[Cadence: Vision] ResNet18 & ResNet50: Optimized, DMA-enabled, functional

.gitignore

@@ -32,7 +32,6 @@ build-profiling/
 *.model
 *.etdump
 tokenizer.json
-*.pte
 *.ptd
 !test_bpe_tokenizer.bin
 !test_tiktoken_tokenizer.model
@@ -69,6 +68,11 @@ xcuserdata/
 /src/executorch/share/
 /src/executorch/version.py
 *_etdump
+/runtime/core/portable_type/c10/CMakeFiles/
+/runtime/core/portable_type/c10/bin/
+/runtime/core/portable_type/c10/Makefile
+/runtime/core/portable_type/c10/cmake_install.cmake
+/runtime/core/portable_type/c10/*.a
 
 # Android
 *.aar

CMakeLists.txt

@@ -241,13 +241,21 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -s")
 endif()
 
-if(EXECUTORCH_OPTIMIZE_SIZE)
-  # -Os: Optimize for size.
-  set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -Os")
+if(NOT EXECUTORCH_BUILD_CADENCE)
+  if(OPTIMIZE_SIZE)
+    # -Os: Optimize for size
+    set(CMAKE_CXX_FLAGS_RELEASE "-Os ${CMAKE_CXX_FLAGS_RELEASE}")
+  else()
+    # -O2: Moderate opt.
+    set(CMAKE_CXX_FLAGS_RELEASE "-O2 ${CMAKE_CXX_FLAGS_RELEASE}")
+  endif()
 else()
-  # -O2: Moderate opt.
-  set(CMAKE_CXX_FLAGS_RELEASE "-O2 ${CMAKE_CXX_FLAGS_RELEASE}")
+  set(CMAKE_CXX_FLAGS_RELEASE
+    "-O3 -mcoproc -mlongcalls -LNO:simd  -ffunction-sections -fsigned-char -fno-exceptions -INLINE:requested -fno-zero-initialized-in-bss -mtext-section-literals -fmessage-length=0")
+  set(CMAKE_C_FLAGS_RELEASE
+    "-O3 -mcoproc -mlongcalls -LNO:simd  -ffunction-sections -fsigned-char -fno-exceptions -INLINE:requested -fno-zero-initialized-in-bss -mtext-section-literals -fmessage-length=0")
 endif()
+set(CMAKE_CXX_FLAGS_DEBUG "-O0 -g")
 
 if(EXECUTORCH_BUILD_TESTS)
   include(CTest)

backends/cadence/CMakeLists.txt

@@ -90,6 +90,10 @@ elseif(EXECUTORCH_FUSION_G3_OPT)
   )
 elseif(EXECUTORCH_VISION_OPT)
   set(TARGET_DIR vision)
+  add_subdirectory(
+    ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party
+    ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10
+  )
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
 else()
   set(TARGET_DIR generic)

backends/cadence/aot/functions_vision.yaml

@@ -85,12 +85,12 @@
 - op: max_pool2d_with_indices.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::max_pool2d_with_indices_out
+      kernel_name: impl::vision::max_pool2d_with_indices_out
 
 - op: mean.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::mean_dim_out
+      kernel_name: impl::vision::mean_dim_out
 
 - op: mul.out
   kernels:
@@ -205,6 +205,16 @@
     - arg_meta: null
       kernel_name: impl::vision::quantized_conv2d_nhwc_out
 
+- func: cadence::quantized_conv2d_nchw.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::vision::quantized_conv2d_nchw_per_tensor_out
+
+- func: cadence::quantized_conv2d_nhwc.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::vision::quantized_conv2d_nhwc_per_tensor_out
+
 - func: cadence::quantized_layer_norm.out(Tensor input, Tensor in_scale, Tensor in_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
@@ -249,10 +259,6 @@
     - arg_meta: null
       kernel_name: impl::vision::im2row_per_tensor_out
 
-- func: cadence::quantized_conv.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
-  kernels:
-    - arg_meta: null
-      kernel_name: impl::vision::quantized_conv_per_tensor_out
 
