Skip to content

GitLab

Commit db775824 authored by Jehandad Khan's avatar Jehandad Khan
Browse files

Merge branch 'develop' into jd/dev_pkg

parents 74397984 7c0b1498
Hide whitespace changes
Inline Side-by-side
Showing with 839 additions and 847 deletions
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
View file @ db775824
......@@ -28,19 +28,19 @@ using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances = std::tuple<
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 128, 32, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 64, 32, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 64, 128, 32, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 64, 32, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 32, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 32, 128, 32, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 32, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 32, 64, 32, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
......
library/src/utility/CMakeLists.txt 0 → 100644
View file @ db775824
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/device
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/element
${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
)
set(CONV_FWD_UTIL_SOURCE
conv_fwd_util.cpp
)
add_library(conv_fwd_util SHARED ${CONV_FWD_UTIL_SOURCE})
target_link_libraries(conv_fwd_util PRIVATE host_tensor)
target_compile_features(conv_fwd_util PUBLIC)
set_target_properties(conv_fwd_util PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(conv_fwd_util SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
clang_tidy_check(conv_fwd_util)
library/src/utility/conv_fwd_util.cpp 0 → 100644
View file @ db775824
#include "conv_fwd_util.hpp"
namespace ck {
namespace utils {
namespace conv {
/**
* @brief Calculate number of FLOPs for Convolution
*
* @param[in] N Batch size.
* @param[in] C Number of input channels.
* @param[in] K Number of output channels.
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
* @param[in] output_spatial_lengths Convolution output spatial dimensions
* lengths.
*
* @return The number of flops.
*/
std::size_t get_flops(ck::index_t N,
ck::index_t C,
ck::index_t K,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths)
{
// 2 * N * K * <output spatial lengths product> * C * <filter spatial lengths product>
return static_cast<std::size_t>(2) * N * K *
std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>()) *
C *
std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>());
}
ConvParams::ConvParams()
: num_dim_spatial(2),
N(128),
K(256),
C(192),
filter_spatial_lengths(2, 3),
input_spatial_lengths(2, 71),
conv_filter_strides(2, 2),
conv_filter_dilations(2, 1),
input_left_pads(2, 1),
input_right_pads(2, 1)
{
}
ConvParams::ConvParams(ck::index_t n_dim,
ck::index_t n_batch,
ck::index_t n_out_channels,
ck::index_t n_in_channels,
const std::vector<ck::index_t>& filters_len,
const std::vector<ck::index_t>& input_len,
const std::vector<ck::index_t>& strides,
const std::vector<ck::index_t>& dilations,
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads)
: num_dim_spatial(n_dim),
N(n_batch),
K(n_out_channels),
C(n_in_channels),
filter_spatial_lengths(filters_len),
input_spatial_lengths(input_len),
conv_filter_strides(strides),
conv_filter_dilations(dilations),
input_left_pads(left_pads),
input_right_pads(right_pads)
{
if(filter_spatial_lengths.size() != num_dim_spatial ||
input_spatial_lengths.size() != num_dim_spatial ||
conv_filter_strides.size() != num_dim_spatial ||
conv_filter_dilations.size() != num_dim_spatial ||
input_left_pads.size() != num_dim_spatial || input_right_pads.size() != num_dim_spatial)
{
throw(std::runtime_error(
"ConvParams::GetOutputSpatialLengths: "
"parameter size is different from number of declared dimensions!"));
}
}
std::vector<ck::index_t> ConvParams::GetOutputSpatialLengths() const
{
if(filter_spatial_lengths.size() != num_dim_spatial ||
input_spatial_lengths.size() != num_dim_spatial ||
conv_filter_strides.size() != num_dim_spatial ||
conv_filter_dilations.size() != num_dim_spatial ||
input_left_pads.size() != num_dim_spatial || input_right_pads.size() != num_dim_spatial)
{
throw(std::runtime_error(
"ConvParams::GetOutputSpatialLengths: "
"parameter size is different from number of declared dimensions!"));
}
std::vector<ck::index_t> out_spatial_len(num_dim_spatial, 0);
for(ck::index_t i = 0; i < num_dim_spatial; ++i)
{
// XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::index_t idx_eff =
(filter_spatial_lengths[i] - 1) * conv_filter_dilations[i] + 1;
out_spatial_len[i] =
(input_spatial_lengths[i] + input_left_pads[i] + input_right_pads[i] - idx_eff) /
conv_filter_strides[i] +
1;
}
return out_spatial_len;
}
ConvParams parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
HostTensorDescriptor get_output_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWK{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWK{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWK{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KZYXC{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KYXC{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KXC{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
