Update on "[Executorch] Make apple take accelerate path for blas"

For some reason fbobjc_exported_preprocessor_flags does not seems to work. Thus
taking 'select' route which seems to work.

Differential Revision: [D64499608](https://our.internmc.facebook.com/intern/diff/D64499608/)

[ghstack-poisoned]

Author: kimishpatel

.github/workflows/_android.yml

@@ -66,6 +66,16 @@ jobs:
           # avoid permission issue
           sudo chown -R "${USER}" /opt/android
 
+      - name: Download Artifacts
+        shell: bash
+        run: |
+          set -eux
+          curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug.apk
+          curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug-androidTest.apk
+          curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
+          unzip model.zip
+          mv *.pte model.pte
+
       - name: Gradle cache
         uses: gradle/actions/setup-gradle@v3
 

.github/workflows/pull.yml

@@ -99,6 +99,8 @@ jobs:
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 900
+      upload-artifact: android-models
+      upload-artifact-to-s3: true
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -107,13 +109,15 @@ jobs:
         DTYPE=${{ matrix.dtype }}
         BUILD_TOOL="cmake"
         MODE=${{ matrix.mode }}
+        ARTIFACTS_DIR_NAME="artifacts-to-be-uploaded/${DTYPE}-${MODE}"
+        ARTIFACTS_DIR_NAME="${ARTIFACTS_DIR_NAME/+/-}"
 
         # Setup executorch
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
         # Install requirements for export_llama
         PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
         # Test llama2
-        PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}" "${ARTIFACTS_DIR_NAME}"
 
   test-llama-runner-linux-android:
     name: test-llama-runner-linux-android
@@ -320,6 +324,7 @@ jobs:
 
   android:
     uses: ./.github/workflows/_android.yml
+    needs: test-llama-runner-linux
 
   unittest:
     uses: ./.github/workflows/_unittest.yml

backends/cadence/aot/functions_hifi.yaml

@@ -25,7 +25,7 @@
 - op: add.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::add_out
+      kernel_name: cadence::impl::HiFi::add_out
 
 - op: bmm.out
   kernels:
@@ -45,12 +45,12 @@
 - op: div.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::div_out
+      kernel_name: cadence::impl::HiFi::div_out
 
 - op: div.out_mode
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::div_out_mode
+      kernel_name: cadence::impl::HiFi::div_out_mode
 
 - op: embedding.out
   kernels:
@@ -65,7 +65,7 @@
 - op: mul.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::mul_out
+      kernel_name: cadence::impl::HiFi::mul_out
 
 - op: permute_copy.out
   kernels:
@@ -75,7 +75,7 @@
 - op: sigmoid.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::sigmoid_out
+      kernel_name: cadence::impl::HiFi::sigmoid_out
 
 - op: slice_copy.Tensor_out
   kernels:
@@ -90,7 +90,12 @@
 - op: sub.out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::sub_out
+      kernel_name: cadence::impl::HiFi::sub_out
+
+- op: tanh.out
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::tanh_out
 
 - op: view_copy.out
   kernels:

backends/cadence/cadence.cmake

@@ -43,6 +43,9 @@ set(CMAKE_CXX_COMPILER ${TOOLCHAIN_HOME}/bin/${CROSS_COMPILE_TARGET}-clang++)
 
 set(CMAKE_C_FLAGS_INIT "-stdlib=libc++ -mtext-section-literals -mlongcalls")
 set(CMAKE_CXX_FLAGS_INIT "-stdlib=libc++ -mtext-section-literals -mlongcalls")
+#workaround for larger compilation time
+set(CMAKE_CXX_FLAGS_INIT "${CMAKE_CXX_FLAGS_INIT} -fno-strict-aliasing")
+
 set(CMAKE_SYSROOT ${TOOLCHAIN_HOME}/${SYSROOT_TARGET})
 set(CMAKE_LINKER ${TOOLCHAIN_HOME}/bin/xt-ld)
 add_link_options(-lm -stdlib=libc++ -Wl,--no-as-needed -static)

backends/cadence/hifi/kernels/CMakeLists.txt

@@ -9,6 +9,10 @@ add_library(
   cadence_kernels
   kernels.cpp
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/matmul_asym8uxasym8u_asym8u.cpp
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_add_f32_broadcast.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_f32_broadcast.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_mode_f32_broadcast.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_mul_f32_broadcast.c
 )
 # Let files say "include <executorch/path/to/header.h>".
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)

backends/cadence/hifi/kernels/kernels.h

@@ -11,6 +11,49 @@
 #include <inttypes.h>
 #include <stddef.h>
 #include <xa_type_def.h>
+/* For NNLIB APIs */
+#include "xa_nnlib_kernels_api.h"
+
+/* Potential NNLIB function/APIs */
+extern "C" WORD32 xa_nn_elm_add_broadcast_4D_f32xf32_f32(
+    FLOAT32* __restrict__ p_out,
+    const WORD32* const p_out_shape,
+    const FLOAT32* __restrict__ p_inp1,
+    const WORD32* const p_inp1_shape,
+    const FLOAT32* __restrict__ p_inp2,
+    const WORD32* const p_inp2_shape);
+
+extern "C" WORD32 xa_nn_elm_div_broadcast_4D_f32xf32_f32(
+    FLOAT32* __restrict__ p_out,
+    const WORD32* const p_out_shape,
+    const FLOAT32* __restrict__ p_inp1,
+    const WORD32* const p_inp1_shape,
+    const FLOAT32* __restrict__ p_inp2,
+    const WORD32* const p_inp2_shape);
+
+extern "C" WORD32 xa_nn_elm_div_mode_f32xf32_f32(
+    FLOAT32* __restrict__ p_out,
+    const FLOAT32* __restrict__ p_inp1,
+    const FLOAT32* __restrict__ p_inp2,
+    WORD32 num_elm,
+    WORD32 mode);
+
+extern "C" WORD32 xa_nn_elm_div_mode_broadcast_4D_f32xf32_f32(
+    FLOAT32* __restrict__ p_out,
+    const WORD32* const p_out_shape,
+    const FLOAT32* __restrict__ p_inp1,
+    const WORD32* const p_inp1_shape,
+    const FLOAT32* __restrict__ p_inp2,
+    const WORD32* const p_inp2_shape,
+    WORD32 mode);
+
+extern "C" WORD32 xa_nn_elm_mul_broadcast_4D_f32xf32_f32(
+    FLOAT32* __restrict__ p_out,
+    const WORD32* const p_out_shape,
+    const FLOAT32* __restrict__ p_inp1,
+    const WORD32* const p_inp1_shape,
+    const FLOAT32* __restrict__ p_inp2,
+    const WORD32* const p_inp2_shape);
 
 namespace cadence {
 namespace impl {

backends/cadence/hifi/operators/CMakeLists.txt

@@ -20,6 +20,12 @@ endif()
 
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_add.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_div.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_mul.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_sigmoid.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_sub.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_tanh.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
@@ -29,24 +35,29 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/repeat_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/slice_util.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_add.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_bmm.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_cat.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_clone.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_div.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_embedding.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_full.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_mul.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_permute_copy.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sigmoid.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_slice_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_softmax.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_split_with_sizes_copy.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sub.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_to_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_view_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_where.cpp"
-)
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/dtype_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/index_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/kernel_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/repeat_util.cpp"
+    )
 add_library(aten_ops_cadence ${_aten_ops__srcs})
 target_link_libraries(aten_ops_cadence PUBLIC executorch)
 target_link_libraries(aten_ops_cadence PRIVATE cadence_kernels)