Add iDMA support for softmax over last dimension within the softmax operator

- Added iDMA support within the softmax operator in vision/operators directory.
- Modified op_add, op_embedding and op_view_copy based on comments from last PR.
- Created a build script as an example on how build and run "add" on Vision 130.
- Replaced IVP_PACKVNRN_2X64W with IVP_PACKVRNRN_2X64W. (Instruction also available on VP6)

Author: cad-rlc

backends/cadence/aot/functions_vision.yaml

@@ -25,7 +25,7 @@
 - op: add.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::add_out
+      kernel_name: cadence::impl::vision::add_out
 
 - op: bmm.out
   kernels:
@@ -55,7 +55,7 @@
 - op: embedding.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::embedding_out
+      kernel_name: cadence::impl::vision::embedding_out
 
 - op: empty.out
   kernels:
@@ -70,7 +70,7 @@
 - op: full.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::full_out
+      kernel_name: cadence::impl::vision::full_out
 
 - op: gelu.out
   kernels:
@@ -135,7 +135,7 @@
 - op: view_copy.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::view_copy_out
+      kernel_name: cadence::impl::vision::view_copy_out
 
 - op: where.self_out
   kernels:

backends/cadence/build_cadence_vision.sh

@@ -0,0 +1,83 @@
+#!/bin/bash
+# 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.
+
+set -euo pipefail
+
+unset CMAKE_PREFIX_PATH
+unset XTENSA_CORE
+export XTENSA_CORE=XRC_Vision_130_AO
+git submodule sync
+git submodule update --init --recursive
+./install_requirements.sh
+./install_executorch.sh
+
+rm -rf cmake-out
+
+STEPWISE_BUILD=false
+
+if $STEPWISE_BUILD; then
+    echo "Building ExecuTorch"
+    CXXFLAGS="-fno-exceptions -fno-rtti" cmake -DCMAKE_INSTALL_PREFIX=cmake-out \
+        -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake  \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DEXECUTORCH_ENABLE_EVENT_TRACER=OFF \
+        -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+        -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
+        -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+        -DEXECUTORCH_BUILD_CPUINFO=OFF \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
+        -DEXECUTORCH_USE_DL=OFF \
+        -DEXECUTORCH_BUILD_CADENCE=OFF \
+        -Bcmake-out .
+
+    echo "Building any Cadence-specific binaries on top"
+    CXXFLAGS="-fno-exceptions -fno-rtti" cmake -DBUCK2="$BUCK" \
+        -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake \
+        -DCMAKE_INSTALL_PREFIX=cmake-out \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON \
+        -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+        -DEXECUTORCH_BUILD_CADENCE=ON \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
+        -DEXECUTORCH_ENABLE_PROGRAM_VERIFICATION=ON \
+        -DEXECUTORCH_USE_DL=OFF \
+        -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
+        -DEXECUTORCH_BUILD_KERNELS_LLM=OFF \
+        -DPYTHON_EXECUTABLE=python3 \
+        -DEXECUTORCH_VISION_OPT=ON \
+        -DHAVE_FNMATCH_H=OFF \
+        -Bcmake-out/backends/cadence \
+        backends/cadence
+    cmake --build cmake-out/backends/cadence  -j8
+else
+    echo "Building Cadence toolchain with ExecuTorch packages"
+    cmake_prefix_path="${PWD}/cmake-out/lib/cmake/ExecuTorch;${PWD}/cmake-out/third-party/gflags"
+    CXXFLAGS="-fno-exceptions -fno-rtti" cmake -DBUCK2="$BUCK" \
+        -DCMAKE_PREFIX_PATH="${cmake_prefix_path}" \
+        -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake \
+        -DCMAKE_INSTALL_PREFIX=cmake-out \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON \
+        -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+        -DEXECUTORCH_BUILD_CPUINFO=OFF \
+        -DEXECUTORCH_BUILD_CADENCE=ON \
+        -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+        -DEXECUTORCH_ENABLE_LOGGING=ON \
+        -DEXECUTORCH_ENABLE_PROGRAM_VERIFICATION=ON \
+        -DEXECUTORCH_USE_DL=OFF \
+        -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
+        -DEXECUTORCH_BUILD_KERNELS_LLM=OFF \
+        -DPYTHON_EXECUTABLE=python3 \
+        -DEXECUTORCH_VISION_OPT=ON \
+        -DHAVE_FNMATCH_H=OFF \
+        -Bcmake-out
+    cmake --build cmake-out --target install --config Release -j8
+fi
+
+echo "Run simple model to verify cmake build"
+python3 -m examples.portable.scripts.export --model_name="add"
+xt-run --turbo cmake-out/executor_runner  --model_path=add.pte

