Skip to content

GitLab

Unverified Commit 55b9438c authored by Charlie Lin's avatar Charlie Lin Committed by GitHub
Browse files

Merge branch 'develop' into refactor_dynamic_compute

parents fbe13c96 83784c52
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2072 additions and 4067 deletions
src/CMakeLists.txt
View file @ 55b9438c
......@@ -81,7 +81,6 @@ add_library(migraphx
replace_allocate.cpp
simplify_qdq.cpp
sqlite.cpp
rewrite_batchnorm.cpp
rewrite_gelu.cpp
rewrite_pooling.cpp
rewrite_quantization.cpp
......@@ -115,7 +114,6 @@ register_migraphx_ops(
as_shape
atanh
atan
batch_norm_inference
broadcast
capture
ceil
......
src/driver/alexnet.cpp
View file @ 55b9438c
......@@ -25,6 +25,7 @@
#include <migraphx/make_op.hpp>
#include <migraphx/program.hpp>
#include <migraphx/generate.hpp>
#include <migraphx/json.hpp>
#include "models.hpp"
namespace migraphx {
namespace driver {
......@@ -39,153 +40,161 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1}}, 1)));
auto x_main_module_2 = mmain->add_literal(migraphx::abs(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1}}, 2)));
auto x_0 = mmain->add_parameter(
"0", migraphx::shape{migraphx::shape::float_type, {batch, 3, 224, 224}});
auto x_data_0 = mmain->add_parameter(
"data_0", migraphx::shape{migraphx::shape::float_type, {batch, 3, 224, 224}});
auto x_main_module_4 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1000}}, 3));
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1000, 4096}}, 3));
auto x_main_module_5 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1000, 4096}}, 4));
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {1000}}, 4));
auto x_main_module_6 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096}}, 5));
auto x_main_module_7 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096, 4096}}, 6));
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096, 4096}}, 5));
auto x_main_module_7 = mmain->add_literal(migraphx::abs(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096}}, 6)));
auto x_main_module_8 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096}}, 7));
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096, 9216}}, 7));
auto x_main_module_9 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096, 9216}}, 8));
auto x_main_module_10 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {256}}, 9));
auto x_main_module_11 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {256, 256, 3, 3}}, 10));
auto x_main_module_12 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {256}}, 11));
auto x_main_module_13 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {256, 384, 3, 3}}, 12));
auto x_main_module_14 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {384}}, 13));
auto x_main_module_15 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {384, 192, 3, 3}}, 14));
auto x_main_module_16 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {192}}, 15));
auto x_main_module_17 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {192, 64, 5, 5}}, 16));
auto x_main_module_18 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {64}}, 17));
auto x_main_module_19 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {64, 3, 11, 11}}, 18));
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {4096}}, 8));
auto x_main_module_10 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {256, 192, 3, 3}}, 9));
auto x_main_module_11 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {256}}, 10));
auto x_main_module_12 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {384, 192, 3, 3}}, 11));
auto x_main_module_13 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {384}}, 12));
auto x_main_module_14 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {384, 256, 3, 3}}, 13));
auto x_main_module_15 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {384}}, 14));
auto x_main_module_16 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {256, 48, 5, 5}}, 15));
auto x_main_module_17 = mmain->add_literal(migraphx::abs(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {256}}, 16)));
auto x_main_module_18 = mmain->add_literal(migraphx::generate_literal(
migraphx::shape{migraphx::shape::float_type, {96, 3, 11, 11}}, 17));
auto x_main_module_19 = mmain->add_literal(
migraphx::generate_literal(migraphx::shape{migraphx::shape::float_type, {96}}, 18));
auto x_main_module_20 = mmain->add_instruction(
migraphx::make_json_op("convolution",
"{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[4,"
"{dilation:[1,1],group:1,padding:[0,0,0,0],padding_mode:0,stride:[4,"
"4],use_dynamic_same_auto_pad:0}"),
x_0,
x_main_module_19);
x_data_0,
x_main_module_18);
auto x_main_module_21 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,64,55,55]}"), x_main_module_18);
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,96,54,54]}"), x_main_module_19);
auto x_main_module_22 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_20, x_main_module_21);
