Skip to content

GitLab

Commit 7ffb0921 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge branch 'develop' into aosewski/ck_tile_universal_gemm_p1

parents 4cf45f1b 0023f01a
Hide whitespace changes
Inline Side-by-side
Showing with 1040 additions and 113 deletions
example/ck_tile/04_img2col/image_to_column.cpp 0 → 100644
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstring>
#include "ck_tile/host.hpp"
#include "image_to_column.hpp"
// Host API implementation
template <>
float image_to_column(const image_to_column_traits& traits,
const image_to_column_args<2>& args,
const ck_tile::stream_config& stream_conf)
{
if(traits.data_type.compare("fp16") == 0)
{
constexpr ck_tile::index_t NDimSpatial = 2;
constexpr ck_tile::index_t VectorSize = 8;
using thread_tile = ck_tile::sequence<8, 8>;
using warp_tile = ck_tile::sequence<64, 64>;
using block_tile = ck_tile::sequence<128, 128>;
using Shape = ck_tile::TileImageToColumnShape<thread_tile, warp_tile, block_tile>;
using InDataType = ck_tile::half_t;
using OutDataType = ck_tile::half_t;
using PipelineProblem = ck_tile::BlockImageToColumnProblem<InDataType,
OutDataType,
Shape,
NDimSpatial,
VectorSize,
VectorSize>;
using Kernel = ck_tile::ImageToColumn<PipelineProblem>;
auto kargs = Kernel::MakeKargs(args.p_in,
args.p_out,
args.G,
args.N,
args.C,
args.input_spatial_lengths,
args.filter_spatial_lengths,
args.output_spatial_lengths,
args.image_g_n_c_wis_strides,
args.gemm_g_m_k_strides,
args.conv_filter_strides,
args.conv_filter_dilations,
args.input_left_pads,
args.input_right_pads);
const dim3 grids = Kernel::GridSize(
args.N * args.output_spatial_lengths[0] * args.output_spatial_lengths[1],
args.filter_spatial_lengths[0] * args.filter_spatial_lengths[1] * args.C,
args.G);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 2;
float ave_time = ck_tile::launch_kernel(
stream_conf,
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
return 0;
}
int main(int argc, char* argv[])
{
constexpr ck_tile::index_t NDimSpatial = 2;
ExecutionConfig config;
ck_tile::conv::ConvParam conv_params = DefaultConvParams;
if(!parse_cmd_args(argc, argv, config, conv_params))
{
return EXIT_FAILURE;
}
if(conv_params.num_dim_spatial_ != NDimSpatial)
{
std::cerr << "unsupported # of spatial dimensions" << std::endl;
return EXIT_FAILURE;
}
using InDataType = ck_tile::half_t;
using OutDataType = ck_tile::half_t;
using ImLayout = ck_tile::tensor_layout::convolution::NHWGC;
const auto G = conv_params.G_;
const auto N = conv_params.N_;
const auto C = conv_params.C_;
const ck_tile::long_index_t NHoWo =
N * std::accumulate(conv_params.output_spatial_lengths_.begin(),
std::next(conv_params.output_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const ck_tile::long_index_t CYX =
C * std::accumulate(conv_params.filter_spatial_lengths_.begin(),
std::next(conv_params.filter_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const auto in_desc =
ck_tile::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
const auto out_desc = ck_tile::HostTensorDescriptor({G, NHoWo, CYX});
// host verify
ck_tile::HostTensor<InDataType> in(in_desc);
ck_tile::HostTensor<OutDataType> out_device(out_desc);
ck_tile::HostTensor<OutDataType> out_host(out_desc);
switch(config.init_method)
{
case 0: break;
case 1: ck_tile::FillUniformDistributionIntegerValue<InDataType>{-5.f, 5.f}(in); break;
default: ck_tile::FillUniformDistribution<InDataType>{-0.5, 0.5}(in); break;
}
ck_tile::DeviceMem in_device_buf(in.get_element_space_size_in_bytes());
ck_tile::DeviceMem out_device_buf(out_device.get_element_space_size_in_bytes());
in_device_buf.ToDevice(in.data());
image_to_column_traits traits{"fp16"};
image_to_column_args<NDimSpatial> args{
in_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
G,
N,
C,
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.filter_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.output_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial + 3>(in_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, 3>(out_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_strides_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_dilations_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_left_pads_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_right_pads_)};
float ave_time =
image_to_column(traits, args, ck_tile::stream_config{nullptr, config.time_kernel});
std::size_t num_btype = G * NHoWo * CYX * (sizeof(OutDataType) + sizeof(InDataType));
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
bool pass = true;
if(config.do_verification)
{
// reference
ck_tile::reference_im2col<InDataType, OutDataType, NDimSpatial>(in, out_host, conv_params);
out_device_buf.FromDevice(out_device.data());
pass = ck_tile::check_err(out_device, out_host);
