ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-05 20:55:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "test gk bias"

Browse Source 
This reverts commit 8e42e29d7b.
This commit is contained in:
Bartlomiej Kocot
2025-06-16 08:28:06 +00:00
parent 8e42e29d7b
commit e73c0550ce
4 changed files with 5 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
profiler/include/profiler/profile_grouped_conv_fwd_bias_clamp_impl.hpp
View File

@@ -25,23 +25,6 @@
namespace ck {
namespace profiler {
template <ck::index_t NDimSpatial>
auto get_bias_desc(ck::index_t G, ck::index_t K)
{
if constexpr(NDimSpatial == 1)
{
return HostTensorDescriptor({G, 1, K, 1}, {K, 0, 1, 0});
}
else if constexpr(NDimSpatial == 2)
{
return HostTensorDescriptor({G, 1, K, 1, 1}, {K, 0, 1, 0, 0});
}
else
{
return HostTensorDescriptor({G, 1, K, 1, 1, 1}, {K, 0, 1, 0, 0, 0});
}
}
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
@@ -51,8 +34,7 @@ template <ck::index_t NDimSpatial,
typename OutDataType,
typename AComputeType = InDataType,
typename BComputeType = AComputeType,
typename IndexType = ck::index_t,
bool BiasGK = false>
typename IndexType = ck::index_t>
bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
int init_method,
bool do_log,
@@ -79,16 +61,12 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
const index_t G = conv_param.G_;
const index_t K = conv_param.K_;
std::array<IndexType, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<IndexType, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<IndexType, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<IndexType, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<IndexType, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<IndexType, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<IndexType, NDimSpatial + 3> d_g_n_k_wos_strides{};
std::array<IndexType, NDimSpatial> conv_filter_strides{};
std::array<IndexType, NDimSpatial> conv_filter_dilations{};
std::array<IndexType, NDimSpatial> input_left_pads{};
@@ -102,7 +80,6 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
@@ -112,8 +89,7 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
const auto bias_desc = BiasGK ? get_bias_desc<NDimSpatial>(G, K) : out_g_n_k_wos_desc;
Tensor<OutDataType> bias(bias_desc);
Tensor<OutDataType> bias(out_g_n_k_wos_desc);
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weight: " << weight.mDesc << std::endl;
@@ -137,11 +113,7 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
const std::size_t bias_dev_buf_size =
BiasGK ? sizeof(OutDataType) * G * K
: sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize();
DeviceMem bias_device_buf(bias_dev_buf_size);
DeviceMem bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(input.mData.data());
wei_device_buf.ToDevice(weight.mData.data());
@@ -272,16 +244,6 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
if constexpr(BiasGK)
{
constexpr ck::index_t spatial_offset = 3;
d_g_n_k_wos_strides[1] = 0;
for(int i = 0; i < NDimSpatial; i++)
{
d_g_n_k_wos_strides[i + spatial_offset] = 0;
}
}
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
@@ -293,7 +255,7 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
{e_g_n_k_wos_lengths},
{d_g_n_k_wos_strides},
{e_g_n_k_wos_strides},
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,

3
test/grouped_convnd_fwd_activation/CMakeLists.txt
View File

@@ -2,9 +2,6 @@ if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_grouped_convnd_fwd_bias_clamp test_grouped_convnd_fwd_bias_clamp.cpp)
target_link_libraries(test_grouped_convnd_fwd_bias_clamp PRIVATE utility device_grouped_conv2d_fwd_bias_clamp_instance device_grouped_conv3d_fwd_bias_clamp_instance)
add_gtest_executable(test_grouped_convnd_fwd_gk_bias_clamp test_grouped_convnd_fwd_gk_bias_clamp.cpp)
target_link_libraries(test_grouped_convnd_fwd_gk_bias_clamp PRIVATE utility device_grouped_conv2d_fwd_bias_clamp_instance device_grouped_conv3d_fwd_bias_clamp_instance)
add_gtest_executable(test_grouped_convnd_fwd_clamp test_grouped_convnd_fwd_clamp.cpp)
target_link_libraries(test_grouped_convnd_fwd_clamp PRIVATE utility device_grouped_conv2d_fwd_clamp_instance device_grouped_conv3d_fwd_clamp_instance)
endif()

3
test/grouped_convnd_fwd_activation/test_grouped_convnd_fwd_bias_clamp.cpp
View File

@@ -41,8 +41,7 @@ class TestGroupedConvndFwd : public ::testing::Test
DataType,
DataType,
DataType,
IndexType,
false /*BiasGK*/>(
IndexType>(
true, // do_verification
1, // init_method: integer value
false, // do_log

93
test/grouped_convnd_fwd_activation/test_grouped_convnd_fwd_gk_bias_clamp.cpp
View File

@@ -1,93 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_fwd_bias_clamp_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using AddClamp = ck::tensor_operation::element_wise::AddClamp;
template <typename Tuple>
class TestGroupedConvndFwd : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using InLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using OutLayout = std::tuple_element_t<3, Tuple>;
using IndexType = ck::index_t;
std::vector<ck::utils::conv::ConvParam> conv_params;
template <ck::index_t NDimSpatial>
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_fwd_bias_clamp_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
DataType,
DataType,
DataType,
DataType,
DataType,
IndexType,
true /*BiasGK*/>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param);
}
EXPECT_TRUE(pass);
}
};
using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>>;
using KernelTypes3d = ::testing::Types<std::tuple<ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK>>;
template <typename Tuple>
class TestGroupedConvndFwd2d : public TestGroupedConvndFwd<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndFwd3d : public TestGroupedConvndFwd<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndFwd2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndFwd3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndFwd2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 32, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 32, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndFwd3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->template Run<3>();
}