Add: Initial backends/cadence/vision module scaffold with optimized softmax (non-iDMA)

Author: cad-rlc

backends/cadence/CMakeLists.txt

@@ -84,10 +84,18 @@ if(EXECUTORCH_NNLIB_OPT)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
 elseif(EXECUTORCH_FUSION_G3_OPT)
   set(TARGET_DIR fusion_g3)
+<<<<<<< HEAD
   add_subdirectory(
     ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party/nnlib
     ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10
   )
+=======
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party/nnlib
+  ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10)
+elseif(EXECUTORCH_VISION_OPT)
+  set(TARGET_DIR vision)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
+>>>>>>> 992b83e373 ( Add: Initial backends/cadence/vision module scaffold with optimized softmax (non-iDMA))
 else()
   set(TARGET_DIR reference)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)

backends/cadence/aot/functions_vision.yaml

@@ -0,0 +1,265 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This yaml file contains operators that are also defined by the ATen library.
+# For lean mode:
+#   - Codegen'd target `executorch_generated_lib` will be reading all the information
+#     from this file, including operator schema and kernel metadata.
+#   - Selective build target `codegen:executorch_defined_ops` now is selecting all the
+#     operators in this file, by dumping all the op names into `selected_operators.yaml`.
+#
+# See the README.md file in executorch/kernels/portable for a description of the syntax used
+# by this file.
+
+
+# aten ops
+- op: _to_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::to_copy_out
+
+- op: _softmax.out
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::_softmax_out
+
+- op: add.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::add_out
+
+- op: bmm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::bmm_out
+
+- op: cat.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::cat_out
+
+- op: clone.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::clone_out
+
+- op: div.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::div_out
+
+- op: div.out_mode
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::div_out_mode
+
+- op: embedding.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::embedding_out
+
+- op: empty.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::empty_out
+
+- op: expand_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::expand_copy_out
+
+- op: full.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::full_out
+
+- op: gelu.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::gelu_out
+
+- op: hardtanh.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::hardtanh_out
+
+- op: max_pool2d_with_indices.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::max_pool2d_with_indices_out
+
+- op: mean.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::mean_dim_out
+
+- op: mul.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::mul_out
+
+- op: mul.Scalar_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::mul_scalar_out
+
+- op: permute_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::permute_copy_out
+
+- op: rsqrt.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::rsqrt_out
+
+- op: sigmoid.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sigmoid_out
+
+- op: slice_copy.Tensor_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::slice_copy_Tensor_out
+
+- op: split_with_sizes_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::split_with_sizes_copy_out
+
+- op: sub.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sub_out
+
+- op: view_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::view_copy_out
+
+- op: where.self_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::where_out
+
+- op: transpose_copy.int_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::transpose_copy_int_out
+
+- op: eq.Scalar_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::eq_scalar_out
+
+- op: logical_not.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::logical_not_out
+
+- op: any.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::any_out
+
+- op: native_group_norm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::native_group_norm_out
+
+- op: sum.IntList_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sum_dim_out
+
+- op: select_copy.int_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::select_copy_int_out
+
+# custom ops
+- func: cadence::quantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantize_per_tensor_out
+
+- func: cadence::dequantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::dequantize_per_tensor_out
+
+- func: cadence::quantized_conv.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_conv_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
+      kernel_name: cadence::impl::vision::quantized_layer_norm_out
+- func: cadence::quantized_layer_norm.per_tensor_out(Tensor input, float in_scale, int 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
+      kernel_name: cadence::impl::vision::quantized_layer_norm_per_tensor_out
+
+- func: cadence::quantized_linear.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:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_linear_out
+
+- func: cadence::quantized_relu.out(Tensor X, Tensor X_zero_point, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_relu_out
+
+- func: cadence::quantized_relu.per_tensor_out(Tensor X, int X_zero_point, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_relu_per_tensor_out
+
+- func: cadence::quantized_matmul.out(Tensor X, int X_zero_point, Tensor Y, int Y_zero_point, Tensor? bias, int out_multiplier, int out_shift, int out_zero_point, bool transposed, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_matmul_out
+
+- func: cadence::quantized_linear.per_tensor_out(Tensor src, Tensor weight, Tensor bias, SymInt src_zero_point, SymInt weight_zero_point, SymInt out_multiplier, SymInt out_shift, SymInt out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_linear_per_tensor_out
+
+- func: cadence::im2row.out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, Tensor in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::im2row_out
+
+- func: cadence::im2row.per_tensor_out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, int in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::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: cadence::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:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_fully_connected_out
+
+- func: cadence::quantized_fully_connected.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::quantized_fully_connected_per_tensor_out
+
+- func: cadence::requantize.out(Tensor input, Tensor in_scale, Tensor in_zero_point, Tensor out_scale, Tensor out_zero_point, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::requantize_out
+
+- func: cadence::requantize.per_tensor_out(Tensor input, float in_scale, int in_zero_point, float out_scale, int out_zero_point, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::vision::requantize_per_tensor_out

backends/cadence/vision/kernels/CMakeLists.txt

@@ -0,0 +1,29 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# lint_cmake: -linelength
+add_library(
+  cadence_kernels 
+  kernels.cpp
+  ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/library/api/tensor_transposef.c 
+  ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/library/api/vsoftmaxf.c 
+  ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/library/tables/expf_tbl.c 
+  ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/library/tables/nanf_tbl.c 
+  ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/library/tables/inff_tbl.c 
+)
+
+# Let files say "include <executorch/path/to/header.h>".
+set(_common_include_directories ${EXECUTORCH_ROOT}/..
+${EXECUTORCH_ROOT}/runtime/core/portable_type/c10)
+
+target_include_directories(
+  cadence_kernels 
+  PUBLIC 
+    .
+    ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/include 
+    ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/include_private 
+  ${_common_include_directories}
+)

backends/cadence/vision/kernels/TARGETS

@@ -0,0 +1,5 @@
+load("targets.bzl", "define_common_targets")
+
+oncall("odai_jarvis")
+
+define_common_targets()