Added HiFi optimizations for add sub mul and div operators

backends/cadence/aot/functions_hifi.yaml

@@ -92,6 +92,11 @@
     - arg_meta: null
       kernel_name: torch::executor::sub_out
 
+- op: tanh.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::tanh_out
+
 - op: view_copy.out
   kernels:
     - arg_meta: null

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
 )
 
 target_include_directories(

backends/cadence/hifi/kernels/kernels.h

@@ -12,6 +12,46 @@
 #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 impl {
 namespace HiFi {
 namespace kernels {

backends/cadence/hifi/operators/CMakeLists.txt

@@ -20,32 +20,34 @@ endif()
 
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
-    "${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"
-    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_add.cpp"
+    "${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/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_realhb_to_floath.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/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)

backends/cadence/hifi/operators/op_add.cpp

@@ -0,0 +1,123 @@
+/*
+ * 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.
+ */
+
+#include <executorch/kernels/portable/cpu/scalar_utils.h>
+#include <executorch/kernels/portable/cpu/util/broadcast_util.h>
+#include <executorch/kernels/portable/cpu/util/functional_util.h>
+#include <executorch/runtime/kernel/kernel_includes.h>
+#include <executorch/runtime/platform/assert.h>
+#include "kernels.h"
+
+namespace torch {
+namespace executor {
+namespace native {
+
+#define NNLIB_MAX_DIM 4  /* Add fallback if broadcast and dim > 4 */
+
+Tensor& add_out(
+    RuntimeContext& ctx,
+    const Tensor& a,
+    const Tensor& b,
+    const Scalar& alpha,
+    Tensor& out) {
+  (void)ctx;
+
+  ScalarType a_type = a.scalar_type();
+  ScalarType b_type = b.scalar_type();
+  ScalarType common_type = promoteTypes(a_type, b_type);
+  ScalarType out_type = out.scalar_type();
+
+  ET_CHECK_MSG(a_type == ScalarType::Float, "Input tensor not a float.\n");
+  ET_CHECK_MSG(b_type == ScalarType::Float, "Input tensor not a float.\n");
+  ET_CHECK_MSG(out_type == ScalarType::Float, "Output tensor not a float.\n");
+
+  ET_CHECK(canCast(common_type, out_type));
+
+  using CTYPE_A = float;
+  using CTYPE_B = float;
+  using CTYPE_IN = float;
+  using CTYPE_OUT = float;
+  CTYPE_IN alpha_val;
+  ET_EXTRACT_SCALAR(alpha, alpha_val);
+
+  int a_dim = a.dim(), b_dim = b.dim(), out_dim = out.dim();
+  int fall_back = 0;
+  /*find broadcast*/
+  const int a_is_broadcasted = !out.sizes().equals(a.sizes());
+  const int b_is_broadcasted = !out.sizes().equals(b.sizes());
+  const int broadcast = (a_is_broadcasted || b_is_broadcasted);
+  int max_dim = a.dim() > b.dim() ? a.dim() : b.dim();
+  max_dim = out.dim() > max_dim ? out.dim() : max_dim;
+
+  if( (out_type != ScalarType::Float) || (alpha_val != 1.0))
+    fall_back = 1;
+
+  if( (a_dim == 0) || (b_dim == 0) )
+    fall_back = 1;
+
+  if((broadcast == 1) && (max_dim > NNLIB_MAX_DIM))
+    fall_back = 1;
+
+
+  if (!fall_back)
+  {
+      const float* const a_data = a.const_data_ptr<float>();
+      const float* const b_data = b.const_data_ptr<float>();
+      float* const out_data = out.mutable_data_ptr<float>();
+      if(broadcast == 1)
+      {
+         int out_shape[NNLIB_MAX_DIM];
+         int inp1_shape[NNLIB_MAX_DIM];
+         int inp2_shape[NNLIB_MAX_DIM];
+
+         for(int i = 0; i < NNLIB_MAX_DIM; i++)
+         {
+            out_shape[i] = 1;
+            inp1_shape[i] = 1;
+            inp2_shape[i] = 1;
+         }
+
+         int off_o = NNLIB_MAX_DIM - out.dim();
+         int off_a = NNLIB_MAX_DIM - a.dim();
+         int off_b = NNLIB_MAX_DIM - b.dim();
+
+         for(int i = 0; i < out.dim(); i++)
+             out_shape[i+off_o] = out.size(i);
+         for(int i = 0; i < a.dim(); i++)
+             inp1_shape[i+off_a] = a.size(i);
+         for(int i = 0; i < b.dim(); i++)
+             inp2_shape[i+off_b] = b.size(i);
+
+         xa_nn_elm_add_broadcast_4D_f32xf32_f32(out_data, out_shape, a_data, inp1_shape,
+                                                b_data, inp2_shape);
+      }                      
+      else
+        xa_nn_elm_add_f32xf32_f32(out_data, a_data, b_data, out.numel());
+
+  }
+  else
+  {
+      apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
+      [alpha_val](const CTYPE_A val_a, const CTYPE_B val_b) {
+        CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
+        CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
+        CTYPE_IN value = a_casted + alpha_val * b_casted;
+
+        return static_cast<CTYPE_OUT>(value);
+      },
+      a,
+      b,
+      out);
+  }
+
+  return out;
+}
+
+} // namespace native
+} // namespace executor
+} // namespace torch