auto x_main_module_23 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_22);
auto x_main_module_24 = mmain->add_instruction(
migraphx::make_json_op("lrn", "{alpha:9.999999747378752e-05,beta:0.75,bias:1.0,size:5}"),
x_main_module_23);
auto x_main_module_25 = mmain->add_instruction(
migraphx::make_json_op(
"pooling",
"{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,0,0],stride:[2,2]}"),
x_main_module_23);
auto x_main_module_25 = mmain->add_instruction(
x_main_module_24);
auto x_main_module_26 = mmain->add_instruction(
migraphx::make_json_op("convolution",
"{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[1,"
"{dilation:[1,1],group:2,padding:[2,2,2,2],padding_mode:0,stride:[1,"
"1],use_dynamic_same_auto_pad:0}"),
x_main_module_24,
x_main_module_17);
auto x_main_module_26 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,192,27,27]}"), x_main_module_16);
auto x_main_module_27 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_25, x_main_module_26);
auto x_main_module_28 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_27);
auto x_main_module_29 = mmain->add_instruction(
x_main_module_25,
x_main_module_16);
auto x_main_module_27 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,256,26,26]}"), x_main_module_17);
auto x_main_module_28 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_26, x_main_module_27);
auto x_main_module_29 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_28);
auto x_main_module_30 = mmain->add_instruction(
migraphx::make_json_op("lrn", "{alpha:9.999999747378752e-05,beta:0.75,bias:1.0,size:5}"),
x_main_module_29);
auto x_main_module_31 = mmain->add_instruction(
migraphx::make_json_op(
"pooling",
"{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,0,0],stride:[2,2]}"),
x_main_module_28);
auto x_main_module_30 = mmain->add_instruction(
x_main_module_30);
auto x_main_module_32 = mmain->add_instruction(
migraphx::make_json_op("convolution",
"{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
"1],use_dynamic_same_auto_pad:0}"),
x_main_module_29,
x_main_module_15);
auto x_main_module_31 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,384,13,13]}"), x_main_module_14);
auto x_main_module_32 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_30, x_main_module_31);
auto x_main_module_33 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_32);
auto x_main_module_34 = mmain->add_instruction(
x_main_module_31,
x_main_module_14);
auto x_main_module_33 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,384,12,12]}"), x_main_module_15);
auto x_main_module_34 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_32, x_main_module_33);
auto x_main_module_35 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_34);
auto x_main_module_36 = mmain->add_instruction(
migraphx::make_json_op("convolution",
"{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
"{dilation:[1,1],group:2,padding:[1,1,1,1],padding_mode:0,stride:[1,"
"1],use_dynamic_same_auto_pad:0}"),
x_main_module_33,
x_main_module_13);
auto x_main_module_35 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,256,13,13]}"), x_main_module_12);
auto x_main_module_36 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_34, x_main_module_35);
auto x_main_module_37 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_36);
auto x_main_module_38 = mmain->add_instruction(
x_main_module_35,
x_main_module_12);
auto x_main_module_37 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,384,12,12]}"), x_main_module_13);
auto x_main_module_38 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_36, x_main_module_37);
auto x_main_module_39 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_38);
auto x_main_module_40 = mmain->add_instruction(
migraphx::make_json_op("convolution",
"{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
"{dilation:[1,1],group:2,padding:[1,1,1,1],padding_mode:0,stride:[1,"
"1],use_dynamic_same_auto_pad:0}"),
x_main_module_37,
x_main_module_11);
auto x_main_module_39 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,256,13,13]}"), x_main_module_10);
auto x_main_module_40 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_38, x_main_module_39);
auto x_main_module_41 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_40);
auto x_main_module_42 = mmain->add_instruction(
x_main_module_39,
x_main_module_10);
auto x_main_module_41 = mmain->add_instruction(
migraphx::make_json_op("broadcast", "{axis:1,out_lens:[1,256,12,12]}"), x_main_module_11);
auto x_main_module_42 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_40, x_main_module_41);
auto x_main_module_43 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_42);
auto x_main_module_44 = mmain->add_instruction(
migraphx::make_json_op(
"pooling",
"{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,0,0],stride:[2,2]}"),
x_main_module_41);
auto x_main_module_43 =
mmain->add_instruction(migraphx::make_json_op("flatten", "{axis:1}"), x_main_module_42);