backends/cadence/vision/kernels/CMakeLists.txt

@@ -27,3 +27,5 @@ target_include_directories(
     ${EXECUTORCH_ROOT}/backends/cadence/vision/third-party/include_private 
   ${_common_include_directories}
 )
+
+target_link_libraries(cadence_kernels PRIVATE idma)

backends/cadence/vision/operators/op_add.cpp

@@ -11,8 +11,19 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
-namespace torch {
-namespace executor {
+using executorch::aten::ScalarType;
+using executorch::aten::Tensor;
+using executorch::aten::Scalar;
+using executorch::runtime::KernelRuntimeContext;
+using executorch::runtime::promoteTypes;
+using torch::executor::Error;
+using torch::executor::apply_binary_elementwise_fn;
+using executorch::runtime::canCast;
+using torch::executor::native::utils::extract_scalar;
+
+namespace cadence {
+namespace impl {
+namespace vision {
 namespace native {
 
 Tensor& add_out(
@@ -23,6 +34,8 @@ Tensor& add_out(
     Tensor& out) {
   (void)ctx;
 
+  using namespace torch::executor::native::utils;
+
   ScalarType a_type = a.scalar_type();
   ScalarType b_type = b.scalar_type();
   ScalarType common_type = promoteTypes(a_type, b_type);
@@ -39,7 +52,7 @@ Tensor& add_out(
   using CTYPE_IN = float;
   using CTYPE_OUT = float;
   CTYPE_IN alpha_val;
-  ET_EXTRACT_SCALAR(alpha, alpha_val);
+  ET_CHECK_MSG(extract_scalar(alpha, &alpha_val), "Could not be extracted: wrong type or out of range");
 
   apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
       [alpha_val](const CTYPE_A val_a, const CTYPE_B val_b) {
@@ -57,5 +70,6 @@ Tensor& add_out(
 }
 
 } // namespace native
-} // namespace executor
-} // namespace torch
+} // namespace vision
+} // namespace impl
+} // namespace cadence

backends/cadence/vision/operators/op_embedding.cpp

@@ -8,13 +8,14 @@
 
 #include <executorch/runtime/kernel/kernel_includes.h>
 
-namespace torch {
-namespace executor {
-namespace native {
-
 using executorch::aten::Tensor;
 using executorch::runtime::KernelRuntimeContext;
 
+namespace cadence {
+namespace impl {
+namespace vision {
+namespace native {
+
 void embedding_out(
     KernelRuntimeContext& ctx,
     const Tensor& weight,
@@ -37,5 +38,6 @@ void embedding_out(
 }
 
 } // namespace native
-} // namespace executor
-} // namespace torch
+} // namespace vision
+} // namespace impl
+} // namespace cadence

backends/cadence/vision/operators/op_full.cpp

@@ -9,12 +9,19 @@
 #include <executorch/kernels/portable/cpu/scalar_utils.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
-namespace torch {
-namespace executor {
-namespace native {
-
 using executorch::aten::ScalarType;
 using executorch::aten::Tensor;
+using executorch::aten::Scalar;
+using executorch::aten::IntArrayRef;
+using executorch::runtime::KernelRuntimeContext;
+using torch::executor::Error;
+using torch::executor::native::utils::get_scalar_dtype;
+using torch::executor::native::utils::extract_scalar;
+
+namespace cadence {
+namespace impl {
+namespace vision {
+namespace native {
 
 Tensor& full_out(
     KernelRuntimeContext& ctx,
@@ -23,15 +30,15 @@ Tensor& full_out(
     Tensor& out) {
   (void)ctx;
 
-  ScalarType val_type = utils::get_scalar_dtype(fill_value);
+  ScalarType val_type = get_scalar_dtype(fill_value);
   ScalarType out_type = out.scalar_type();
 
   Error err = resize_tensor(out, sizes);
   ET_CHECK_MSG(err == Error::Ok, "Could not resize out");
 
   ET_SWITCH_REAL_TYPES_AND(Bool, val_type, ctx, "full", CTYPE_VAL, [&] {
     CTYPE_VAL val;
-    ET_EXTRACT_SCALAR(fill_value, val);
+    ET_CHECK_MSG(extract_scalar(fill_value, &val), "Could not be extracted: wrong type or out of range");
 
     ET_SWITCH_REAL_TYPES_AND(Bool, out_type, ctx, "full", CTYPE_OUT, [&] {
       CTYPE_OUT val_casted = static_cast<CTYPE_OUT>(val);
@@ -46,5 +53,6 @@ Tensor& full_out(
 }
 
 } // namespace native
-} // namespace executor
-} // namespace torch
+} // namespace vision
+} // namespace impl
+} // namespace cadence