std::cout << "valid:" << (pass ? "y" : "n") << std::endl;
}
return !pass;
}
example/ck_tile/04_img2col/image_to_column.hpp 0 → 100644
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/image_to_column.hpp"
#include <string>
#define DefaultConvParams \
ck_tile::conv::ConvParam \
{ \
2, 2, 32, 32, 32, {4, 4}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, { 0, 0 } \
}
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
inline void print_help_msg()
{
std::cerr << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck_tile::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ExecutionConfig& config,
ck_tile::conv::ConvParam& conv_params)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1)
{
// use default
config = ExecutionConfig{};
}
// catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
// catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
const ck_tile::index_t num_dim_spatial = std::stoi(argv[4]);
conv_params =
ck_tile::conv::parse_conv_param(num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
}
return true;
}
struct image_to_column_traits
{
std::string data_type;
};
template <ck_tile::index_t NDimSpatial>
struct image_to_column_args
{
const void* p_in;
void* p_out;
const ck_tile::long_index_t G;
const ck_tile::long_index_t N;
const ck_tile::long_index_t C;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_spatial_lengths;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> filter_spatial_lengths;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> output_spatial_lengths;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial + 3> image_g_n_c_wis_strides;
const ck_tile::array<ck_tile::long_index_t, 3> gemm_g_m_k_strides;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> conv_filter_strides;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> conv_filter_dilations;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_left_pads;
const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_right_pads;
};
// host API
template <ck_tile::index_t NDimSpatial>
float image_to_column(const image_to_column_traits&,
const image_to_column_args<NDimSpatial>&,
const ck_tile::stream_config&);
example/ck_tile/CMakeLists.txt
View file @ 7ffb0921
......@@ -5,3 +5,4 @@ include_directories(AFTER
add_subdirectory(01_fmha)
add_subdirectory(02_layernorm2d)
add_subdirectory(03_gemm)
add_subdirectory(04_img2col)
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp
View file @ 7ffb0921
......@@ -406,7 +406,7 @@ struct BlockwiseGemmXdlops_pipeline_v4
}
template <>
__device__ static constexpr auto TailScheduler<1>()
__device__ constexpr auto TailScheduler<1>()
{
// schedule
constexpr auto num_ds_read_inst =
......@@ -433,7 +433,7 @@ struct BlockwiseGemmXdlops_pipeline_v4
}
template <>
__device__ static constexpr auto TailScheduler<2>()
__device__ constexpr auto TailScheduler<2>()
{
// schedule
constexpr auto num_ds_read_inst =
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp
View file @ 7ffb0921
......@@ -308,7 +308,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -390,9 +390,10 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp
View file @ 7ffb0921
......@@ -350,7 +350,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -443,7 +443,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -518,9 +518,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......@@ -575,9 +576,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp
View file @ 7ffb0921
......@@ -427,7 +427,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -504,9 +504,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp
View file @ 7ffb0921
......@@ -64,7 +64,7 @@ __global__ void
const index_t N = gemm_desc_ptr[group_id].N;
const index_t K = gemm_desc_ptr[group_id].K;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
return;
const auto StrideAs = gemm_desc_ptr[group_id].StrideAs;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_splitk_xdl_cshuffle_two_stage.hpp
View file @ 7ffb0921
......@@ -345,7 +345,7 @@ struct DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage
const index_t N = gemm_descs[i].N_;
const index_t K = gemm_descs[i].K_;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
{
skipped_group_count_++;
continue;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp
View file @ 7ffb0921
......@@ -109,7 +109,7 @@ __global__ void
N = gemm_desc_ptr[group_id].N;
K = gemm_desc_ptr[group_id].K;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
{
grid_size_grp = 0;
continue;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
View file @ 7ffb0921
......@@ -68,7 +68,7 @@ __global__ void
const index_t N = gemm_desc_ptr[group_id].N;
const index_t K = gemm_desc_ptr[group_id].K;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
return;
const auto StrideA = gemm_desc_ptr[group_id].StrideA;
......
