Skip to content

GitLab

Commit f6922d3f authored by Chao Liu's avatar Chao Liu
Browse files

fix reference conv bwd data bug; update conv bwd data test

parent 8f722700
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 295 additions and 959 deletions
example/09_convnd_fwd/convnd_fwd_common.hpp
View file @ f6922d3f
......@@ -188,17 +188,26 @@ int run_conv_fwd_nhwc(bool do_verification,
if(do_verification)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::KZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWK,
ck::tensor_layout::convolution::NHWK,
ck::tensor_layout::convolution::NDHWK>>,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
......
example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
View file @ f6922d3f
......@@ -194,17 +194,28 @@ int run_conv_bwd_data_nhwc(bool do_verification,
if(do_verification)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
std::cout << "before ref" << std::endl;
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::KZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWK,
ck::tensor_layout::convolution::NHWK,
ck::tensor_layout::convolution::NDHWK>>,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
......@@ -219,8 +230,12 @@ int run_conv_bwd_data_nhwc(bool do_verification,
wei_element_op,
out_element_op);
std::cout << "before ref" << std::endl;
ref_invoker.Run(ref_argument);
std::cout << "after ref" << std::endl;
in_device_buf.FromDevice(in_n_hi_wi_c_device.mData.data());
return ck::utils::check_err(in_n_hi_wi_c_device.mData, in_n_hi_wi_c_host.mData) ? 0 : 1;
......
example/20_convnd_bwd_weight/convnd_bwd_weight_common.hpp
View file @ f6922d3f
......@@ -15,7 +15,7 @@
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
void print_helper_msg()
{
......@@ -197,17 +197,26 @@ int run_conv_bwd_weight_nhwc(bool do_verification,
if(do_verification)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvBwdWeight<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>{};
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<
2,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::KZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWK,
ck::tensor_layout::convolution::NHWK,
ck::tensor_layout::convolution::NDHWK>>,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>{};
auto ref_invoker = ref_conv.MakeInvoker();
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View file @ f6922d3f
......@@ -142,14 +142,14 @@ struct ReferenceConvBwdData : public device::BaseOperator
for(std::size_t x = 0; x < X; ++x)
{
auto w_tmp = ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[0]);
auto w_tmp = static_cast<ck::long_index_t>(wi) +
static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]);
if(w_tmp % arg.conv_strides_[0] == 0)
{
auto wo = ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
auto wo = static_cast<ck::long_index_t>(w_tmp) /
static_cast<ck::long_index_t>(arg.conv_strides_[0]);
if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
{
......@@ -209,27 +209,26 @@ struct ReferenceConvBwdData : public device::BaseOperator
for(std::size_t y = 0; y < Y; ++y)
{
auto h_tmp = ck::type_convert<ck::long_index_t>(hi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]);
auto h_tmp = static_cast<ck::long_index_t>(hi) +
static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]);
if(h_tmp % arg.conv_strides_[0] == 0)
{
auto ho = ck::type_convert<ck::long_index_t>(h_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
auto ho = static_cast<ck::long_index_t>(h_tmp) /
static_cast<ck::long_index_t>(arg.conv_strides_[0]);
if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
{
for(std::size_t x = 0; x < X; ++x)
{
auto w_tmp =
ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]) -
ck::type_convert<ck::long_index_t>(x *
arg.conv_dilations_[1]);
static_cast<ck::long_index_t>(wi) +
static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]);
if(w_tmp % arg.conv_strides_[1] == 0)
{
auto wo = ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[1]);
auto wo =
static_cast<ck::long_index_t>(w_tmp) /
static_cast<ck::long_index_t>(arg.conv_strides_[1]);
if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
{
for(std::size_t k = 0; k < K; ++k)
......@@ -296,44 +295,41 @@ struct ReferenceConvBwdData : public device::BaseOperator
for(std::size_t z = 0; z < Z; ++z)
{
auto d_tmp = ck::type_convert<ck::long_index_t>(di) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(z * arg.conv_dilations_[0]);
auto d_tmp = static_cast<ck::long_index_t>(di) +
static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]);
if(d_tmp % arg.conv_strides_[0] == 0)
{
auto do_ = ck::type_convert<ck::long_index_t>(d_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