} // namespace conv
} // namespace utils
} // namespace ck
std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParams& p)
{
os << "ConvParams {"
<< "\nnum_dim_spatial: " << p.num_dim_spatial << "\nN: " << p.N << "\nK: " << p.K
<< "\nC: " << p.C << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths
<< "\ninput_spatial_lengths: " << p.input_spatial_lengths
<< "\nconv_filter_strides: " << p.conv_filter_strides
<< "\nconv_filter_dilations: " << p.conv_filter_dilations
<< "\ninput_left_pads: " << p.input_left_pads
<< "\ninput_right_pads: " << p.input_right_pads;
return os;
}
profiler/CMakeLists.txt
View file @ db775824
......@@ -29,10 +29,10 @@ set(PROFILER_SOURCE
src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp
src/profile_batched_gemm.cpp
src/profile_conv_fwd.cpp
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_convnd_fwd.cpp
src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
......@@ -43,13 +43,16 @@ set(PROFILER_SOURCE
add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE conv_fwd_util)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv1d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv3d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
......@@ -57,7 +60,6 @@ target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
set_target_properties(ckProfiler PROPERTIES EXCLUDE_FROM_ALL 1)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
profiler/include/profile_conv_fwd_impl.hpp deleted 100644 → 0
View file @ 74397984
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_fwd.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_fwd.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_fwd_instance {
using DeviceConvFwdNoOpPtr = DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv2d_fwd_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
void profile_conv_fwd_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_host_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_device_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.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_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.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_device_result.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvFwdNoOpPtr> 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::device_conv2d_fwd_instance::
add_device_conv2d_fwd_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::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_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::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
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,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_host : ", out_n_k_ho_wo_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_convnd_fwd.hpp 0 → 100644
View file @ db775824
#pragma once
namespace ck {
namespace profiler {
int profile_convnd_fwd(int argc, char* argv[]);
} // namespace profiler
} // namespace ck
profiler/include/profile_reduce_impl.hpp
View file @ db775824
......@@ -380,13 +380,9 @@ void profile_reduce_impl_impl(bool do_verification,
if(do_verification)
{
using HostInDataType = typename type_mapping<InDataType>::OutType;
using HostOutDataType = typename type_mapping<OutDataType>::OutType;
using HostAccDataType = typename type_mapping<AccDataType>::OutType;
ReductionHost<HostInDataType,
HostAccDataType,
HostOutDataType,
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
Rank,
NumReduceDim,
......@@ -394,11 +390,8 @@ void profile_reduce_impl_impl(bool do_verification,
NeedIndices>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha,
reinterpret_cast<const HostInDataType*>(in.mData.data()),
beta,
reinterpret_cast<HostOutDataType*>(out_ref.mData.data()),
out_indices_ref.mData.data());
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
};
const auto i_inLengths = to_int_vector(inLengths);
......
profiler/src/profile_conv_fwd.cpp deleted 100644 → 0
View file @ 74397984
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv_fwd_impl.hpp"
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int profile_conv_fwd(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_fwd: ForwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
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;
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_fwd_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_fwd_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>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::BF16_BF16_BF16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_fwd_impl<2,
uint16_t,
uint16_t,
uint16_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::INT8_INT8_INT8 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_fwd_impl<2,
int8_t,
int8_t,
int8_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
}
profiler/src/profile_convnd_bwd_data.cpp
View file @ db775824
......@@ -7,6 +7,8 @@
#include "profile_convnd_bwd_data_impl.hpp"
namespace {
enum struct ConvDataType
{
F32_F32_F32, // 0
......@@ -76,6 +78,8 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[],
return params;
}
} // namespace
int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
{
const int preParams = 10;
......