auto x_main_module_44 = mmain->add_instruction(migraphx::make_op("identity"), x_main_module_43);
"{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,1,1],stride:[2,2]}"),
x_main_module_43);
auto x_main_module_45 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_9);
auto x_main_module_46 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_44, x_main_module_45);
auto x_main_module_47 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_8);
migraphx::make_json_op("reshape", "{dims:[1,9216]}"), x_main_module_44);
auto x_main_module_46 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_8);
auto x_main_module_47 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_45, x_main_module_46);
auto x_main_module_48 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_9);
auto x_main_module_49 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_2);
auto x_main_module_49 =
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_47, x_main_module_48);
auto x_main_module_50 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_46, x_main_module_49);
auto x_main_module_51 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_50);
auto x_main_module_52 = mmain->add_instruction(migraphx::make_op("identity"), x_main_module_51);
auto x_main_module_53 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_7);
auto x_main_module_54 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_52, x_main_module_53);
auto x_main_module_55 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_6);
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_48, x_main_module_49);
auto x_main_module_51 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_47, x_main_module_50);
auto x_main_module_52 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_51);
auto x_main_module_53 = mmain->add_instruction(migraphx::make_op("identity"), x_main_module_52);
auto x_main_module_54 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_6);
auto x_main_module_55 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_53, x_main_module_54);
auto x_main_module_56 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_7);
auto x_main_module_57 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,4096]}"), x_main_module_1);
auto x_main_module_57 =
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_55, x_main_module_56);
auto x_main_module_58 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_54, x_main_module_57);
auto x_main_module_59 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_58);
auto x_main_module_60 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_5);
auto x_main_module_61 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_59, x_main_module_60);
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_56, x_main_module_57);
auto x_main_module_59 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_55, x_main_module_58);
auto x_main_module_60 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_59);
auto x_main_module_61 = mmain->add_instruction(migraphx::make_op("identity"), x_main_module_60);
auto x_main_module_62 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,1000]}"), x_main_module_4);
auto x_main_module_63 = mmain->add_instruction(
migraphx::make_json_op("transpose", "{permutation:[1,0]}"), x_main_module_4);
auto x_main_module_63 =
mmain->add_instruction(migraphx::make_op("dot"), x_main_module_61, x_main_module_62);
auto x_main_module_64 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,1000]}"), x_main_module_5);
auto x_main_module_65 = mmain->add_instruction(
migraphx::make_json_op("multibroadcast", "{out_lens:[1,1000]}"), x_main_module_0);
auto x_main_module_64 =
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_62, x_main_module_63);
auto x_main_module_65 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_61, x_main_module_64);
mmain->add_return({x_main_module_65});
auto x_main_module_66 =
mmain->add_instruction(migraphx::make_op("mul"), x_main_module_64, x_main_module_65);
auto x_main_module_67 =
mmain->add_instruction(migraphx::make_op("add"), x_main_module_63, x_main_module_66);
auto x_main_module_68 =
mmain->add_instruction(migraphx::make_json_op("softmax", "{axis:1}"), x_main_module_67);
mmain->add_return({x_main_module_68});
return p;
}
......
src/driver/inceptionv3.cpp
View file @ 55b9438c
This source diff could not be displayed because it is too large. You can view the blob instead.
src/driver/main.cpp
View file @ 55b9438c
......@@ -44,7 +44,6 @@
#include <migraphx/propagate_constant.hpp>
#include <migraphx/quantization.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/simplify_algebra.hpp>
#include <migraphx/simplify_reshapes.hpp>
#include <migraphx/register_target.hpp>
......@@ -221,7 +220,6 @@ struct loader
{
migraphx::run_passes(*p.get_main_module(),
{
migraphx::rewrite_batchnorm{},
migraphx::eliminate_identity{},
migraphx::dead_code_elimination{},
migraphx::simplify_algebra{},
......