include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
View file @ 7ffb0921
......@@ -324,55 +324,55 @@ struct DppSelector
static constexpr auto GetDpp();
template <>
static constexpr auto GetDpp<half_t, 8, 32>()
constexpr auto GetDpp<half_t, 8, 32>()
{
return DppInstr::dpp8_f16_8x32x2;
}
template <>
static constexpr auto GetDpp<half_t, 8, 16>()
constexpr auto GetDpp<half_t, 8, 16>()
{
return DppInstr::dpp8_f16_8x16x2;
}
template <>
static constexpr auto GetDpp<half_t, 16, 16>()
constexpr auto GetDpp<half_t, 16, 16>()
{
return DppInstr::dpp8_f16_16x16x2;
}
template <>
static constexpr auto GetDpp<half_t, 32, 8>()
constexpr auto GetDpp<half_t, 32, 8>()
{
return DppInstr::dpp8_f16_32x8x2;
}
template <>
static constexpr auto GetDpp<half_t, 1, 32>()
constexpr auto GetDpp<half_t, 1, 32>()
{
return DppInstr::dpp8_f16_1x32x2;
}
template <>
static constexpr auto GetDpp<half_t, 2, 32>()
constexpr auto GetDpp<half_t, 2, 32>()
{
return DppInstr::dpp8_f16_2x32x2;
}
template <>
static constexpr auto GetDpp<half_t, 2, 16>()
constexpr auto GetDpp<half_t, 2, 16>()
{
return DppInstr::dpp8_f16_2x16x2;
}
template <>
static constexpr auto GetDpp<half_t, 4, 16>()
constexpr auto GetDpp<half_t, 4, 16>()
{
return DppInstr::dpp8_f16_4x16x2;
}
template <>
static constexpr auto GetDpp<half_t, 4, 32>()
constexpr auto GetDpp<half_t, 4, 32>()
{
return DppInstr::dpp8_f16_4x32x2;
}
......
include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
View file @ 7ffb0921
......@@ -415,7 +415,7 @@ struct WmmaSelector
static constexpr auto GetWmma();
template <>
static constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_f32_16x16x16_f16_gfx12;
......@@ -425,7 +425,7 @@ struct WmmaSelector
}
template <>
static constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_f32_16x16x16_bf16_gfx12;
......@@ -435,19 +435,19 @@ struct WmmaSelector
}
template <>
static constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
{
return WmmaInstr::wmma_f16_16x16x16_f16;
}
template <>
static constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
{
return WmmaInstr::wmma_bf16_16x16x16_bf16;
}
template <>
static constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_i32_16x16x16_iu8_gfx12;
......@@ -458,7 +458,7 @@ struct WmmaSelector
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
template <>
static constexpr auto GetWmma<int4_t, int4_t, int, 16, 16>()
constexpr auto GetWmma<int4_t, int4_t, int, 16, 16>()
{
return WmmaInstr::wmma_i32_16x16x16_iu4;
}
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 7ffb0921
......@@ -651,97 +651,97 @@ struct MfmaSelector
static constexpr auto GetMfma();
template <>
static constexpr auto GetMfma<double, 16, 16>()
constexpr auto GetMfma<double, 16, 16>()
{
return MfmaInstr::mfma_f64_16x16x4f64;
}
template <>
static constexpr auto GetMfma<float, 64, 64>()
constexpr auto GetMfma<float, 64, 64>()
{
return MfmaInstr::mfma_f32_32x32x1xf32;
}
template <>
static constexpr auto GetMfma<float, 32, 64>()
constexpr auto GetMfma<float, 32, 64>()
{
return MfmaInstr::mfma_f32_32x32x1xf32;
}
template <>
static constexpr auto GetMfma<float, 16, 64>()
constexpr auto GetMfma<float, 16, 64>()
{
return MfmaInstr::mfma_f32_16x16x1xf32;
}
template <>
static constexpr auto GetMfma<float, 8, 64>()
constexpr auto GetMfma<float, 8, 64>()
{
return MfmaInstr::mfma_f32_4x4x1xf32;
}
template <>
static constexpr auto GetMfma<float, 4, 64>()
constexpr auto GetMfma<float, 4, 64>()
{
return MfmaInstr::mfma_f32_4x4x1xf32;
}
template <>
static constexpr auto GetMfma<float, 32, 32>()
constexpr auto GetMfma<float, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x2xf32;
}
template <>
static constexpr auto GetMfma<float, 16, 16>()
constexpr auto GetMfma<float, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x4xf32;
}
template <>
static constexpr auto GetMfma<half_t, 64, 64>()
