Skip to content

Commit 71d78a1

Browse files
cad-audiocad-audio
authored
Merge pull request #7 from dijopaul/main
Adding add sub mul and div optimizations
2 parents 8ef6c79 + 4ed8bd6 commit 71d78a1

File tree

11 files changed

+2874
-13
lines changed

11 files changed

+2874
-13
lines changed

‎backends/cadence/hifi/kernels/CMakeLists.txt‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,10 @@ add_library(
99
cadence_kernels
1010
kernels.cpp
1111
${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/matmul_asym8uxasym8u_asym8u.cpp
12+
${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_add_f32_broadcast.c
13+
${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_f32_broadcast.c
14+
${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_mode_f32_broadcast.c
15+
${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_mul_f32_broadcast.c
1216
)
1317

1418
target_include_directories(

‎backends/cadence/hifi/kernels/kernels.h‎

Lines changed: 40 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -12,6 +12,46 @@
1212
#include <stddef.h>
1313
#include <xa_type_def.h>
1414

15+
/* For NNLIB APIs */
16+
#include "xa_nnlib_kernels_api.h"
17+
18+
/* Potential NNLIB function/APIs */
19+
extern "C" WORD32 xa_nn_elm_add_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
20+
const WORD32 *const p_out_shape,
21+
const FLOAT32 * __restrict__ p_inp1,
22+
const WORD32 *const p_inp1_shape,
23+
const FLOAT32 * __restrict__ p_inp2,
24+
const WORD32 *const p_inp2_shape);
25+
26+
extern "C" WORD32 xa_nn_elm_div_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
27+
const WORD32 *const p_out_shape,
28+
const FLOAT32 * __restrict__ p_inp1,
29+
const WORD32 *const p_inp1_shape,
30+
const FLOAT32 * __restrict__ p_inp2,
31+
const WORD32 *const p_inp2_shape);
32+
33+
extern "C" WORD32 xa_nn_elm_div_mode_f32xf32_f32(FLOAT32 * __restrict__ p_out,
34+
const FLOAT32 * __restrict__ p_inp1,
35+
const FLOAT32 * __restrict__ p_inp2,
36+
WORD32 num_elm,
37+
WORD32 mode);
38+
39+
extern "C" WORD32 xa_nn_elm_div_mode_broadcast_4D_f32xf32_f32(
40+
FLOAT32 * __restrict__ p_out,
41+
const WORD32 *const p_out_shape,
42+
const FLOAT32 * __restrict__ p_inp1,
43+
const WORD32 *const p_inp1_shape,
44+
const FLOAT32 * __restrict__ p_inp2,
45+
const WORD32 *const p_inp2_shape,
46+
WORD32 mode);
47+
48+
extern "C" WORD32 xa_nn_elm_mul_broadcast_4D_f32xf32_f32(FLOAT32 * __restrict__ p_out,
49+
const WORD32 *const p_out_shape,
50+
const FLOAT32 * __restrict__ p_inp1,
51+
const WORD32 *const p_inp1_shape,
52+
const FLOAT32 * __restrict__ p_inp2,
53+
const WORD32 *const p_inp2_shape);
54+
1555
namespace impl {
1656
namespace HiFi {
1757
namespace kernels {

‎backends/cadence/hifi/operators/CMakeLists.txt‎

Lines changed: 13 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -20,32 +20,32 @@ endif()
2020

2121
# ATen compliant ops that are needed to run this model.
2222
set(_aten_ops__srcs
23-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
24-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
25-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
26-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/index_util.cpp"
27-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/kernel_ops_util.cpp"
28-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"
29-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
30-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/repeat_util.cpp"
31-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_add.cpp"
23+
"${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_add.cpp"
24+
"${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_div.cpp"
25+
"${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_mul.cpp"
26+
"${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_sub.cpp"
3227
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_bmm.cpp"
3328
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_cat.cpp"
3429
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_clone.cpp"
35-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_div.cpp"
3630
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_embedding.cpp"
3731
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_full.cpp"
38-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_mul.cpp"
3932
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_permute_copy.cpp"
4033
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sigmoid.cpp"
4134
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_slice_copy.cpp"
4235
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_softmax.cpp"
4336
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_split_with_sizes_copy.cpp"
44-
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sub.cpp"
4537
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_to_copy.cpp"
4638
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_view_copy.cpp"
4739
"${EXECUTORCH_ROOT}/kernels/portable/cpu/op_where.cpp"
48-
)
40+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
41+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
42+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
43+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/index_util.cpp"
44+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/kernel_ops_util.cpp"
45+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"
46+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
47+
"${EXECUTORCH_ROOT}/kernels/portable/cpu/util/repeat_util.cpp"
48+
)
4949
add_library(aten_ops_cadence ${_aten_ops__srcs})
5050
target_link_libraries(aten_ops_cadence PUBLIC executorch)
5151
target_link_libraries(aten_ops_cadence PRIVATE cadence_kernels)