src/driver/resnet50.cpp
View file @ 55b9438c
This diff is collapsed.
src/include/migraphx/op/batch_norm_inference.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_OPERATORS_BATCH_NORM_HPP
#define MIGRAPHX_GUARD_OPERATORS_BATCH_NORM_HPP
#include <migraphx/check_shapes.hpp>
#include <migraphx/config.hpp>
#include <cmath>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
struct batch_norm_inference
{
float epsilon = 1.0e-6f;
float momentum = 0.9f;
std::string name() const { return "batch_norm_inference"; }
enum bn_infer_mode_t
{
per_activation,
spatial,
};
bn_infer_mode_t bn_mode = spatial;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(
f(self.epsilon, "epsilon"), f(self.momentum, "momentum"), f(self.bn_mode, "bn_mode"));
}
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this}.has(5);
check_shapes{inputs.data(), inputs.data() + 1, *this}.same_ndims();
check_shapes{inputs.data() + 1, inputs.data() + inputs.size(), *this}.same_shape();
return inputs.front();
}
};
} // namespace op
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/operators.hpp
View file @ 55b9438c
......@@ -35,7 +35,6 @@
#include <migraphx/op/as_shape.hpp>
#include <migraphx/op/atan.hpp>
#include <migraphx/op/atanh.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/binary.hpp>
#include <migraphx/op/broadcast.hpp>
#include <migraphx/op/capture.hpp>
......
src/include/migraphx/rewrite_batchnorm.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
#define MIGRAPHX_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
#include <string>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/config.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
struct module;
/**
* Rewrite batchnorm to a multiply and add.
*/
struct rewrite_batchnorm
{
std::string name() const { return "rewrite_batchnorm"; }
void apply(module& m) const;
};
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/onnx/parse_batchnorm.cpp
View file @ 55b9438c
......@@ -54,18 +54,19 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
MIGRAPHX_THROW("PARSE_BATCHNORM: argument scale, bias, mean, or var rank != 1");
}
if(x_lens.size() == 1)
auto x_rank = x_lens.size();
if(x_rank == 1 or x_rank == 2)
{
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto n0 = info.add_broadcastable_binary_op("sub", args[0], args[3]);
auto d0 = info.add_broadcastable_binary_op("add", args[4], eps);
auto d1 = info.add_broadcastable_binary_op("pow", d0, rt);
auto div0 = info.add_broadcastable_binary_op("div", n0, d1);
auto r0 = info.add_broadcastable_binary_op("mul", div0, args[1]);
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto numer = info.add_broadcastable_binary_op("sub", args[0], args[3]);
auto var_eps = info.add_broadcastable_binary_op("add", args[4], eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, args[1]);
return info.add_broadcastable_binary_op("add", r0, args[2]);
}
else if(x_lens.size() > 2)
else if(x_rank > 2)
{
// unsqueeze tensors of shape (C) to broadcast correctly
std::vector<int64_t> unsqueeze_axes(x_lens.size() - 2);
......@@ -89,7 +90,7 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
}
else
{
// num dims either 0 or 2
// rank == 0
MIGRAPHX_THROW("PARSE_BATCHNORM: rank " + std::to_string(x_lens.size()) +
" input tensor, unhandled data format");
}
......
