Skip to content

GitLab

Commit b9ba2813 authored by Bartlomiej Kocot's avatar Bartlomiej Kocot
Browse files

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel... 

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel into barkocot/fix-instance-fwd-xdl-log
parents a0049dde f2398f61
Hide whitespace changes
Inline Side-by-side
Showing with 455 additions and 95 deletions
client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
View file @ b9ba2813
...@@ -100,18 +100,18 @@ int main() ...@@ -100,18 +100,18 @@ int main()
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
View file @ b9ba2813
...@@ -71,18 +71,18 @@ int main() ...@@ -71,18 +71,18 @@ int main()
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ b9ba2813
...@@ -80,7 +80,7 @@ int main(int argc, char* argv[]) ...@@ -80,7 +80,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ b9ba2813
...@@ -78,18 +78,18 @@ int main(int argc, char* argv[]) ...@@ -78,18 +78,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
View file @ b9ba2813
...@@ -83,7 +83,7 @@ int main(int argc, char* argv[]) ...@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
View file @ b9ba2813
...@@ -79,18 +79,18 @@ int main(int argc, char* argv[]) ...@@ -79,18 +79,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ b9ba2813
...@@ -76,19 +76,19 @@ int main(int argc, char* argv[]) ...@@ -76,19 +76,19 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<RequantScaleLayout>, ck::Tuple<RequantScaleLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<RequantScaleDataType>, ck::Tuple<RequantScaleDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ b9ba2813
...@@ -72,18 +72,18 @@ int main(int argc, char* argv[]) ...@@ -72,18 +72,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ b9ba2813
...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f ...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_operations)
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ b9ba2813
...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial, ...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial,
typename OutDataType, typename OutDataType,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
typename OutLayout> typename OutLayout,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_bwd_weight( bool run_grouped_conv_bwd_weight(
const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths, const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
const std::array<ck::index_t, NumDimSpatial + 3>& input_strides, const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight( ...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough,
AComputeType,
BComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using AComputeType = ck::bf8_t;
using BComputeType = ck::f8_t;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
AComputeType,
BComputeType>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/16_convnd_fwd/common.hpp
View file @ b9ba2813
...@@ -11,7 +11,7 @@ ...@@ -11,7 +11,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
...@@ -174,19 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD ...@@ -174,19 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths); std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size; std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough, PassThrough,
ComputeType>; ComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu.inc
View file @ b9ba2813
...@@ -63,7 +63,7 @@ int execute_conv_fwd_scaleadd_scaleadd_relu() ...@@ -63,7 +63,7 @@ int execute_conv_fwd_scaleadd_scaleadd_relu()
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C}; K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo}; std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, 6> out_strides{ std::array<ck::index_t, 6> out_strides{
C, Do * Ho * Wo * G * C, 1, Ho * Wo * G * C, Wo * G * C, G * C}; K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1}; std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1}; std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
...@@ -76,7 +76,7 @@ int execute_conv_fwd_scaleadd_scaleadd_relu() ...@@ -76,7 +76,7 @@ int execute_conv_fwd_scaleadd_scaleadd_relu()
SimpleDeviceMem d0(sizeof(std::tuple_element_t<0, DDataTypes>) * N * Do * Ho * Wo * G * K); SimpleDeviceMem d0(sizeof(std::tuple_element_t<0, DDataTypes>) * N * Do * Ho * Wo * G * K);
SimpleDeviceMem d1(sizeof(std::tuple_element_t<1, DDataTypes>) * N * Do * Ho * Wo * G * K); SimpleDeviceMem d1(sizeof(std::tuple_element_t<1, DDataTypes>) * N * Do * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/24_grouped_convnd_fwd_scaleadd_ab/CMakeLists.txt 0 → 100644
View file @ b9ba2813
add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp32 grouped_conv_fwd_scaleadd_ab_fp32.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp32 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp16 grouped_conv_fwd_scaleadd_ab_fp16.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp16 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_ab_bf16 grouped_conv_fwd_scaleadd_ab_bf16.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_bf16 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_ab_int8 grouped_conv_fwd_scaleadd_ab_int8.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_operations)
client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab.inc 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 64; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Z = 3; // filter D
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Di = 14; // input D
static constexpr ck::index_t Hi = 14; // input H
static constexpr ck::index_t Wi = 14; // input W
static constexpr ck::index_t Do = 14; // output D
static constexpr ck::index_t Ho = 14; // output H
static constexpr ck::index_t Wo = 14; // output W
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int execute_conv_fwd_scaleadd_ab()
{
constexpr ck::index_t NumAs = 2;
constexpr ck::index_t NumBs = 2;
constexpr float scale = 1.5f;
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space.
// However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
// Hence, we need to adjust the order of strides.
std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, 6> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, 6> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, 6> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
using InputDtype = ck::tuple_element_t<0, InDataType>;
using InputBiasDtype = ck::tuple_element_t<1, InDataType>;
using WeightDtype = ck::tuple_element_t<0, WeiDataType>;
using WeightBiasDtype = ck::tuple_element_t<1, WeiDataType>;
SimpleDeviceMem in(sizeof(InputDtype) * N * Di * Hi * Wi * G * C);
SimpleDeviceMem in_bias(sizeof(InputBiasDtype) * N * Di * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeightDtype) * G * K * Z * Y * X * C);
SimpleDeviceMem wei_bias(sizeof(WeightBiasDtype) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Do * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
ScaleAdd,
ScaleAdd,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
std::array<const void*, NumAs> as = {in.GetDeviceBuffer(), in_bias.GetDeviceBuffer()};
std::array<const void*, NumBs> bs = {wei.GetDeviceBuffer(), wei_bias.GetDeviceBuffer()};
std::array<const void*, 0> ds{};
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(as,
bs,
ds,
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
ScaleAdd{scale},
ScaleAdd{scale},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Z * Y * X +
N * Di * Hi * Wi * G * C + G * K * Z * Y * X * C;
std::size_t num_bytes = 2 * sizeof(InDataType) * N * Di * Hi * Wi * G * C +
2 * sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(as,
bs,
ds,
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
ScaleAdd{scale},
ScaleAdd{scale},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
using InDataType = ck::Tuple<ck::bhalf_t, ck::bhalf_t>;
using WeiDataType = ck::Tuple<ck::bhalf_t, ck::bhalf_t>;
using OutDataType = ck::bhalf_t;
#include "grouped_conv_fwd_scaleadd_ab.inc"
int main() { return execute_conv_fwd_scaleadd_ab(); }
client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
using InDataType = ck::Tuple<ck::half_t, ck::half_t>;
using WeiDataType = ck::Tuple<ck::half_t, ck::half_t>;
using OutDataType = ck::half_t;
#include "grouped_conv_fwd_scaleadd_ab.inc"
int main() { return execute_conv_fwd_scaleadd_ab(); }
client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
using InDataType = ck::Tuple<float, float>;
using WeiDataType = ck::Tuple<float, float>;
using OutDataType = float;
#include "grouped_conv_fwd_scaleadd_ab.inc"
int main() { return execute_conv_fwd_scaleadd_ab(); }
client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp 0 → 100644
View file @ b9ba2813
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
using InDataType = ck::Tuple<int8_t, int8_t>;
using WeiDataType = ck::Tuple<int8_t, int8_t>;
using OutDataType = int8_t;
#include "grouped_conv_fwd_scaleadd_ab.inc"
int main() { return execute_conv_fwd_scaleadd_ab(); }
example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
View file @ b9ba2813
...@@ -3,7 +3,7 @@ ...@@ -3,7 +3,7 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
...@@ -27,7 +27,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio ...@@ -27,7 +27,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout> template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance = using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial, NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
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