auto do_ = static_cast<ck::long_index_t>(d_tmp) /
static_cast<ck::long_index_t>(arg.conv_strides_[0]);
if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
{
for(std::size_t y = 0; y < Y; ++y)
{
auto h_tmp =
ck::type_convert<ck::long_index_t>(hi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]) -
ck::type_convert<ck::long_index_t>(y *
arg.conv_dilations_[1]);
static_cast<ck::long_index_t>(hi) +
static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]);
if(h_tmp % arg.conv_strides_[1] == 0)
{
auto ho = ck::type_convert<ck::long_index_t>(h_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[1]);
auto ho =
static_cast<ck::long_index_t>(h_tmp) /
static_cast<ck::long_index_t>(arg.conv_strides_[1]);
if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
{
for(std::size_t x = 0; x < X; ++x)
{
auto w_tmp =
ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(
arg.in_left_pads_[2]) -
ck::type_convert<ck::long_index_t>(
x * arg.conv_dilations_[2]);
auto w_tmp = static_cast<ck::long_index_t>(wi) +
static_cast<ck::long_index_t>(
arg.in_left_pads_[2]) -
static_cast<ck::long_index_t>(
x * arg.conv_dilations_[2]);
if(w_tmp % arg.conv_strides_[2] == 0)
{
auto wo =
ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[2]);
auto wo = static_cast<ck::long_index_t>(w_tmp) /
static_cast<ck::long_index_t>(
arg.conv_strides_[2]);
if(wo >= 0 &&
ck::type_convert<std::size_t>(wo) < Wo)
{
......@@ -381,7 +377,7 @@ struct ReferenceConvBwdData : public device::BaseOperator
arg.in_element_op_(v_in, v_acc);
// FIXME hacky
arg.input_.mData[in_desc.GetOffsetFromMultiIndex(n, c, wi)] =
arg.input_.mData[in_desc.GetOffsetFromMultiIndex(n, c, di, hi, wi)] =
ck::type_convert<InDataType>(v_acc);
};
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp → library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
View file @ f6922d3f
......@@ -139,10 +139,9 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
{
for(std::size_t wo = 0; wo < out_desc.GetLengths()[2]; ++wo)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < in_desc.GetLengths()[2])
......@@ -195,17 +194,16 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
{
for(std::size_t ho = 0; ho < out_desc.GetLengths()[2]; ++ho)
{
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t wo = 0; wo < out_desc.GetLengths()[3]; ++wo)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) < in_desc.GetLengths()[2] &&
......@@ -261,24 +259,21 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
{
for(std::size_t do_ = 0; do_ < out_desc.GetLengths()[2]; ++do_)
{
auto di =
ck::type_convert<ck::long_index_t>(do_ * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(z * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto di = static_cast<ck::long_index_t>(do_ * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t ho = 0; ho < out_desc.GetLengths()[3]; ++ho)
{
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(std::size_t wo = 0; wo < out_desc.GetLengths()[4]; ++wo)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo *
arg.conv_strides_[2]) +
ck::type_convert<ck::long_index_t>(x *
arg.conv_dilations_[2]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[2]);
static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
if(di >= 0 &&
ck::type_convert<std::size_t>(di) <
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ f6922d3f
......@@ -157,10 +157,9 @@ struct ReferenceConvFwd : public device::BaseOperator
{
for(std::size_t x = 0; x < wei_desc.GetLengths()[2]; ++x)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < in_desc.GetLengths()[2])
......@@ -213,17 +212,16 @@ struct ReferenceConvFwd : public device::BaseOperator
{
for(std::size_t y = 0; y < wei_desc.GetLengths()[2]; ++y)
{
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t x = 0; x < wei_desc.GetLengths()[3]; ++x)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) < in_desc.GetLengths()[2] &&
......@@ -280,24 +278,21 @@ struct ReferenceConvFwd : public device::BaseOperator
{
for(std::size_t z = 0; z < wei_desc.GetLengths()[2]; ++z)
{
auto di =
ck::type_convert<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(z * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t y = 0; y < wei_desc.GetLengths()[3]; ++y)
{
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(std::size_t x = 0; x < wei_desc.GetLengths()[4]; ++x)
{
auto wi =
ck::type_convert<ck::long_index_t>(wo *
arg.conv_strides_[2]) +
ck::type_convert<ck::long_index_t>(x *
arg.conv_dilations_[2]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[2]);
static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
if(di >= 0 &&
ck::type_convert<std::size_t>(di) <
in_desc.GetLengths()[2] &&
......
profiler/include/profile_conv_bwd_weight_impl.hpp
View file @ f6922d3f
......@@ -20,7 +20,7 @@
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
namespace ck {
namespace profiler {
......
test/CMakeLists.txt
View file @ f6922d3f
......@@ -41,9 +41,8 @@ add_subdirectory(gemm_reduce)
add_subdirectory(batched_gemm)
add_subdirectory(batched_gemm_reduce)
add_subdirectory(grouped_gemm)
add_subdirectory(convnd_fwd)
add_subdirectory(reduce)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(convnd_fwd)
add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(block_to_ctile_map)
......
test/batched_gemm/CMakeLists.txt
View file @ f6922d3f
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_fp16 PRIVATE host_tensor)
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
test/batched_gemm_reduce/CMakeLists.txt
View file @ f6922d3f
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE host_tensor)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
test/conv2d_bwd_data/CMakeLists.txt deleted 100644 → 0
View file @ 8f722700
add_test_executable(test_conv2d_bwd_data conv2d_bwd_data.cpp)
target_link_libraries(test_conv2d_bwd_data PRIVATE host_tensor)
target_link_libraries(test_conv2d_bwd_data PRIVATE device_conv2d_bwd_data_instance)
test/conv2d_bwd_data/conv2d_bwd_data.cpp deleted 100644 → 0
View file @ 8f722700
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
template <typename T>
static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
float max_diff = 1e-6;
for(int i = 0; i < ref.mData.size(); ++i)
{
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
return false;
}
}
return true;
}
int main(int argc, char* argv[])
{
int data_type = 0;
int init_method = 0;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 1)
{
data_type = 1;
init_method = 1;
}
else if(argc == 3)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
}
else if(argc == 18)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
N = std::stoi(argv[3]);
K = std::stoi(argv[4]);
C = std::stoi(argv[5]);
Y = std::stoi(argv[6]);
X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]);
conv_stride_h = std::stoi(argv[10]);
conv_stride_w = std::stoi(argv[11]);
conv_dilation_h = std::stoi(argv[12]);
conv_dilation_w = std::stoi(argv[13]);
in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
}
else
{
printf("arg1: data type (0=fp32, 1=fp16, 2= bfp16, 3= int8_t )\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
auto Run = [&](auto input_type, auto wei_type, auto out_type, auto acc_type) {
using InDataType = decltype(input_type);
using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
using AccDataType = decltype(acc_type);
using ReferenceConvBwdInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
};
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi));
Tensor<InDataType> in_n_c_hi_wi_device_result(f_host_tensor_descriptor(N, C, Hi, Wi));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
// reset input to zero
in_n_c_hi_wi_device_result.GenerateTensorValue(GeneratorTensor_1<InDataType>{0});
in_device_buf.ToDevice(in_n_c_hi_wi_device_result.mData.data());
// get host result
{
auto ref_conv = ReferenceConvBwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdDataNoOpPtr = ck::tensor_operation::device::
DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
// profile device Conv instances
bool success = true;
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), 1);
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
if(!check_out(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result))
{
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
success = false;
}
else
{
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
}
}
else
{
std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
}
}
if(success)
{
std::cout << "test conv2d bwd : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd: Fail " << std::endl;
return -1;
}
};
if(data_type == 0)
{
return Run(F32(), F32(), F32(), F32());
}
else if(data_type == 1)
{
return Run(F16(), F16(), F16(), F32());
}
else if(data_type == 2)
{
return Run(BF16(), BF16(), BF16(), F32());
}
else if(data_type == 3)
{
return Run(INT8(), INT8(), INT8(), int());
}
else
{
return 1;
}
}
test/conv2d_bwd_weight/CMakeLists.txt deleted 100644 → 0
View file @ 8f722700
add_test_executable(test_conv2d_bwd_weight conv2d_bwd_weight.cpp)
target_link_libraries(test_conv2d_bwd_weight PRIVATE host_tensor device_conv2d_bwd_weight_instance conv_util)
test/conv2d_bwd_weight/conv2d_bwd_weight.cpp deleted 100644 → 0
View file @ 8f722700
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include "test/convnd_fwd/conv_util.hpp"
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
int test_self()
{
bool pass = true;
std::vector<ck::utils::conv::ConvParams> params;
params.push_back({2, 128, 256, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
params.push_back({2, 128, 256, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
params.push_back({2, 128, 256, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : params)
{
// f32
pass &= ck::profiler::profile_conv_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// fp16
pass &= ck::profiler::profile_conv_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
return pass;
}
int main(int argc, char* argv[])
{
int data_type = 1;
int init_method = 1;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
ck::index_t split_k = 1;
bool pass = true;
if(argc == 1)
{
pass = test_self();
}
else
{
if(argc == 3)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
}
else if(argc == 19)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
N = std::stoi(argv[3]);
K = std::stoi(argv[4]);
C = std::stoi(argv[5]);
Y = std::stoi(argv[6]);
X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]);
conv_stride_h = std::stoi(argv[10]);
conv_stride_w = std::stoi(argv[11]);
conv_dilation_h = std::stoi(argv[12]);
conv_dilation_w = std::stoi(argv[13]);
in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
split_k = std::stoi(argv[18]);
}
else
{
printf("arg1: data type (0=fp32, 1=fp16, 2= bfp16, 3= int8_t )\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
ck::utils::conv::ConvParams param{2,
N,
K,
C,
{Y, X},
{Hi, Wi},
{conv_stride_h, conv_stride_w},
{conv_dilation_h, conv_dilation_w},
{in_left_pad_h, in_left_pad_w},
{in_right_pad_h, in_right_pad_w}};
if(data_type == 0)
{
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
init_method,
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
split_k);
}
else if(data_type == 1)
{
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
init_method,
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
split_k);
}
else
{
std::cout << "Not support data type" << std::endl;
return 1;
}
}
if(pass)
{
std::cout << "test conv2d bwd weight : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd weight: Fail " << std::endl;
return -1;
}
}
test/conv_util/CMakeLists.txt
View file @ f6922d3f
add_gtest_executable(test_conv_util conv_util.cpp)
target_link_libraries(test_conv_util PRIVATE host_tensor conv_util)
target_link_libraries(test_conv_util PRIVATE utility)
test/conv_util/conv_util.cpp
View file @ f6922d3f
......@@ -10,7 +10,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
namespace {
......
test/convnd_bwd_data/CMakeLists.txt
View file @ f6922d3f
add_test_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
target_link_libraries(test_convnd_bwd_data PRIVATE host_tensor device_convnd_bwd_data_instance conv_util)
target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance)
test/convnd_bwd_data/convnd_bwd_data.cpp
View file @ f6922d3f
This diff is collapsed.
test/convnd_bwd_weight/CMakeLists.txt
View file @ f6922d3f
add_test_executable(test_convnd_bwd_weight convnd_bwd_weight.cpp)
target_link_libraries(test_convnd_bwd_weight PRIVATE host_tensor device_convnd_bwd_weight_instance conv_util)
target_link_libraries(test_convnd_bwd_weight PRIVATE utility device_convnd_bwd_weight_instance)
test/convnd_fwd/CMakeLists.txt
View file @ f6922d3f
add_custom_target(test_convnd_fwd)
add_gtest_executable(test_conv1d_fwd conv1d_fwd.cpp)
target_link_libraries(test_conv1d_fwd PRIVATE host_tensor device_conv1d_fwd_instance conv_util)
target_link_libraries(test_conv1d_fwd PRIVATE utility device_conv1d_fwd_instance)
add_dependencies(test_convnd_fwd test_conv1d_fwd)
add_gtest_executable(test_conv2d_fwd conv2d_fwd.cpp)
target_link_libraries(test_conv2d_fwd PRIVATE host_tensor device_conv2d_fwd_instance device_convnd_2d_fwd_instance conv_util)
target_link_libraries(test_conv2d_fwd PRIVATE utility device_conv2d_fwd_instance)
add_dependencies(test_convnd_fwd test_conv2d_fwd)
add_gtest_executable(test_conv3d_fwd conv3d_fwd.cpp)
target_link_libraries(test_conv3d_fwd PRIVATE host_tensor device_conv3d_fwd_instance conv_util)
target_link_libraries(test_conv3d_fwd PRIVATE utility device_conv3d_fwd_instance)
add_dependencies(test_convnd_fwd test_conv3d_fwd)
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