src/rewrite_batchnorm.cpp deleted 100644 → 0
View file @ fbe13c96
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/program.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/broadcast.hpp>
#include <migraphx/op/add.hpp>
#include <migraphx/op/mul.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/dfor.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
void rewrite_batchnorm::apply(module& m) const
{
for(auto ins : iterator_for(m))
{
if(ins->name() != "batch_norm_inference")
continue;
// Get scale, bias, mean, variance from inputs
auto gamma = ins->inputs()[1]->eval();
auto bias = ins->inputs()[2]->eval();
auto mean = ins->inputs()[3]->eval();
auto variance = ins->inputs()[4]->eval();
if(any_of({gamma, bias, mean, variance}, [](auto arg) { return arg.empty(); }))
continue;
std::vector<std::size_t> lens = ins->inputs()[1]->get_shape().lens();
shape s{ins->get_shape().type(), lens};
// Get epsilon
auto bn_op = any_cast<op::batch_norm_inference>(ins->get_operator());
auto epsilon = bn_op.epsilon;
argument a{s};
argument b{s};
visit_all(gamma, bias, mean, variance, a, b)(
[&](auto gamma2, auto bias2, auto mean2, auto variance2, auto a2, auto b2) {
dfor(a.get_shape().elements())(
[&](std::size_t c) { a2[c] = gamma2[c] / std::sqrt(variance2[c] + epsilon); });
dfor(b.get_shape().elements())([&](std::size_t c) {
b2[c] = bias2[c] - (gamma2[c] * mean2[c] / std::sqrt(variance2[c] + epsilon));
});
});
auto broadcast = op::broadcast{1, ins->get_shape().lens()};
auto a_ins = m.add_literal({a.get_shape(), a.data()});
auto a_broadcast = m.insert_instruction(ins, broadcast, a_ins);
auto mul = m.insert_instruction(ins, make_op("mul"), ins->inputs().front(), a_broadcast);
auto b_ins = m.add_literal({b.get_shape(), b.data()});
auto b_broadcast = m.insert_instruction(ins, broadcast, b_ins);
auto add = m.insert_instruction(ins, make_op("add"), mul, b_broadcast);
m.replace_instruction(ins, add);
}
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/cpu/lowering.cpp
View file @ 55b9438c
......@@ -26,7 +26,6 @@
#include <migraphx/instruction.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
......
src/targets/cpu/target.cpp
View file @ 55b9438c
......@@ -37,7 +37,6 @@
#include <migraphx/propagate_constant.hpp>
#include <migraphx/register_target.hpp>
#include <migraphx/replace_allocate.hpp>
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/rewrite_pooling.hpp>
#include <migraphx/rewrite_quantization.hpp>
#include <migraphx/rewrite_rnn.hpp>
......@@ -78,8 +77,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
eliminate_identity{},
eliminate_pad{},
dead_code_elimination{},
rewrite_batchnorm{},
dead_code_elimination{},
rewrite_rnn{},
dead_code_elimination{},
eliminate_common_subexpression{},
......
src/targets/gpu/CMakeLists.txt 100755 → 100644
View file @ 55b9438c
......@@ -78,7 +78,6 @@ add_library(migraphx_gpu
allocation_model.cpp
argmax.cpp
argmin.cpp
batch_norm_inference.cpp
code_object_op.cpp
compile_ops.cpp
compile_gen.cpp
......@@ -144,7 +143,6 @@ register_migraphx_gpu_ops(hip_
)
register_migraphx_gpu_ops(miopen_
abs
batch_norm_inference
contiguous
convolution
deconvolution
......
src/targets/gpu/batch_norm_inference.cpp deleted 100644 → 0
View file @ fbe13c96
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <migraphx/gpu/batch_norm_inference.hpp>
#include <migraphx/gpu/context.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape miopen_batch_norm_inference::compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(6);
check_shapes{inputs.data(), inputs.data() + 1, *this}.same_ndims().max_ndims(5);
return op.compute_shape({inputs.at(0), inputs.at(1), inputs.at(2), inputs.at(3), inputs.at(4)});
}
inline shape reshape_to_2d(const shape& input)