profiler/src/profile_convnd_fwd.cpp 0 → 100644
View file @ db775824
#include <cstdlib>
#include <iostream>
#include <memory>
#include <string>
#include <vector>
#include <half.hpp>
#include "conv_fwd_util.hpp"
#include "element_wise_operation.hpp"
#include "fill.hpp"
#include "profile_convnd_fwd.hpp"
#include "tensor_layout.hpp"
namespace {
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum struct ConvDataLayout
{
NCHW, // 0
NHWC, // 1
};
namespace ctl = ck::tensor_layout::convolution;
template <int NDim, ConvDataLayout DataLayout>
struct ConvolutionLayouts;
template <>
struct ConvolutionLayouts<1, ConvDataLayout::NHWC>
{
typedef ctl::NWC Input;
typedef ctl::KXC Weight;
typedef ctl::NWK Output;
};
template <>
struct ConvolutionLayouts<2, ConvDataLayout::NHWC>
{
typedef ctl::NHWC Input;
typedef ctl::KYXC Weight;
typedef ctl::NHWK Output;
};
template <>
struct ConvolutionLayouts<3, ConvDataLayout::NHWC>
{
typedef ctl::NDHWC Input;
typedef ctl::KZYXC Weight;
typedef ctl::NDHWK Output;
};
template <>
struct ConvolutionLayouts<1, ConvDataLayout::NCHW>
{
typedef ctl::NCW Input;
typedef ctl::KCX Weight;
typedef ctl::NKW Output;
};
template <>
struct ConvolutionLayouts<2, ConvDataLayout::NCHW>
{
typedef ctl::NCHW Input;
typedef ctl::KCYX Weight;
typedef ctl::NKHW Output;
};
template <>
struct ConvolutionLayouts<3, ConvDataLayout::NCHW>
{
typedef ctl::NCDHW Input;
typedef ctl::KCZYX Weight;
typedef ctl::NKDHW Output;
};
void print_use_msg()
{
std::cout << "arg1: tensor operation (conv_fwd: ForwardConvolution)\n"
<< "arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n"
<< "arg3: data layout (0: NCHW; 1: NHWC)\n"
<< "arg4: verification (0=no, 1=yes)\n"
<< "arg5: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: run kernel # of times (>1)\n"
<< "arg8: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
<< " <strides>, (ie Sy, Sx for 2D)\n"
<< " <dilations>, (ie Dy, Dx for 2D)\n"
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
}
ck::utils::conv::ConvParams parse_params(int num_dim_spatial, int argc, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 9;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(1);
}
int arg_idx = 9;
return ck::utils::conv::parse_conv_params(num_dim_spatial, arg_idx, argv);
}
template <int NDim,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename ConvLayouts>
void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
bool do_verification,
bool do_log,
int nrepeat,
int init_method,
ConvLayouts)
{
using namespace std::placeholders;
using namespace ck::utils;
std::unique_ptr<OpInstance<OutDataType, InDataType, WeiDataType>> conv_instance;
switch(init_method)
{
case 0:
conv_instance =
std::make_unique<conv::ConvFwdOpInstance<InDataType,
WeiDataType,
OutDataType,
typename ConvLayouts::Input,
typename ConvLayouts::Weight,
typename ConvLayouts::Output>>(params, false);
break;
case 1:
conv_instance = std::make_unique<
conv::ConvFwdOpInstance<InDataType,
WeiDataType,
OutDataType,
typename ConvLayouts::Input,
typename ConvLayouts::Weight,
typename ConvLayouts::Output,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::utils::FillUniform<int>,
ck::utils::FillUniform<int>>>(
params, true, ck::utils::FillUniform<int>{}, ck::utils::FillUniform<int>{});
break;
case 2:
conv_instance = std::make_unique<
conv::ConvFwdOpInstance<InDataType,
WeiDataType,
OutDataType,
typename ConvLayouts::Input,
typename ConvLayouts::Weight,
typename ConvLayouts::Output,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::utils::FillUniform<InDataType>,
ck::utils::FillUniform<WeiDataType>>>(
params,
true,
ck::utils::FillUniform<InDataType>{},
ck::utils::FillUniform<WeiDataType>{});
break;
default: throw std::runtime_error("Unsupported init method!");
}
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<NDim, InDataType, WeiDataType, OutDataType>,
params,
_1,
_2,
_3);
OpInstanceRunEngine<InDataType, WeiDataType, OutDataType> run_engine(*conv_instance,
reference_conv_fwd_fun);
auto best_conf = run_engine.Profile(
conv::ConvolutionFwdInstances<InDataType, WeiDataType, OutDataType>::template Get<NDim>(),
nrepeat,
do_verification,
do_log);
std::cout << "Best configuration parameters:"
<< "\nname: " << best_conf.best_op_name << "\navg_time: " << best_conf.best_avg_time
<< "\ntflops: " << best_conf.best_tflops << "\nGB/s: " << best_conf.best_gb_per_sec
<< std::endl;
}
template <int NDim>
void profile_convnd_instances(ConvDataType data_type,
ConvDataLayout data_layout,
const ck::utils::conv::ConvParams& params,
bool do_verification,
bool do_log,
int nrepeat,
int init_method)
{
switch(data_layout)
{
case ConvDataLayout::NHWC: {
switch(data_type)
{
case ConvDataType::F32_F32_F32:
profile_convnd_instances_impl<NDim, float, float, float>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
case ConvDataType::F16_F16_F16:
profile_convnd_instances_impl<NDim, ck::half_t, ck::half_t, ck::half_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
case ConvDataType::BF16_BF16_BF16:
profile_convnd_instances_impl<NDim, ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
case ConvDataType::INT8_INT8_INT8:
profile_convnd_instances_impl<NDim, int8_t, int8_t, int8_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
}
break;
}
case ConvDataLayout::NCHW: {
switch(data_type)
{
case ConvDataType::F32_F32_F32:
profile_convnd_instances_impl<NDim, float, float, float>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
case ConvDataType::F16_F16_F16:
profile_convnd_instances_impl<NDim, ck::half_t, ck::half_t, ck::half_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
case ConvDataType::BF16_BF16_BF16:
profile_convnd_instances_impl<NDim, ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
case ConvDataType::INT8_INT8_INT8:
profile_convnd_instances_impl<NDim, int8_t, int8_t, int8_t>(
params,
do_verification,
do_log,
nrepeat,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
}
break;
}
}
}
} // namespace
int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
{
using namespace ck::utils::conv;
ConvDataType data_type{ConvDataType::F32_F32_F32};
ConvDataLayout data_layout{ConvDataLayout::NHWC};
bool do_verification{true};
int init_method{2};
bool do_log{false};
int nrepeat{100};
int num_dim_spatial{2};
ConvParams params;
if(argc >= 4)
{
data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
data_layout = static_cast<ConvDataLayout>(std::stoi(argv[3]));
}
if(argc >= 9)
{
do_verification = std::stoi(argv[4]);
init_method = std::stoi(argv[5]);
do_log = std::stoi(argv[6]);
nrepeat = std::stoi(argv[7]);
num_dim_spatial = std::stoi(argv[8]);
}
if(argc >= 10)
{
params = parse_params(num_dim_spatial, argc, argv);
}
// TODO Print nice message what is being profiled.
switch(num_dim_spatial)
{
case 1:
profile_convnd_instances<1>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
break;
case 2:
profile_convnd_instances<2>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
break;
case 3:
profile_convnd_instances<3>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
break;
default:
throw std::runtime_error("profile_conv_fwd: unsupported num_dim_spatial value: " +
std::to_string(num_dim_spatial));
}
return 1;
}
profiler/src/profiler.cpp
View file @ db775824
......@@ -4,6 +4,8 @@
#include <cstdlib>
#include <cstring>
#include "profile_convnd_fwd.hpp"
int profile_gemm(int, char*[]);
int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
......@@ -11,7 +13,6 @@ int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
......@@ -56,7 +57,7 @@ int main(int argc, char* argv[])
}
else if(strcmp(argv[1], "conv_fwd") == 0)
{
return profile_conv_fwd(argc, argv);
return ck::profiler::profile_convnd_fwd(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0)
{
......
script/clang-format-overwrite.sh
View file @ db775824
find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
test/conv2d_bwd_weight/CMakeLists.txt
View file @ db775824
......@@ -4,5 +4,4 @@ include_directories(BEFORE
)
add_test_executable(test_conv2d_bwd_weight conv2d_bwd_weight.cpp)
target_link_libraries(test_conv2d_bwd_weight PRIVATE host_tensor)
target_link_libraries(test_conv2d_bwd_weight PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(test_conv2d_bwd_weight PRIVATE host_tensor device_conv2d_bwd_weight_instance conv_fwd_util)
test/conv_util/CMakeLists.txt
View file @ db775824
add_test_executable(test_conv_util conv_util.cpp)
target_link_libraries(test_conv_util PRIVATE host_tensor)
target_link_libraries(test_conv_util PRIVATE host_tensor conv_fwd_util)
test/convnd_bwd_data/CMakeLists.txt
View file @ db775824
......@@ -4,5 +4,4 @@ include_directories(BEFORE
)
add_test_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
target_link_libraries(test_convnd_bwd_data PRIVATE host_tensor)
target_link_libraries(test_convnd_bwd_data PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(test_convnd_bwd_data PRIVATE host_tensor device_convnd_bwd_data_instance conv_fwd_util)
test/convnd_fwd/CMakeLists.txt
View file @ db775824
add_custom_target(test_convnd_fwd)
add_test_executable(test_conv1d_fwd conv1d_fwd.cpp)
target_link_libraries(test_conv1d_fwd PRIVATE host_tensor)
target_link_libraries(test_conv1d_fwd PRIVATE device_conv1d_fwd_instance)
target_link_libraries(test_conv1d_fwd PRIVATE host_tensor device_conv1d_fwd_instance conv_fwd_util)
target_link_libraries(test_conv1d_fwd PRIVATE )
add_dependencies(test_convnd_fwd test_conv1d_fwd)
add_test_executable(test_conv2d_fwd conv2d_fwd.cpp)
target_link_libraries(test_conv2d_fwd PRIVATE host_tensor)
target_link_libraries(test_conv2d_fwd PRIVATE device_conv2d_fwd_instance)
target_link_libraries(test_conv2d_fwd PRIVATE host_tensor device_conv2d_fwd_instance conv_fwd_util)
add_dependencies(test_convnd_fwd test_conv2d_fwd)
add_test_executable(test_conv3d_fwd conv3d_fwd.cpp)
target_link_libraries(test_conv3d_fwd PRIVATE host_tensor)
target_link_libraries(test_conv3d_fwd PRIVATE device_conv3d_fwd_instance)
target_link_libraries(test_conv3d_fwd PRIVATE host_tensor device_conv3d_fwd_instance conv_fwd_util)
add_dependencies(test_convnd_fwd test_conv3d_fwd)
test/convnd_fwd/conv1d_fwd.cpp
View file @ db775824
......@@ -7,37 +7,15 @@
#include "element_wise_operation.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "host_tensor.hpp"
#include "tensor_layout.hpp"
#include "check_err.hpp"
// Forward declarations for conv instances.
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv1d_fwd_instance {
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv1d_fwd_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace {
bool test_conv1D_nwc()
{
bool res{true};
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 2;
......@@ -50,30 +28,26 @@ bool test_conv1D_nwc()
params.input_left_pads = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1};
auto host_tensors =
ck::utils::conv::get_host_tensors<float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& host_output = std::get<2>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
ck::utils::conv::run_reference_convolution_forward<1>(params, input, weights, host_output);
test::conv::RunConv<1>(params, input, weights, device_output);
res = res &&
ck::utils::check_err(
device_output.mData, host_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return res;
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<1>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(
params);
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<1, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_conv1d_nwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_ptrs)
bool test_conv1d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
......@@ -83,52 +57,36 @@ bool test_conv1d_nwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_ptr
params.input_left_pads = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1};
auto host_tensors =
ck::utils::conv::get_host_tensors<T,
T,
T,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
const Tensor<T>& input = std::get<0>(host_tensors);
const Tensor<T>& weights = std::get<1>(host_tensors);
Tensor<T>& host_output = std::get<2>(host_tensors);
Tensor<T>& device_output = std::get<3>(host_tensors);
ck::utils::conv::run_reference_convolution_forward<1>(params, input, weights, host_output);
return ck::utils::conv::run_convolution_forward_instances<1>(
params, conv_ptrs, input, weights, device_output, host_output);
conv::ConvFwdOpInstance<T, T, T, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(params);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
bool test_conv1d_nwc_bf16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv1d_fwd_instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
return test_conv1d_nwc_instances<ck::bhalf_t>(conv_ptrs);
return test_conv1d_nwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<1>());
}
bool test_conv1d_nwc_f16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv1d_fwd_instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
return test_conv1d_nwc_instances<ck::half_t>(conv_ptrs);
return test_conv1d_nwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<1>());
}
bool test_conv1d_nwc_f32_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv1d_fwd_instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
return test_conv1d_nwc_instances<float>(conv_ptrs);
return test_conv1d_nwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<1>());
}
bool test_conv1d_nwc_int8_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv1d_fwd_instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
return test_conv1d_nwc_instances<int8_t>(conv_ptrs);
return test_conv1d_nwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<1>());
}
} // anonymous namespace
......@@ -149,7 +107,7 @@ int main()
std::cout << "\ntest_conv1d_nwc_f32_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv1d_nwc_int8_instances();
std::cout << "\ntes_tconv1_dnw_cint_8instances ..... " << (res ? "SUCCESS" : "FAILURE")
std::cout << "\ntest_conv1d_nwc_int8_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
......
test/convnd_fwd/conv2d_fwd.cpp
View file @ db775824
#include <half.hpp>
#include <iostream>
#include <stdexcept>
#include <tuple>
#include <vector>
......@@ -8,38 +7,14 @@
#include "element_wise_operation.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "host_tensor.hpp"
#include "tensor_layout.hpp"
#include "check_err.hpp"
// Forward declarations for conv instances.
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_fwd_instance {
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv2d_fwd_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace {
bool test_conv2d_nhwc()
{
bool res{true};
using namespace std::placeholders;
using namespace ck::utils;
ck::utils::conv::ConvParams params;
params.N = 2;
params.K = 16;
......@@ -47,25 +22,25 @@ bool test_conv2d_nhwc()
params.input_spatial_lengths = std::vector<ck::index_t>{16, 16};
params.conv_filter_strides = std::vector<ck::index_t>{1, 1};
auto host_tensors = ck::utils::conv::get_host_tensors(params);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& host_output = std::get<2>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
ck::utils::conv::run_reference_convolution_forward<2>(params, input, weights, host_output);
test::conv::RunConv<2>(params, input, weights, device_output);
res = res &&
ck::utils::check_err(
device_output.mData, host_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<2>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float> conv_instance(params);
return res;
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<2, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_conv2d_nhwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_ptrs)
bool test_conv2d_nhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
ck::utils::conv::ConvParams params;
using namespace std::placeholders;
using namespace ck::utils;
conv::ConvParams params;
params.num_dim_spatial = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3};
params.input_spatial_lengths = std::vector<ck::index_t>{71, 71};
......@@ -74,55 +49,36 @@ bool test_conv2d_nhwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_pt
params.input_left_pads = std::vector<ck::index_t>{1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1};
auto host_tensors =
ck::utils::conv::get_host_tensors<T,
T,
T,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(params);
const Tensor<T>& input = std::get<0>(host_tensors);
const Tensor<T>& weights = std::get<1>(host_tensors);
Tensor<T>& host_output = std::get<2>(host_tensors);
Tensor<T>& device_output = std::get<3>(host_tensors);
ck::utils::conv::run_reference_convolution_forward<2>(params, input, weights, host_output);
return ck::utils::conv::run_convolution_forward_instances<2>(
params, conv_ptrs, input, weights, device_output, host_output);
conv::ConvFwdOpInstance<T, T, T> conv_instance(params);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<2, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
bool test_conv2d_nhwc_bf16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
return test_conv2d_nhwc_instances<ck::bhalf_t>(conv_ptrs);
return test_conv2d_nhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<2>());
}
bool test_conv2d_nhwc_f16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
return test_conv2d_nhwc_instances<ck::half_t>(conv_ptrs);
return test_conv2d_nhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<2>());
}
bool test_conv2d_nhwc_f32_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
return test_conv2d_nhwc_instances<float>(conv_ptrs);
return test_conv2d_nhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<2>());
}
bool test_conv2d_nhwc_int8_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
return test_conv2d_nhwc_instances<int8_t>(conv_ptrs);
return test_conv2d_nhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<2>());
}
} // anonymous namespace
......