 - func: cadence::quantized_fully_connected.out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
   kernels:

backends/cadence/vision/config_generator/README.md

@@ -0,0 +1,107 @@
+# Config Generator Python
+
+Python tools for extracting convolution layer parameters from neural network models and generating optimized C header configurations for DMA-tiled execution on the Xtensa XRC Vision DSP (XAI CNN runtime).
+
+## Prerequisites
+
+The script requires the Python venv in the executorch tree and must be run from a **bash** terminal (not csh):
+
+```bash
+# The venv is at <executorch>/.venv/
+# All paths below are relative to the executorch root.
+
+# Option 1: call the venv python directly (works from any shell)
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py ...
+
+# Option 2: activate the venv in a bash shell
+bash
+source .venv/bin/activate
+python3 backends/cadence/vision/config_generator/generate_layer_configs.py ...
+```
+
+> **Note:** The default terminal on this machine is `csh`. Inline python commands
+> and `source ... && ...` chains will fail in csh. Always use `bash` or invoke
+> the venv python by its full path.
+
+## Quick Start
+
+```bash
+# Run from the executorch root directory: cd <executorch>
+
+# From a single ExecuTorch .pte binary
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py \
+    --pte operator_and_model_testing/resnet18/pte/resnet18_quantized.pte \
+    --output backends/cadence/vision/config_generator/conv_layer_configs.h \
+    --dram0 62976 --dram1 62976
+
+# From multiple .pte files (layers are deduplicated automatically)
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py \
+    --pte operator_and_model_testing/resnet18/pte/resnet18_quantized.pte \
+         operator_and_model_testing/resnet50/pte/resnet50_quantized.pte \
+    --output backends/cadence/vision/config_generator/conv_layer_configs_combined.h \
+    --dram0 62976 --dram1 62976
+
+# From a torchvision model (requires torchvision installed in venv)
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py \
+    --model resnet18 --input-size 1,3,64,64 \
+    --output backends/cadence/vision/config_generator/conv_layer_configs.h \
+    --dram0 32768 --dram1 32768
+```
+
+### Full working commands
+
+```bash
+# cd to the executorch root first
+cd <path-to-executorch>
+
+# ResNet18 with 62976 bytes per DRAM bank
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py \
+    --pte operator_and_model_testing/resnet18/pte/resnet18_quantized.pte \
+    --output backends/cadence/vision/config_generator/conv_layer_configs_62k_pte.h \
+    --dram0 62976 --dram1 62976
+
+# ResNet18 + ResNet50 combined
+.venv/bin/python3 backends/cadence/vision/config_generator/generate_layer_configs.py \
+    --pte operator_and_model_testing/resnet18/pte/resnet18_quantized.pte \
+         operator_and_model_testing/resnet50/pte/resnet50_quantized.pte \
+    --output backends/cadence/vision/config_generator/conv_layer_configs_62k_combined.h \
+    --dram0 62976 --dram1 62976
+```
+
+---
+
+## `generate_layer_configs.py` — Arguments
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--model`, `-m` | — | Comma or `+`-separated torchvision model names (e.g. `resnet18+resnet50`) |
+| `--pte` | — | Path to an ExecuTorch `.pte` binary; bootstraps `exir._serialize` from the local source tree — no pip install needed |
+| `--flatc` | cmake-out default | Path to `flatc` binary (auto-detected; only relevant with `--pte`) |
+| `--input-size` | `1,3,64,64` | Input tensor shape `N,C,H,W` (only used with `--model`) |
+| `--output`, `-o` | `conv_layer_configs.h` | Output C header file |
+| `--dram0` | `32768` | DRAM0 size in bytes |
+| `--dram1` | `32768` | DRAM1 size in bytes |
+| `--cache-mode` | off | Append `_cache` to every kernel name |
+
+---
+
+## Output
+
+The generated header contains:
+
+- `conv_layer_config_t` struct with ~60 fields (buffer sizes, tile dimensions, DRAM0/1 placement, kernel name, quantization params)
+- `CONV_LAYER_CONFIGS[]` static array — one entry per unique layer
+- `get_layer_config()`, `get_layer_config_by_params()`, `get_layer_config_by_key()` inline accessors
+
+---
+
+## Directory Structure
+
+```
+config_generator_python/
+├── generate_layer_configs.py    # Main entry point
+├── generate_idma_buffers.py     # Core tiling / buffer sizing engine
+├── extract_layers_from_pte.py   # .pte/.onnx → JSON (intermediate step)
+├── config/                      # Pre-generated headers
+└── bin/                         # Compare / test utilities
+```