{
auto dims = input.lens();
if(dims.size() >= 4)
return input;
std::vector<size_t> new_dims(dims.begin(), dims.end());
std::size_t num = 4 - dims.size();
new_dims.insert(new_dims.end(), num, 1);
return {input.type(), new_dims};
}
argument miopen_batch_norm_inference::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
{
shape x_shape = args[0].get_shape();
shape y_shape = output_shape;
shape bn_shape = args[3].get_shape();
auto x_desc = make_tensor(reshape_to_2d(x_shape));
auto y_desc = make_tensor(reshape_to_2d(y_shape));
auto bn_desc = make_tensor(reshape_to_2d(bn_shape));
float alpha = 1.0;
float beta = 0.0f;
miopenBatchNormalizationForwardInference(ctx.get_stream().get_miopen(),
miopenBatchNormMode_t(op.bn_mode),
&alpha,
&beta,
x_desc.get(),
args[0].implicit(),
y_desc.get(),
args[5].implicit(),
bn_desc.get(),
args[1].implicit(),
args[2].implicit(),
args[3].implicit(),
args[4].implicit(),
op.epsilon);
return args[5];
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/include/migraphx/gpu/batch_norm_inference.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_BATCHNORM_HPP
#define MIGRAPHX_GUARD_RTGLIB_BATCHNORM_HPP
#include <migraphx/argument.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/reflect.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_batch_norm_inference
{
op::batch_norm_inference op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "gpu::batch_norm_inference"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/lowering.cpp
View file @ 55b9438c
......@@ -37,7 +37,6 @@
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/quant_dot.hpp>
#include <migraphx/gpu/batch_norm_inference.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/convolution.hpp>
#include <migraphx/gpu/deconvolution.hpp>
......@@ -113,7 +112,6 @@ struct miopen_apply
add_extend_op("scatter_none");
add_extend_op("topk");
add_batch_norm_inference_op();
add_convolution_op();
add_deconvolution_op();
add_gemm_op<op::dot>("dot");
......@@ -334,43 +332,6 @@ struct miopen_apply
});
}
void add_batch_norm_inference_op()
{
apply_map.emplace("batch_norm_inference", [=](instruction_ref ins) {
auto&& op = any_cast<op::batch_norm_inference>(ins->get_operator());
auto output = insert_allocation(ins, ins->get_shape());
shape old_shape = ins->inputs().at(1)->get_shape();
auto input = ins->inputs()[0];
auto input_lens = input->get_shape().lens();
std::vector<int64_t> rsp_lens(input_lens.size(), 1);
// for per_activation case, also need to reshape input
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
std::copy(input_lens.begin() + 1, input_lens.end(), rsp_lens.begin() + 1);
}
else
{
rsp_lens[1] = static_cast<int64_t>(old_shape.elements());
}
auto reshape_op = op::reshape{rsp_lens};
std::vector<instruction_ref> reshapes;
std::transform(ins->inputs().begin() + 1,
ins->inputs().end(),
std::back_inserter(reshapes),
[&](auto i) { return mod->insert_instruction(ins, reshape_op, i); });
return mod->replace_instruction(ins,
miopen_batch_norm_inference{op},
input,
reshapes[0],
reshapes[1],
reshapes[2],
reshapes[3],
output);
});
}
// use 0 - input to represent neg
void add_neg_op()
{
......
src/targets/gpu/target.cpp
View file @ 55b9438c
......@@ -41,7 +41,6 @@
#include <migraphx/propagate_constant.hpp>
#include <migraphx/register_target.hpp>
#include <migraphx/replace_allocate.hpp>
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/rewrite_gelu.hpp>
#include <migraphx/rewrite_pooling.hpp>
#include <migraphx/rewrite_quantization.hpp>
......@@ -110,8 +109,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
insert_pad{},
dead_code_elimination{},
rewrite_batchnorm{},
dead_code_elimination{},
rewrite_rnn{},
dead_code_elimination{},
inline_module{},
......
