Merge branch 'dijopaul_int' into dijopaul_optimization2

dijopaul · Jul 31, 2024 · d344254 · d344254
2 parents 0b58c5d + 44b46df
commit d344254
Show file tree

Hide file tree

Showing 10 changed files with 2,129 additions and 2 deletions.
diff --git a/backends/cadence/aot/functions_hifi.yaml b/backends/cadence/aot/functions_hifi.yaml
@@ -101,6 +101,11 @@
   kernels:
     - arg_meta: null
       kernel_name: torch::executor::where_out
+
+- op: rsqrt.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::rsqrt_out        
 
 - op: empty.out
   kernels:

diff --git a/backends/cadence/hifi/kernels/CMakeLists.txt b/backends/cadence/hifi/kernels/CMakeLists.txt
@@ -11,6 +11,9 @@ add_library(
   ${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_mul_f32_broadcast.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_where_f32xf32_f32.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_broadcast_f32.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_floor_div_broadcast_f32.c
 )
 
 target_include_directories(

diff --git a/backends/cadence/hifi/kernels/kernels.h b/backends/cadence/hifi/kernels/kernels.h
@@ -14,7 +14,40 @@
 
 /* For NNLIB APIs */
 #include "xa_nnlib_kernels_api.h"
-
+
+
+extern "C" WORD32 xa_nn_elm_floor_div_f32xf32_f32(
+                                    FLOAT32 * __restrict__ p_out,
+                                    const FLOAT32 * __restrict__ p_inp1,
+                                    const FLOAT32 * __restrict__ p_inp2,
+                                    WORD32 num_elm);
+
+extern "C" WORD32 xa_nn_elm_floor_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_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_where_f32xf32_f32(FLOAT32 * __restrict__ p_out, const FLOAT32 * __restrict__ p_inp1, const FLOAT32 * __restrict__ p_inp2, const unsigned char *__restrict__ p_condition, WORD32 num_elm);
+
+extern "C" WORD32 xa_nn_elm_where_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,
+                      const unsigned char *__restrict__ p_condition,
+                      const WORD32 *const p_condition_shape
+                      );
 
 /* new functions in nnlib */
 extern "C" WORD32 xa_nn_elm_add_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,

diff --git a/backends/cadence/hifi/operators/CMakeLists.txt b/backends/cadence/hifi/operators/CMakeLists.txt
@@ -23,6 +23,7 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_add.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_embedding.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_full.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_where.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_view_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
@@ -37,6 +38,9 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_clone.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_div.cpp"
     "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_mul.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_div.cpp"
+    "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_rsqrt.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"
@@ -45,7 +49,9 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sub.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_to_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_where.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_empty.cpp")
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_empty.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/pattern/unary_ufunc_realhb_to_floath.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)