‎backends/cadence/hifi/operators/op_add.cpp‎

Lines changed: 123 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,123 @@
1+
/*
2+
* Copyright (c) Meta Platforms, Inc. and affiliates.
3+
* All rights reserved.
4+
*
5+
* This source code is licensed under the BSD-style license found in the
6+
* LICENSE file in the root directory of this source tree.
7+
*/
8+
9+
#include <executorch/kernels/portable/cpu/scalar_utils.h>
10+
#include <executorch/kernels/portable/cpu/util/broadcast_util.h>
11+
#include <executorch/kernels/portable/cpu/util/functional_util.h>
12+
#include <executorch/runtime/kernel/kernel_includes.h>
13+
#include <executorch/runtime/platform/assert.h>
14+
#include "kernels.h"
15+
16+
namespace torch {
17+
namespace executor {
18+
namespace native {
19+
20+
#define NNLIB_MAX_DIM 4 /* Add fallback if broadcast and dim > 4 */
21+
22+
Tensor& add_out(
23+
RuntimeContext& ctx,
24+
const Tensor& a,
25+
const Tensor& b,
26+
const Scalar& alpha,
27+
Tensor& out) {
28+
(void)ctx;
29+
30+
ScalarType a_type = a.scalar_type();
31+
ScalarType b_type = b.scalar_type();
32+
ScalarType common_type = promoteTypes(a_type, b_type);
33+
ScalarType out_type = out.scalar_type();
34+
35+
ET_CHECK_MSG(a_type == ScalarType::Float, "Input tensor not a float.\n");
36+
ET_CHECK_MSG(b_type == ScalarType::Float, "Input tensor not a float.\n");
37+
ET_CHECK_MSG(out_type == ScalarType::Float, "Output tensor not a float.\n");
38+
39+
ET_CHECK(canCast(common_type, out_type));
40+
41+
using CTYPE_A = float;
42+
using CTYPE_B = float;
43+
using CTYPE_IN = float;
44+
using CTYPE_OUT = float;
45+
CTYPE_IN alpha_val;
46+
ET_EXTRACT_SCALAR(alpha, alpha_val);
47+
48+
int a_dim = a.dim(), b_dim = b.dim(), out_dim = out.dim();
49+
int fall_back = 0;
50+
/*find broadcast*/
51+
const int a_is_broadcasted = !out.sizes().equals(a.sizes());
52+
const int b_is_broadcasted = !out.sizes().equals(b.sizes());
53+
const int broadcast = (a_is_broadcasted || b_is_broadcasted);
54+
int max_dim = a.dim() > b.dim() ? a.dim() : b.dim();
55+
max_dim = out.dim() > max_dim ? out.dim() : max_dim;
56+
57+
if( (out_type != ScalarType::Float) || (alpha_val != 1.0))
58+
fall_back = 1;
59+
60+
if( (a_dim == 0) || (b_dim == 0) )
61+
fall_back = 1;
62+
63+
if((broadcast == 1) && (max_dim > NNLIB_MAX_DIM))
64+
fall_back = 1;
65+
66+
67+
if (!fall_back)
68+
{
69+
const float* const a_data = a.const_data_ptr<float>();
70+
const float* const b_data = b.const_data_ptr<float>();
71+
float* const out_data = out.mutable_data_ptr<float>();
72+
if(broadcast == 1)
73+
{
74+
int out_shape[NNLIB_MAX_DIM];
75+
int inp1_shape[NNLIB_MAX_DIM];
76+
int inp2_shape[NNLIB_MAX_DIM];
77+
78+
for(int i = 0; i < NNLIB_MAX_DIM; i++)
79+
{
80+
out_shape[i] = 1;
81+
inp1_shape[i] = 1;
82+
inp2_shape[i] = 1;
83+
}
84+
85+
int off_o = NNLIB_MAX_DIM - out.dim();
86+
int off_a = NNLIB_MAX_DIM - a.dim();
87+
int off_b = NNLIB_MAX_DIM - b.dim();
88+
89+
for(int i = 0; i < out.dim(); i++)
90+
out_shape[i+off_o] = out.size(i);
91+
for(int i = 0; i < a.dim(); i++)
92+
inp1_shape[i+off_a] = a.size(i);
93+
for(int i = 0; i < b.dim(); i++)
94+
inp2_shape[i+off_b] = b.size(i);
95+
96+
xa_nn_elm_add_broadcast_4D_f32xf32_f32(out_data, out_shape, a_data, inp1_shape,
97+
b_data, inp2_shape);
98+
}
99+
else
100+
xa_nn_elm_add_f32xf32_f32(out_data, a_data, b_data, out.numel());
101+
102+
}
103+
else
104+
{
105+
apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
106+
[alpha_val](const CTYPE_A val_a, const CTYPE_B val_b) {
107+
CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
108+
CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
109+
CTYPE_IN value = a_casted + alpha_val * b_casted;
110+
111+
return static_cast<CTYPE_OUT>(value);
112+
},
113+
a,
114+
b,
115+
out);
116+
}
117+
118+
return out;
119+
}
120+
121+
} // namespace native
122+
} // namespace executor
123+
} // namespace torch

0 commit comments

Comments
 (0)