test/convnd_fwd/conv3d_fwd.cpp
View file @ db775824
......@@ -8,37 +8,16 @@
#include "element_wise_operation.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "host_tensor.hpp"
#include "tensor_layout.hpp"
#include "check_err.hpp"
// Forward declarations for conv instances.
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv3d_fwd_instance {
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv3d_fwd_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace {
bool test_conv3d_ndhwc()
{
bool res{true};
ck::utils::conv::ConvParams params;
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 2;
params.K = 16;
......@@ -50,31 +29,26 @@ bool test_conv3d_ndhwc()
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
auto host_tensors =
ck::utils::conv::get_host_tensors<float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(params);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& host_output = std::get<2>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
ck::utils::conv::run_reference_convolution_forward<3>(params, input, weights, host_output);
test::conv::RunConv<3>(params, input, weights, device_output);
res = res &&
ck::utils::check_err(
device_output.mData, host_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return res;
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float, ctl::NDHWC, ctl::KZYXC, ctl::NDHWK> conv_instance(
params);
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<3, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
return run_engine.Test(conv_ptrs);
}
bool test_conv3d_ndhwc_2gb_input()
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
// >2GB Input
ck::utils::conv::ConvParams params;
conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 2;
params.K = 16;
......@@ -86,39 +60,35 @@ bool test_conv3d_ndhwc_2gb_input()
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
auto host_tensors =
ck::utils::conv::get_host_tensors<float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(params, false);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
try
{
test::conv::RunConv<3>(params, input, weights, device_output);
}
catch(const std::runtime_error& err)
{
std::string err_msg{"Error! device_conv with the specified compilation parameters does "
"not support this Conv problem"};
if(err.what() != err_msg)
{
return false;
}
return true;
}
std::cout << "Error: Failure checking oversized tensor!" << std::endl;
return false;
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
bool test_conv3d_ndhwc_2gb_filters()
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
// >2GB Filters
ck::utils::conv::ConvParams params;
conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 2;
params.K = 16;
......@@ -130,39 +100,35 @@ bool test_conv3d_ndhwc_2gb_filters()
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
auto host_tensors =
ck::utils::conv::get_host_tensors<float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(params, false);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
try
{
test::conv::RunConv<3>(params, input, weights, device_output);
}
catch(const std::runtime_error& err)
{
std::string err_msg{"Error! device_conv with the specified compilation parameters does "
"not support this Conv problem"};
if(err.what() != err_msg)
{
return false;
}
return true;
}
std::cout << "Error: Failure checking oversized tensor!" << std::endl;
return false;
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
bool test_conv3d_ndhwc_2gb_output()
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
// >2GB Output
ck::utils::conv::ConvParams params;
conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 2;
params.K = 16;
......@@ -174,39 +140,35 @@ bool test_conv3d_ndhwc_2gb_output()
params.input_left_pads = std::vector<ck::index_t>{2, 2, 2};
params.input_right_pads = std::vector<ck::index_t>{2, 2, 2};
auto host_tensors =
ck::utils::conv::get_host_tensors<float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(params, false);
const Tensor<float>& input = std::get<0>(host_tensors);
const Tensor<float>& weights = std::get<1>(host_tensors);
Tensor<float>& device_output = std::get<3>(host_tensors);
try
{
test::conv::RunConv<3>(params, input, weights, device_output);
}
catch(const std::runtime_error& err)
{
std::string err_msg{"Error! device_conv with the specified compilation parameters does "
"not support this Conv problem"};
if(err.what() != err_msg)
{
return false;
}
return true;
}
std::cout << "Error: Failure checking oversized tensor!" << std::endl;
return false;
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
template <typename T>
bool test_conv3d_ndhwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_ptrs)
bool test_conv3d_ndhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
ck::utils::conv::ConvParams params;
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvParams params;
params.N = 64;
params.num_dim_spatial = 3;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 2};
......@@ -216,53 +178,36 @@ bool test_conv3d_ndhwc_instances(const std::vector<DeviceConvFwdNoOpPtr>& conv_p
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
auto host_tensors =
ck::utils::conv::get_host_tensors<T,
T,
T,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(params);
const Tensor<T>& input = std::get<0>(host_tensors);
const Tensor<T>& weights = std::get<1>(host_tensors);
Tensor<T>& host_output = std::get<2>(host_tensors);
Tensor<T>& device_output = std::get<3>(host_tensors);
conv::ConvFwdOpInstance<T, T, T, ctl::NDHWC, ctl::KZYXC, ctl::NDHWK> conv_instance(params);
ck::utils::conv::run_reference_convolution_forward<3>(params, input, weights, host_output);
return ck::utils::conv::run_convolution_forward_instances<3>(
params, conv_ptrs, input, weights, device_output, host_output);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
bool test_conv3d_ndhwc_bf16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv3d_fwd_instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
return test_conv3d_ndhwc_instances<ck::bhalf_t>(conv_ptrs);
return test_conv3d_ndhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<3>());
}
bool test_conv3d_ndhwc_f16_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv3d_fwd_instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
return test_conv3d_ndhwc_instances<ck::half_t>(conv_ptrs);
return test_conv3d_ndhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<3>());
}
bool test_conv3d_ndhwc_f32_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv3d_fwd_instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
return test_conv3d_ndhwc_instances<float>(conv_ptrs);
return test_conv3d_ndhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<3>());
}
bool test_conv3d_ndhwc_int8_instances()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
ck::tensor_operation::device::device_conv3d_fwd_instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
return test_conv3d_ndhwc_instances<int8_t>(conv_ptrs);
return test_conv3d_ndhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<3>());
}
} // anonymous namespace
......@@ -293,7 +238,7 @@ int main()
std::cout << "\ntest_conv3d_ndhwc_f32_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_int8_instances();
std::cout << "\ntest_conv3d_ndhw_cint_8instances ..... " << (res ? "SUCCESS" : "FAILURE")
std::cout << "\ntest_conv3d_ndhwc_int8_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
......
test/convnd_fwd/conv_util.hpp
View file @ db775824
......@@ -10,7 +10,8 @@
#include "host_tensor.hpp"
#include "sequence.hpp"
namespace {
namespace test {
namespace conv {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -19,6 +20,9 @@ using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<InElementOp, WeiElementOp, OutElementOp>;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
......@@ -62,26 +66,14 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::
1>; // CThreadTransferDstScalarPerVector
// clang-format on
} // namespace
namespace test {
namespace conv {
template <ck::index_t NDim,
typename InDataType = float,
typename WeiDataType = float,
typename OutDataType = float>
void RunConv(const ck::utils::conv::ConvParams& params,
const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weights,
Tensor<OutDataType>& output)
void get_test_convolution_fwd_instance(std::vector<DeviceConvFwdNoOpPtr>& instances)
{
ck::utils::conv::run_convolution_forward<NDim,
InDataType,
WeiDataType,
OutDataType,
DeviceConvNDFwdInstance>(
params, input, weights, output);
using ConvInstanceT = DeviceConvNDFwdInstance<NDim, InDataType, WeiDataType, OutDataType>;
instances.emplace_back(std::make_unique<ConvInstanceT>());
}
} // namespace conv
......
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