src/targets/ref/lowering.cpp
View file @ 55b9438c
......@@ -26,7 +26,6 @@
#include <migraphx/instruction.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
......@@ -73,84 +72,6 @@ typename std::conditional_t<std::is_integral<T>{}, std::make_signed<T>, std::ena
return x;
}
//
// ref implemenataion of batch norm for inference
//
// inputs are:
// args[0] -> input data buffer
// args[1] -> mini batch mean
// args[2] -> mini batch variance
// args[3] -> gamma
// args[4] -> bias
//
// The equation to compute batch norm for inference is:
//
// output[i] = bias + gamma * (input[i] + mean) / sqrt(variance + epsilon)
//
// the input data format should be nchw
//
struct ref_batch_norm_inference
{
op::batch_norm_inference op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "ref::batch_norm_inference"; }
shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument output{output_shape};
double epsilon = op.epsilon;
auto input = args[0];
auto arg_gamma = args[1];
auto arg_bias = args[2];
auto mini_batch_mean = args[3];
auto mini_batch_variance = args[4];
if(op.bn_mode == op::batch_norm_inference::spatial)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
auto c = idx[1];
assert((variance[c] + epsilon) > 0);
result[i] =
gamma[c] * (buffer[i] - mean[c]) / std::sqrt(variance[c] + epsilon) +
bias[c];
});
});
}
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
idx[0] = 0;
auto index = output_shape.index(idx);
assert((variance[index] + epsilon) > 0);
result[i] = gamma[index] * (buffer[i] - mean[index]) /
std::sqrt(variance[index] + epsilon) +
bias[index];
});
});
}
return output;
}
};
MIGRAPHX_REGISTER_OP(ref_batch_norm_inference)
struct ref_lrn
{
op::lrn op;
......@@ -643,8 +564,6 @@ struct ref_apply
void init()
{
apply_map["batch_norm_inference"] =
extend_op<ref_batch_norm_inference, op::batch_norm_inference>();
apply_map["convolution"] = extend_op<ref_convolution<op::convolution>, op::convolution>();
apply_map["dot"] = extend_op<ref_gemm, op::dot>();
apply_map["quant_dot"] = extend_op<ref_quant_gemm, op::quant_dot>();
......
src/tf/parse_batchnorm.cpp
View file @ 55b9438c
......@@ -23,6 +23,7 @@
*/
#include <migraphx/tf/op_parser.hpp>
#include <migraphx/tf/tf_parser.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/make_op.hpp>
......@@ -38,16 +39,37 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
instruction_ref parse(const op_desc& /*opd*/,
const tf_parser& /*parser*/,
tf_parser::node_info info,
const std::vector<instruction_ref>& args) const
std::vector<instruction_ref> args) const
{
float epsilon = 1e-5f;
float momentum = 0.9f;
// different default epsilon than from ONNX
float epsilon = 1e-4f;
if(contains(info.attributes, "epsilon"))
{
epsilon = info.attributes.at("epsilon").f();
}
auto op = make_op("batch_norm_inference", {{"epsilon", epsilon}, {"momentum", momentum}});
return info.add_instruction(op, args);
auto x_lens = args[0]->get_shape().lens();
auto x_type = args[0]->get_shape().type();
// unsqueeze tensors of shape (C) to broadcast correctly
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto scale_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[1]);
auto bias_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[2]);
auto mean_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[3]);
auto var_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[4]);
auto numer = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
};
......
test/api/test_custom_op_gpu.cpp
View file @ 55b9438c
......@@ -55,7 +55,8 @@ struct half_copy_host final : migraphx::experimental_custom_op_base
hipMemcpyHostToHost,
ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
MIGRAPHX_HIP_ASSERT(hipMemset(output_buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(output_buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[1];
}
......@@ -97,7 +98,8 @@ struct half_copy_device final : migraphx::experimental_custom_op_base
hipMemcpyDeviceToDevice,
ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
MIGRAPHX_HIP_ASSERT(hipMemset(output_buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(output_buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[1];
}
......@@ -124,7 +126,7 @@ struct half_copy_device_same_buffer final : migraphx::experimental_custom_op_bas
virtual bool runs_on_offload_target() const override { return true; }
virtual migraphx::argument
compute(migraphx::context, migraphx::shape, migraphx::arguments inputs) const override
compute(migraphx::context ctx, migraphx::shape, migraphx::arguments inputs) const override
{
// This custom op simply sets first half size_bytes of the input 0, and rest of the half
// bytes are copied. for this custom_op, it does its computation on the "device". Therefore,
......@@ -133,7 +135,8 @@ struct half_copy_device_same_buffer final : migraphx::experimental_custom_op_bas
auto input_bytes = inputs[0].get_shape().bytes();
auto copy_bytes = input_bytes / 2;
MIGRAPHX_HIP_ASSERT(hipSetDevice(0));
MIGRAPHX_HIP_ASSERT(hipMemset(buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[0];
}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment