hasRepeatedTuple(getWeights(), weightsPerLane))
+  for (unsigned index = 0; index < getWeights().size(); ++index)
-    printValueTupleRun(printer, getWeights(), weightsPerLane, ListDelimiter::Paren);
+    weightArgs.push_back(getWeightArgument(index));
-  else
+
-    printCompressedValueList(printer, getWeights(), ListDelimiter::Paren);
   printer << " ";
-  printCompressedValueList(printer, getInputs(), ListDelimiter::Square);
+  printBoundValueList(printer, weightArgs, getWeights(), ListDelimiter::Square);
+  printer << " coreId " << getCoreId();
+  printer.printOptionalAttrDict((*this)->getAttrs(), {getCoreIdAttrName().getValue()});
+  printer << " : ";
+  printCompressedTypeList(printer, TypeRange(getWeights()), ListDelimiter::Square);
+  printer << " -> () ";
+  printer.printRegion(getBody(), /*printEntryBlockArgs=*/false);
+}
+
+ParseResult PimCoreOp::parse(OpAsmParser& parser, OperationState& result) {
+  SmallVector<OpAsmParser::Argument> weightArgs;
+  SmallVector<OpAsmParser::UnresolvedOperand> weights;
+  SmallVector<Type> weightTypes;
+  int32_t coreId = 0;
+
+  if (parseBoundValueList(parser, ListDelimiter::Square, weightArgs, weights))
+    return failure();
+
+  bool hasCoreId = parseOptionalKeywordAlias(parser, "coreId", "core_id");
+  if (hasCoreId && parser.parseInteger(coreId))
+    return failure();
+
+  if (parser.parseOptionalAttrDict(result.attributes) || parser.parseColon()
+      || parseCompressedRepeatedList(
+        parser, ListDelimiter::Square, weightTypes, [&](Type& type) { return parser.parseType(type); })
+      || parser.parseArrow() || parser.parseLParen() || parser.parseRParen())
+    return failure();
+
+  if (weights.size() != weightTypes.size())
+    return parser.emitError(parser.getCurrentLocation(), "number of weights and weight types must match");
+  if (weightArgs.size() != weights.size())
+    return parser.emitError(parser.getCurrentLocation(), "number of weight bindings and weight operands must match");
+  if (hasCoreId && result.attributes.get("coreId"))
+    return parser.emitError(parser.getCurrentLocation(),
+                            "coreId cannot be specified both positionally and in attr-dict");
+
+  if (parser.resolveOperands(weights, weightTypes, parser.getCurrentLocation(), result.operands))
+    return failure();
+
+  if (hasCoreId)
+    result.addAttribute("coreId", getI32Attr(parser, coreId));
+
+  Region* body = result.addRegion();
+  applyArgumentTypes(weightTypes, weightArgs);
+  return parser.parseRegion(*body, weightArgs);
+}
+
+void PimCoreBatchOp::print(OpAsmPrinter& printer) {
+  printer << " ";
+  printer.printOperand(getLaneArgument());
+  printer << " = 0 to " << getLaneCount() << " ";
+
+  SmallVector<Value> weightArgs;
+  weightArgs.reserve(getWeights().size());
+  for (unsigned index = 0; index < getWeights().size(); ++index)
+    weightArgs.push_back(getWeightArgument(index));
+  printBoundValueList(printer, weightArgs, getWeights(), ListDelimiter::Square);
+  printer << " ";
+  SmallVector<Value> inputArgs;
+  inputArgs.reserve(getInputs().size());
+  for (unsigned index = 0; index < getInputs().size(); ++index)
+    inputArgs.push_back(getInputArgument(index));
+  printBoundValueList(printer, inputArgs, getInputs(), ListDelimiter::Paren);
 
   if (auto coreIdsAttr = (*this)->getAttrOfType<DenseI32ArrayAttr>(onnx_mlir::kCoreIdsAttrName))
     printCoreIdList(printer, "coreIds", coreIdsAttr.asArrayRef());
@@ -49,51 +149,57 @@ void PimCoreBatchOp::print(OpAsmPrinter& printer) {
   printer.printOptionalAttrDict(
     (*this)->getAttrs(),
     {getLaneCountAttrName().getValue(), getOperandSegmentSizesAttrName().getValue(), onnx_mlir::kCoreIdsAttrName});
-  printer << " ";
-  printer.printRegion(getBody(), /*printEntryBlockArgs=*/false);
   printer << " : ";
-  if (getLaneCount() > 1 && hasRepeatedTuple(TypeRange(getWeights()), weightsPerLane))
+  printCompressedTypeList(printer, TypeRange(getWeights()), ListDelimiter::Square);
-    printTypeTupleRun(printer, TypeRange(getWeights()), weightsPerLane, ListDelimiter::Paren);
-  else
-    printCompressedTypeList(printer, TypeRange(getWeights()), ListDelimiter::Paren);
   printer << " ";
-  printCompressedTypeList(printer, TypeRange(getInputs()), ListDelimiter::Square);
+  printCompressedTypeList(printer, TypeRange(getInputs()), ListDelimiter::Paren);
   printer << " -> () ";
+  printer.printRegion(getBody(), /*printEntryBlockArgs=*/false);
 }
 
 ParseResult PimCoreBatchOp::parse(OpAsmParser& parser, OperationState& result) {
+  int64_t lowerBound = 0;
   int32_t laneCount = 0;
+  OpAsmParser::Argument laneArg;
+  SmallVector<OpAsmParser::Argument> weightArgs;
+  SmallVector<OpAsmParser::Argument> inputArgs;
+  SmallVector<OpAsmParser::Argument> regionArgs;
   SmallVector<OpAsmParser::UnresolvedOperand> weights;
   SmallVector<OpAsmParser::UnresolvedOperand> inputs;
   SmallVector<Type> weightTypes;
   SmallVector<Type> inputTypes;
   SmallVector<int32_t> coreIds;
 
-  if (parser.parseKeyword("lanes") || parser.parseInteger(laneCount)
+  if (parser.parseArgument(laneArg) || parser.parseEqual() || parser.parseInteger(lowerBound)
-      || parseCompressedOrTupleOperandList(parser, ListDelimiter::Paren, weights)
+      || parser.parseKeyword("to") || parser.parseInteger(laneCount))
-      || parseCompressedOperandList(parser, ListDelimiter::Square, inputs))
+    return failure();
+  if (lowerBound != 0)
+    return parser.emitError(parser.getCurrentLocation(), "core_batch currently requires a zero lower bound");
+
+  if (parseBoundValueList(parser, ListDelimiter::Square, weightArgs, weights)
+      || parseBoundValueList(parser, ListDelimiter::Paren, inputArgs, inputs))
     return failure();
 
-  bool hasCoreIds = succeeded(parser.parseOptionalKeyword("coreIds"));
+  bool hasCoreIds = parseOptionalKeywordAlias(parser, "coreIds", "core_ids");
   if (hasCoreIds && parseCompressedIntegerList(parser, coreIds))
     return failure();
 
-  if (parser.parseOptionalAttrDict(result.attributes))
+  if (parser.parseOptionalAttrDict(result.attributes) || parser.parseColon()
-    return failure();
+      || parseCompressedRepeatedList(
-
+        parser, ListDelimiter::Square, weightTypes, [&](Type& type) { return parser.parseType(type); })
-  Region* body = result.addRegion();
+      || parseCompressedRepeatedList(
-  if (parser.parseRegion(*body))
+        parser, ListDelimiter::Paren, inputTypes, [&](Type& type) { return parser.parseType(type); })
-    return failure();
+      || parser.parseArrow() || parser.parseLParen() || parser.parseRParen())
-
-  if (parser.parseColon() || parseCompressedOrTupleTypeList(parser, ListDelimiter::Paren, weightTypes)
-      || parseCompressedTypeList(parser, ListDelimiter::Square, inputTypes) || parser.parseArrow()
-      || parser.parseLParen() || parser.parseRParen())
     return failure();
 
   if (weights.size() != weightTypes.size())
     return parser.emitError(parser.getCurrentLocation(), "number of weights and weight types must match");
+  if (weightArgs.size() != weights.size())
+    return parser.emitError(parser.getCurrentLocation(), "number of weight bindings and weight operands must match");
   if (inputs.size() != inputTypes.size())
     return parser.emitError(parser.getCurrentLocation(), "number of inputs and input types must match");
+  if (inputArgs.size() != inputs.size())
+    return parser.emitError(parser.getCurrentLocation(), "number of input bindings and input operands must match");
   if (hasCoreIds && result.attributes.get(onnx_mlir::kCoreIdsAttrName))
     return parser.emitError(parser.getCurrentLocation(),
                             "coreIds cannot be specified both positionally and in attr-dict");
@@ -110,233 +216,15 @@ ParseResult PimCoreBatchOp::parse(OpAsmParser& parser, OperationState& result) {
       || parser.resolveOperands(inputs, inputTypes, parser.getCurrentLocation(), result.operands)) {
     return failure();
   }
-  return success();
-}
 
-void PimYieldOp::print(OpAsmPrinter& printer) {
+  Region* body = result.addRegion();
-  printer << " ";
+  laneArg.type = builder.getIndexType();
-  printCompressedValueSequence(printer, getOutputs());
+  regionArgs.push_back(laneArg);
-  printer.printOptionalAttrDict((*this)->getAttrs());
+  applyArgumentTypes(weightTypes, weightArgs);
-  printer << " : ";
+  llvm::append_range(regionArgs, weightArgs);
-  printCompressedTypeSequence(printer, getOutputs().getTypes());
+  applyArgumentTypes(inputTypes, inputArgs);
-}
+  llvm::append_range(regionArgs, inputArgs);
-
+  return parser.parseRegion(*body, regionArgs);
-ParseResult PimYieldOp::parse(OpAsmParser& parser, OperationState& result) {
-  SmallVector<OpAsmParser::UnresolvedOperand> outputs;
-  SmallVector<Type> outputTypes;
-
-  OpAsmParser::UnresolvedOperand firstOutput;
-  OptionalParseResult firstOutputResult = parser.parseOptionalOperand(firstOutput);
-  if (firstOutputResult.has_value()) {
-    if (failed(*firstOutputResult))
-      return failure();
-    if (parseCompressedOperandEntryWithFirst(parser, firstOutput, outputs))
-      return failure();
-    while (succeeded(parser.parseOptionalComma()))
-      if (parseOneCompressedOperandEntry(parser, outputs))
-        return failure();
-  }
-
-  if (parser.parseOptionalAttrDict(result.attributes) || parser.parseColon()
-      || parseCompressedTypeSequence(parser, outputTypes, /*allowEmpty=*/true))
-    return failure();
-
-  if (outputs.size() != outputTypes.size())
-    return parser.emitError(parser.getCurrentLocation(), "number of outputs and output types must match");
-
-  return parser.resolveOperands(outputs, outputTypes, parser.getCurrentLocation(), result.operands);
-}
-
-void PimSendBatchOp::print(OpAsmPrinter& printer) {
-  printer << " ";
-  printer.printOperand(getInput());
-  printCoreIdList(printer, "to", getTargetCoreIds());
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getTargetCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getInput().getType());
-}
-
-ParseResult PimSendBatchOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand input;
-  Type inputType;
-  SmallVector<int32_t> targetCoreIds;
-
-  if (parser.parseOperand(input) || parseOptionalCoreIdList(parser, "to", targetCoreIds)
-      || parser.parseOptionalAttrDict(result.attributes) || parser.parseColon() || parser.parseType(inputType))
-    return failure();
-
-  if (!targetCoreIds.empty() && result.attributes.get("targetCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "targetCoreIds cannot be specified both positionally and in attr-dict");
-  if (!targetCoreIds.empty())
-    result.addAttribute("targetCoreIds", getDenseI32ArrayAttr(parser, targetCoreIds));
-
-  return parser.resolveOperand(input, inputType, result.operands);
-}
-
-void PimSendTensorBatchOp::print(OpAsmPrinter& printer) {
-  printer << " ";
-  printer.printOperand(getInput());
-  printCoreIdList(printer, "to", getTargetCoreIds());
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getTargetCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getInput().getType());
-}
-
-ParseResult PimSendTensorBatchOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand input;
-  Type inputType;
-  SmallVector<int32_t> targetCoreIds;
-
-  if (parser.parseOperand(input) || parseOptionalCoreIdList(parser, "to", targetCoreIds)
-      || parser.parseOptionalAttrDict(result.attributes) || parser.parseColon() || parser.parseType(inputType))
-    return failure();
-
-  if (!targetCoreIds.empty() && result.attributes.get("targetCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "targetCoreIds cannot be specified both positionally and in attr-dict");
-  if (!targetCoreIds.empty())
-    result.addAttribute("targetCoreIds", getDenseI32ArrayAttr(parser, targetCoreIds));
-
-  return parser.resolveOperand(input, inputType, result.operands);
-}
-
-void PimSendTensorOp::print(OpAsmPrinter& printer) {
-  printer << " ";
-  printer.printOperand(getInput());
-  printCoreIdList(printer, "to", getTargetCoreIds());
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getTargetCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getInput().getType());
-}
-
-ParseResult PimSendTensorOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand input;
-  Type inputType;
-  SmallVector<int32_t> targetCoreIds;
-
-  if (parser.parseOperand(input) || parseOptionalCoreIdList(parser, "to", targetCoreIds)
-      || parser.parseOptionalAttrDict(result.attributes) || parser.parseColon() || parser.parseType(inputType))
-    return failure();
-
-  if (!targetCoreIds.empty() && result.attributes.get("targetCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "targetCoreIds cannot be specified both positionally and in attr-dict");
-  if (!targetCoreIds.empty())
-    result.addAttribute("targetCoreIds", getDenseI32ArrayAttr(parser, targetCoreIds));
-
-  return parser.resolveOperand(input, inputType, result.operands);
-}
-
-void PimReceiveTensorOp::print(OpAsmPrinter& printer) {
-  printCoreIdList(printer, "from", getSourceCoreIds());
-  printer << " into ";
-  printOpenDelimiter(printer, ListDelimiter::Paren);
-  printer.printOperand(getOutputBuffer());
-  printCloseDelimiter(printer, ListDelimiter::Paren);
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getSourceCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getOutputBuffer().getType());
-  printer << " -> ";
-  printer.printType(getOutput().getType());
-}
-
-ParseResult PimReceiveTensorOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand outputBuffer;
-  Type outputBufferType;
-  Type outputType;
-  SmallVector<int32_t> sourceCoreIds;
-
-  if (parseOptionalCoreIdList(parser, "from", sourceCoreIds) || parser.parseKeyword("into") || parser.parseLParen()
-      || parser.parseOperand(outputBuffer) || parser.parseRParen() || parser.parseOptionalAttrDict(result.attributes)
-      || parser.parseColon() || parser.parseType(outputBufferType) || parser.parseArrow()
-      || parser.parseType(outputType))
-    return failure();
-
-  if (!sourceCoreIds.empty() && result.attributes.get("sourceCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "sourceCoreIds cannot be specified both positionally and in attr-dict");
-  if (!sourceCoreIds.empty())
-    result.addAttribute("sourceCoreIds", getDenseI32ArrayAttr(parser, sourceCoreIds));
-
-  if (parser.resolveOperand(outputBuffer, outputBufferType, result.operands))
-    return failure();
-  result.addTypes(outputType);
-  return success();
-}
-
-void PimReceiveBatchOp::print(OpAsmPrinter& printer) {
-  printCoreIdList(printer, "from", getSourceCoreIds());
-  printer << " into ";
-  printOpenDelimiter(printer, ListDelimiter::Paren);
-  printer.printOperand(getOutputBuffer());
-  printCloseDelimiter(printer, ListDelimiter::Paren);
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getSourceCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getOutputBuffer().getType());
-  printer << " -> ";
-  printer.printType(getOutput().getType());
-}
-
-ParseResult PimReceiveBatchOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand outputBuffer;
-  Type outputBufferType;
-  Type outputType;
-  SmallVector<int32_t> sourceCoreIds;
-
-  if (parseOptionalCoreIdList(parser, "from", sourceCoreIds) || parser.parseKeyword("into") || parser.parseLParen()
-      || parser.parseOperand(outputBuffer) || parser.parseRParen() || parser.parseOptionalAttrDict(result.attributes)
-      || parser.parseColon() || parser.parseType(outputBufferType) || parser.parseArrow()
-      || parser.parseType(outputType))
-    return failure();
-
-  if (!sourceCoreIds.empty() && result.attributes.get("sourceCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "sourceCoreIds cannot be specified both positionally and in attr-dict");
-  if (!sourceCoreIds.empty())
-    result.addAttribute("sourceCoreIds", getDenseI32ArrayAttr(parser, sourceCoreIds));
-
-  if (parser.resolveOperand(outputBuffer, outputBufferType, result.operands))
-    return failure();
-  result.addTypes(outputType);
-  return success();
-}
-
-void PimReceiveTensorBatchOp::print(OpAsmPrinter& printer) {
-  printCoreIdList(printer, "from", getSourceCoreIds());
-  printer << " into ";
-  printOpenDelimiter(printer, ListDelimiter::Paren);
-  printer.printOperand(getOutputBuffer());
-  printCloseDelimiter(printer, ListDelimiter::Paren);
-  printer.printOptionalAttrDict((*this)->getAttrs(), {getSourceCoreIdsAttrName().getValue()});
-  printer << " : ";
-  printer.printType(getOutputBuffer().getType());
-  printer << " -> ";
-  printer.printType(getOutput().getType());
-}
-
-ParseResult PimReceiveTensorBatchOp::parse(OpAsmParser& parser, OperationState& result) {
-  OpAsmParser::UnresolvedOperand outputBuffer;
-  Type outputBufferType;
-  Type outputType;
-  SmallVector<int32_t> sourceCoreIds;
-
-  if (parseOptionalCoreIdList(parser, "from", sourceCoreIds) || parser.parseKeyword("into") || parser.parseLParen()
-      || parser.parseOperand(outputBuffer) || parser.parseRParen() || parser.parseOptionalAttrDict(result.attributes)
-      || parser.parseColon() || parser.parseType(outputBufferType) || parser.parseArrow()
-      || parser.parseType(outputType))
-    return failure();
-
-  if (!sourceCoreIds.empty() && result.attributes.get("sourceCoreIds"))
-    return parser.emitError(parser.getCurrentLocation(),
-                            "sourceCoreIds cannot be specified both positionally and in attr-dict");
-  if (!sourceCoreIds.empty())
-    result.addAttribute("sourceCoreIds", getDenseI32ArrayAttr(parser, sourceCoreIds));
-
-  if (parser.resolveOperand(outputBuffer, outputBufferType, result.operands))
-    return failure();
-  result.addTypes(outputType);
-  return success();
 }
 
 void PimConcatOp::print(OpAsmPrinter& printer) {

src/PIM/Dialect/Pim/PimOpsVerify.cpp

@@ -1,9 +1,14 @@
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/IR/Block.h"
+#include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/BuiltinTypeInterfaces.h"
 #include "mlir/IR/Diagnostics.h"
+#include "mlir/IR/OpDefinition.h"
 #include "mlir/IR/TypeUtilities.h"
 
 #include "llvm/Support/LogicalResult.h"
 
+#include "src/Accelerators/PIM/Common/IR/AddressAnalysis.hpp"
 #include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
 #include "src/Accelerators/PIM/Dialect/Pim/PimOps.hpp"
 
@@ -14,6 +19,63 @@ namespace pim {
 
 namespace {
 
+static bool isExplicitHostOperand(Operation* op, unsigned operandIndex) {
+  if (isa<PimMemCopyHostToDevOp>(op))
+    return operandIndex == 3;
+  if (isa<PimMemCopyHostToDevBatchOp>(op))
+    return operandIndex == 1;
+  if (isa<PimMemCopyDevToHostOp>(op))
+    return operandIndex == 2;
+  return false;
+}
+
+static Region* getParentRegion(Value value) {
+  if (auto blockArgument = dyn_cast<BlockArgument>(value))
+    return blockArgument.getParentRegion();
+  Operation* definingOp = value.getDefiningOp();
+  return definingOp ? definingOp->getParentRegion() : nullptr;
+}
+
+static bool isDefinedInsideRegion(Value value, Region& region) {
+  Region* parentRegion = getParentRegion(value);
+  return parentRegion && (&region == parentRegion || region.isAncestor(parentRegion));
+}
+
+static bool isConstantExternalValue(Value value) {
+  Operation* definingOp = value.getDefiningOp();
+  if (!definingOp)
+    return false;
+  if (definingOp->hasTrait<OpTrait::ConstantLike>())
+    return true;
+
+  auto getGlobalOp = dyn_cast<memref::GetGlobalOp>(definingOp);
+  if (!getGlobalOp)
+    return false;
+
+  auto moduleOp = definingOp->getParentOfType<ModuleOp>();
+  auto globalOp = lookupGlobalForGetGlobal(moduleOp, getGlobalOp);
+  return globalOp && globalOp.getConstant();
+}
+
+static LogicalResult verifyOnlyConstantExternalValues(Operation* ownerOp, Region& region, StringRef kind) {
+  bool hasFailure = false;
+  region.walk([&](Operation* op) {
+    for (OpOperand& operand : op->getOpOperands()) {
+      Value value = operand.get();
+      if (isDefinedInsideRegion(value, region) || isConstantExternalValue(value)
+          || isExplicitHostOperand(op, operand.getOperandNumber()))
+        continue;
+
+      InFlightDiagnostic diagnostic = ownerOp->emitOpError()
+                                   << kind << " body may only directly reference external constants";
+      diagnostic.attachNote(op->getLoc())
+        << "non-constant external operand #" << operand.getOperandNumber() << " is used by " << op->getName();
+      hasFailure = true;
+    }
+  });
+  return success(!hasFailure);
+}
+
 static bool haveSameShapedContainerKind(Type lhs, Type rhs) {
   return (isa<RankedTensorType>(lhs) && isa<RankedTensorType>(rhs)) || (isa<MemRefType>(lhs) && isa<MemRefType>(rhs));
 }
@@ -28,97 +90,41 @@ static LogicalResult verifyCompatibleShapedTypes(Operation* op, Type lhs, Type r
   return success();
 }
 
-static LogicalResult verifyTensorCommunication(Operation* op, Type type, ArrayRef<int32_t> coreIds, StringRef kind) {
+static FailureOr<ArrayRef<int64_t>> getWeightShapeForVMM(Value weight) {
-  if (coreIds.empty())
+  auto shapedType = dyn_cast<ShapedType>(weight.getType());
-    return op->emitError() << kind << " must carry at least one chunk";
+  if (!shapedType)
-
-  auto shapedType = dyn_cast<ShapedType>(type);
-  if (!shapedType || !shapedType.hasStaticShape())
-    return op->emitError() << kind << " requires a static shaped tensor or memref";
-
-  int64_t elementBits = shapedType.getElementTypeBitWidth();
-  if (elementBits <= 0 || elementBits % 8 != 0)
-    return op->emitError() << kind << " requires byte-sized elements";
-
-  int64_t totalBytes = shapedType.getNumElements() * elementBits / 8;
-  if (totalBytes % static_cast<int64_t>(coreIds.size()) != 0)
-    return op->emitError() << kind << " tensor byte size must be divisible by the number of core ids";
-
-  return success();
-}
-
-static LogicalResult
-verifyTensorBatchCommunication(Operation* op, Type type, ArrayRef<int32_t> coreIds, StringRef kind) {
-  if (coreIds.empty())
-    return op->emitError() << kind << " must carry at least one chunk";
-
-  auto coreBatchOp = op->getParentOfType<PimCoreBatchOp>();
-  if (!coreBatchOp)
-    return op->emitError() << kind << " must be nested inside pim.core_batch";
-
-  int32_t laneCount = coreBatchOp.getLaneCount();
-  if (laneCount <= 0)
-    return op->emitError() << kind << " requires a positive parent laneCount";
-  if (coreIds.size() % static_cast<size_t>(laneCount) != 0)
-    return op->emitError() << kind << " core id count must be divisible by the parent laneCount";
-
-  auto shapedType = dyn_cast<ShapedType>(type);
-  if (!shapedType || !shapedType.hasStaticShape())
-    return op->emitError() << kind << " requires a static shaped tensor or memref";
-
-  int64_t elementBits = shapedType.getElementTypeBitWidth();
-  if (elementBits <= 0 || elementBits % 8 != 0)
-    return op->emitError() << kind << " requires byte-sized elements";
-
-  int64_t chunkCount = static_cast<int64_t>(coreIds.size()) / laneCount;
-  int64_t totalBytes = shapedType.getNumElements() * elementBits / 8;
-  if (totalBytes % chunkCount != 0)
-    return op->emitError() << kind << " tensor byte size must be divisible by the chunk count per lane";
-
-  return success();
-}
-
-static FailureOr<ArrayRef<int64_t>> getWeightShapeForVMM(Operation* op, size_t weightIndex) {
-  if (auto coreOp = op->getParentOfType<PimCoreOp>()) {
-    if (weightIndex >= coreOp.getWeights().size())
-      return failure();
-    return cast<ShapedType>(coreOp.getWeights()[weightIndex].getType()).getShape();
-  }
-
-  if (auto coreBatchOp = op->getParentOfType<PimCoreBatchOp>()) {
-    if (weightIndex >= coreBatchOp.getWeights().size())
-      return failure();
-    return cast<ShapedType>(coreBatchOp.getWeights()[weightIndex].getType()).getShape();
-  }
-
     return failure();
+  return shapedType.getShape();
 }
 
 } // namespace
 
-LogicalResult PimSendTensorOp::verify() {
+LogicalResult PimCoreOp::verify() {
-  return verifyTensorCommunication(getOperation(), getInput().getType(), getTargetCoreIds(), "send_tensor");
+  Block& block = getBody().front();
+  if (block.getNumArguments() != getWeights().size())
+    return emitError("core body must have one block argument per weight");
+  for (auto [weightIndex, weight] : llvm::enumerate(getWeights()))
+    if (getWeightArgument(weightIndex).getType() != weight.getType())
+      return emitError("core weight block argument types must match weight operand types exactly");
+  return verifyOnlyConstantExternalValues(getOperation(), getBody(), "pim.core");
 }
 
-LogicalResult PimSendTensorBatchOp::verify() {
+LogicalResult PimCoreBatchOp::verify() {
-  return verifyTensorBatchCommunication(getOperation(), getInput().getType(), getTargetCoreIds(), "send_tensor_batch");
+  if (getLaneCount() <= 0)
-}
+    return emitError("laneCount must be positive");
-
+  Block& block = getBody().front();
-LogicalResult PimReceiveTensorOp::verify() {
+  unsigned expectedArgCount = 1 + getWeights().size() + getInputs().size();
-  if (failed(verifyCompatibleShapedTypes(
+  if (block.getNumArguments() != expectedArgCount)
-        getOperation(), getOutputBuffer().getType(), getOutput().getType(), "output buffer and output must match")))
+    return emitError("core_batch body must have lane, weight, and input block arguments");
-    return failure();
+  if (!getLaneArgument().getType().isIndex())
-
+    return emitError("core_batch first block argument must have index type");
-  return verifyTensorCommunication(getOperation(), getOutput().getType(), getSourceCoreIds(), "receive_tensor");
+  for (auto [weightIndex, weight] : llvm::enumerate(getWeights()))
-}
+    if (getWeightArgument(weightIndex).getType() != weight.getType())
-
+      return emitError("core_batch weight block argument types must match weight operand types exactly");
-LogicalResult PimReceiveTensorBatchOp::verify() {
+  for (auto [inputIndex, input] : llvm::enumerate(getInputs()))
-  if (failed(verifyCompatibleShapedTypes(
+    if (getInputArgument(inputIndex).getType() != input.getType())
-        getOperation(), getOutputBuffer().getType(), getOutput().getType(), "output buffer and output must match")))
+      return emitError("core_batch input block argument types must match input operand types exactly");
-    return failure();
+  return verifyOnlyConstantExternalValues(getOperation(), getBody(), "pim.core_batch");
-
-  return verifyTensorBatchCommunication(
-    getOperation(), getOutput().getType(), getSourceCoreIds(), "receive_tensor_batch");
 }
 
 LogicalResult PimVMMOp::verify() {
@@ -126,9 +132,9 @@ LogicalResult PimVMMOp::verify() {
         getOperation(), getOutputBuffer().getType(), getOutput().getType(), "output buffer and output must match")))
     return failure();
 
-  auto matrixShapeOpt = getWeightShapeForVMM(getOperation(), getWeightIndex());
+  auto matrixShapeOpt = getWeightShapeForVMM(getWeight());
   if (failed(matrixShapeOpt))
-    return emitError("must be nested inside pim.core or pim.core_batch with a valid weightIndex");
+    return emitError("weight must be a shaped value");
   ArrayRef<int64_t> matrixShape = *matrixShapeOpt;
 
   auto vectorType = dyn_cast<ShapedType>(getInput().getType());

src/PIM/Dialect/Pim/Transforms/Bufferization/BufferizationUtils.cpp

@@ -17,7 +17,7 @@ Value materializeContiguousMemRef(Value memrefValue, Location loc, RewriterBase&
   auto shapedType = cast<ShapedType>(memrefValue.getType());
   auto contiguousType = MemRefType::get(shapedType.getShape(), shapedType.getElementType());
   Value contiguousBuffer = memref::AllocOp::create(rewriter, loc, contiguousType);
-  auto sizeInBytes = shapedType.getNumElements() * shapedType.getElementTypeBitWidth() / 8;
+  auto sizeInBytes = getShapedTypeSizeInBytes(shapedType);
 
   return PimMemCopyOp::create(rewriter,
                               loc,
@@ -30,6 +30,15 @@ Value materializeContiguousMemRef(Value memrefValue, Location loc, RewriterBase&
     .getOutput();
 }
 
+Value allocateContiguousMemRefLike(Value memrefValue, Location loc, RewriterBase& rewriter) {
+  if (succeeded(resolveContiguousAddress(memrefValue)))
+    return memrefValue;
+
+  auto shapedType = cast<ShapedType>(memrefValue.getType());
+  auto contiguousType = MemRefType::get(shapedType.getShape(), shapedType.getElementType());
+  return memref::AllocOp::create(rewriter, loc, contiguousType);
+}
+
 FailureOr<Value>
 getBufferOrValue(RewriterBase& rewriter, Value value, const BufferizationOptions& options, BufferizationState& state) {
   if (isa<BufferLikeType>(value.getType()))

src/PIM/Dialect/Pim/Transforms/Bufferization/BufferizationUtils.hpp

@@ -6,6 +6,7 @@
 namespace onnx_mlir::pim {
 
 mlir::Value materializeContiguousMemRef(mlir::Value memrefValue, mlir::Location loc, mlir::RewriterBase& rewriter);
+mlir::Value allocateContiguousMemRefLike(mlir::Value memrefValue, mlir::Location loc, mlir::RewriterBase& rewriter);
 
 llvm::FailureOr<mlir::Value> getBufferOrValue(mlir::RewriterBase& rewriter,
                                               mlir::Value value,

src/PIM/Dialect/Pim/Transforms/Bufferization/Common.cpp

@@ -1,9 +1,10 @@
 #include "Dialect/Pim/Transforms/Bufferization/Common.hpp"
+#include "src/Accelerators/PIM/Common/PimCommon.hpp"
 
 using namespace mlir;
 
 IntegerAttr onnx_mlir::pim::getMemRefSizeInBytesAttr(OpBuilder& builder, Value memref) {
   auto type = mlir::cast<MemRefType>(memref.getType());
-  int32_t sizeInBytes = static_cast<int32_t>(type.getNumElements() * type.getElementTypeBitWidth() / 8);
+  int32_t sizeInBytes = static_cast<int32_t>(getShapedTypeSizeInBytes(type));
   return builder.getI32IntegerAttr(sizeInBytes);
 }