constexpr auto GetMfma<half_t, 64, 64>()
{
return MfmaInstr::mfma_f32_32x32x4f16;
}
template <>
static constexpr auto GetMfma<half_t, 32, 64>()
constexpr auto GetMfma<half_t, 32, 64>()
{
return MfmaInstr::mfma_f32_32x32x4f16;
}
template <>
static constexpr auto GetMfma<half_t, 32, 32>()
constexpr auto GetMfma<half_t, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x8f16;
}
template <>
static constexpr auto GetMfma<half_t, 16, 16>()
constexpr auto GetMfma<half_t, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x16f16;
}
template <>
static constexpr auto GetMfma<half_t, 16, 64>()
constexpr auto GetMfma<half_t, 16, 64>()
{
return MfmaInstr::mfma_f32_16x16x4f16;
}
template <>
static constexpr auto GetMfma<half_t, 8, 64>()
constexpr auto GetMfma<half_t, 8, 64>()
{
return MfmaInstr::mfma_f32_4x4x4f16;
}
template <>
static constexpr auto GetMfma<half_t, 4, 64>()
constexpr auto GetMfma<half_t, 4, 64>()
{
return MfmaInstr::mfma_f32_4x4x4f16;
}
template <>
static constexpr auto GetMfma<bhalf_t, 32, 32>()
constexpr auto GetMfma<bhalf_t, 32, 32>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_32x32x8bf16_1k;
......@@ -751,7 +751,7 @@ struct MfmaSelector
}
template <>
static constexpr auto GetMfma<bhalf_t, 16, 16>()
constexpr auto GetMfma<bhalf_t, 16, 16>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_16x16x16bf16_1k;
......@@ -762,72 +762,72 @@ struct MfmaSelector
#if defined(CK_USE_AMD_MFMA_GFX940)
template <>
static constexpr auto GetMfma<int8_t, 32, 32>()
constexpr auto GetMfma<int8_t, 32, 32>()
{
return MfmaInstr::mfma_i32_32x32x16i8;
}
template <>
static constexpr auto GetMfma<int8_t, 16, 16>()
constexpr auto GetMfma<int8_t, 16, 16>()
{
return MfmaInstr::mfma_i32_16x16x32i8;
}
#else
template <>
static constexpr auto GetMfma<int8_t, 32, 32>()
constexpr auto GetMfma<int8_t, 32, 32>()
{
return MfmaInstr::mfma_i32_32x32x8i8;
}
template <>
static constexpr auto GetMfma<int8_t, 16, 16>()
constexpr auto GetMfma<int8_t, 16, 16>()
{
return MfmaInstr::mfma_i32_16x16x16i8;
}
#endif
template <>
static constexpr auto GetMfma<f8_t, 32, 32>()
constexpr auto GetMfma<f8_t, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x16f8f8;
}
template <>
static constexpr auto GetMfma<f8_t, 16, 16>()
constexpr auto GetMfma<f8_t, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x32f8f8;
}
template <>
static constexpr auto GetMfma<bf8_t, 32, 32>()
constexpr auto GetMfma<bf8_t, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x16bf8bf8;
}
template <>
static constexpr auto GetMfma<bf8_t, 16, 16>()
constexpr auto GetMfma<bf8_t, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x32bf8bf8;
}
template <>
static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
{
return MfmaInstr::mfma_f32_32x32x16f8bf8;
}
template <>
static constexpr auto GetMfma<f8_t, 16, 16, bf8_t>()
constexpr auto GetMfma<f8_t, 16, 16, bf8_t>()
{
return MfmaInstr::mfma_f32_16x16x32f8bf8;
}
template <>
static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
{
return MfmaInstr::mfma_f32_32x32x16bf8f8;
}
template <>
static constexpr auto GetMfma<bf8_t, 16, 16, f8_t>()
constexpr auto GetMfma<bf8_t, 16, 16, f8_t>()
{
return MfmaInstr::mfma_f32_16x16x32bf8f8;
}
......
include/ck_tile/core/container/array.hpp
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <initializer_list>
#include <vector>
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/numeric/integer.hpp"
......@@ -236,6 +237,16 @@ CK_TILE_HOST_DEVICE constexpr bool operator!=(const array<T, Size>& a, const arr
return !(a == b);
}
template <typename T, index_t N, typename X>
CK_TILE_HOST_DEVICE constexpr auto to_array(const std::vector<X>& x)
{
array<T, N> arr;
static_for<0, N, 1>{}([&x, &arr](auto i) { arr(i) = x[i]; });
return arr;
}
template <typename T, index_t N, typename X>
CK_TILE_HOST_DEVICE constexpr auto to_array(const X& x)
{
......
include/ck_tile/host.hpp
View file @ 7ffb0921
......@@ -5,6 +5,8 @@
#include "ck_tile/host/arg_parser.hpp"
#include "ck_tile/host/check_err.hpp"
#include "ck_tile/host/convolution_host_tensor_descriptor_helper.hpp"
#include "ck_tile/host/convolution_parameter.hpp"
#include "ck_tile/host/device_memory.hpp"
#include "ck_tile/host/fill.hpp"
#include "ck_tile/host/hip_check_error.hpp"
......
include/ck_tile/host/convolution_host_tensor_descriptor_helper.hpp 0 → 100644
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/host/convolution_parameter.hpp"
#include "ck_tile/host/host_tensor.hpp"
namespace ck_tile {
namespace conv {
namespace detail {
template <typename OldLayout>
CK_TILE_HOST std::vector<std::size_t> get_layout_transpose_gnchw_to_old()
{
if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKW>)
{
return {0, 1, 2, 3};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKHW>)
{
return {0, 1, 2, 3, 4};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCDHW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCZYX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
{
return {0, 1, 2, 3, 4, 5};
}
if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWK>)
{
return {0, 1, 3, 2};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWK>)
{
return {0, 1, 4, 2, 3};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKZYXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
{
return {0, 1, 5, 2, 3, 4};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGK>)
{
return {2, 0, 3, 1};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGK>)
{
return {3, 0, 4, 1, 2};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KZYXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
{
return {4, 0, 5, 1, 2, 3};
}
else
{
printf("%s\n", __func__);
throw std::runtime_error("wrong! unsupported layout");
}
}
} // namespace detail
// make tensor descriptor for packed input tensor, and order the dimension in the order of GNCHW
// regardless of physical layout
template <typename InLayout>
CK_TILE_HOST HostTensorDescriptor
make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck_tile::conv::ConvParam& param)
{
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCW> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCDHW>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.end(),
param.input_spatial_lengths_.begin(),
param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNWC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNDHWC>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.begin() + 2,
param.input_spatial_lengths_.begin(),
param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NWGC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NDHWGC>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.begin() + 1,
param.input_spatial_lengths_.begin(),
param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else
{
printf("%s\n", __func__);
printf("%s\n", InLayout::name);
throw std::runtime_error("wrong! unsupported layout");
}
return transpose_host_tensor_descriptor_given_new2old(
HostTensorDescriptor(physical_lengths),
detail::get_layout_transpose_gnchw_to_old<InLayout>());
}
// make tensor descriptor for packed weight tensor, and order the dimension in the order of GKCYX
// regardless of physical layout
template <typename WeiLayout>
CK_TILE_HOST HostTensorDescriptor
make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvParam& param)
{
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXC>)
{
if(param.G_ != 1)
{
throw std::runtime_error("wrong! G != 1");
}
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.end(),
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCX> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCZYX>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.K_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.end(),
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKZYXC>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.K_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.begin() + 2,
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXGC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXGC>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.C_)};
physical_lengths.insert(physical_lengths.begin() + 1,
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else
{
printf("%s\n", __func__);
printf("%s\n", WeiLayout::name);
throw std::runtime_error("wrong! unsupported layout");
}
return transpose_host_tensor_descriptor_given_new2old(
HostTensorDescriptor(physical_lengths),
detail::get_layout_transpose_gnchw_to_old<WeiLayout>());
}
// make tensor descriptor for packed output tensor, and order the dimension in the order of GNKHW
// regardless of physical layout
template <typename OutLayout>
CK_TILE_HOST HostTensorDescriptor
make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck_tile::conv::ConvParam& param)
{
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKW> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKHW> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.K_)};
physical_lengths.insert(physical_lengths.end(),
param.output_spatial_lengths_.begin(),
param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
}
// separate from legacy code above
else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNWK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNHWK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.K_)};
physical_lengths.insert(physical_lengths.begin() + 2,
param.output_spatial_lengths_.begin(),
param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NWGK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NHWGK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.K_)};
physical_lengths.insert(physical_lengths.begin() + 1,
param.output_spatial_lengths_.begin(),
param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else
{
printf("%s\n", __func__);
printf("%s\n", OutLayout::name);
throw std::runtime_error("wrong! unsupported layout");
}
return transpose_host_tensor_descriptor_given_new2old(
HostTensorDescriptor(physical_lengths),
detail::get_layout_transpose_gnchw_to_old<OutLayout>());
}
} // namespace conv
} // namespace ck_tile
include/ck_tile/host/convolution_parameter.hpp 0 → 100644
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <numeric>
#include <iterator>
#include <vector>
namespace ck_tile {
namespace conv {
struct ConvParam
{
ConvParam();
ConvParam(ck_tile::index_t n_dim,
ck_tile::index_t group_count,
ck_tile::index_t n_batch,
ck_tile::index_t n_out_channels,
ck_tile::index_t n_in_channels,
const std::vector<ck_tile::index_t>& filters_len,
const std::vector<ck_tile::index_t>& input_len,
const std::vector<ck_tile::index_t>& strides,
const std::vector<ck_tile::index_t>& dilations,
const std::vector<ck_tile::index_t>& left_pads,
const std::vector<ck_tile::index_t>& right_pads)
: num_dim_spatial_(static_cast<ck_tile::long_index_t>(n_dim)),
G_(static_cast<ck_tile::long_index_t>(group_count)),
N_(static_cast<ck_tile::long_index_t>(n_batch)),
K_(static_cast<ck_tile::long_index_t>(n_out_channels)),
C_(static_cast<ck_tile::long_index_t>(n_in_channels)),
filter_spatial_lengths_(num_dim_spatial_),
input_spatial_lengths_(num_dim_spatial_),
output_spatial_lengths_(num_dim_spatial_),
conv_filter_strides_(num_dim_spatial_),
conv_filter_dilations_(num_dim_spatial_),
input_left_pads_(num_dim_spatial_),
input_right_pads_(num_dim_spatial_)
{
if(static_cast<ck_tile::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_right_pads_.size()) != num_dim_spatial_)
{
throw(std::runtime_error(
"ConvParam::ConvParam: "
"parameter size is different from number of declared dimensions!"));
}
for(ck_tile::index_t i = 0; i < num_dim_spatial_; ++i)
{
filter_spatial_lengths_[i] = static_cast<ck_tile::long_index_t>(filters_len[i]);
input_spatial_lengths_[i] = static_cast<ck_tile::long_index_t>(input_len[i]);
conv_filter_strides_[i] = static_cast<ck_tile::long_index_t>(strides[i]);
conv_filter_dilations_[i] = static_cast<ck_tile::long_index_t>(dilations[i]);
input_left_pads_[i] = static_cast<ck_tile::long_index_t>(left_pads[i]);
input_right_pads_[i] = static_cast<ck_tile::long_index_t>(right_pads[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_tile::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
output_spatial_lengths_[i] =
(input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
conv_filter_strides_[i] +
1;
}
}
ConvParam(ck_tile::long_index_t n_dim,
ck_tile::long_index_t group_count,
ck_tile::long_index_t n_batch,
ck_tile::long_index_t n_out_channels,
ck_tile::long_index_t n_in_channels,
const std::vector<ck_tile::long_index_t>& filters_len,
const std::vector<ck_tile::long_index_t>& input_len,
const std::vector<ck_tile::long_index_t>& strides,
const std::vector<ck_tile::long_index_t>& dilations,
const std::vector<ck_tile::long_index_t>& left_pads,
const std::vector<ck_tile::long_index_t>& right_pads)
: num_dim_spatial_(n_dim),
G_(group_count),
N_(n_batch),
K_(n_out_channels),
C_(n_in_channels),
filter_spatial_lengths_(filters_len),
input_spatial_lengths_(input_len),
output_spatial_lengths_(num_dim_spatial_),
conv_filter_strides_(strides),
conv_filter_dilations_(dilations),
input_left_pads_(left_pads),
input_right_pads_(right_pads)
{
if(static_cast<ck_tile::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
static_cast<ck_tile::index_t>(input_right_pads_.size()) != num_dim_spatial_)
{
throw(std::runtime_error(
"ConvParam::ConvParam: "
"parameter size is different from number of declared dimensions!"));
}
for(ck_tile::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_tile::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
output_spatial_lengths_[i] =
(input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
conv_filter_strides_[i] +
1;
}
}
ck_tile::long_index_t num_dim_spatial_;
ck_tile::long_index_t G_;
ck_tile::long_index_t N_;
ck_tile::long_index_t K_;
ck_tile::long_index_t C_;
std::vector<ck_tile::long_index_t> filter_spatial_lengths_;
std::vector<ck_tile::long_index_t> input_spatial_lengths_;
std::vector<ck_tile::long_index_t> output_spatial_lengths_;
std::vector<ck_tile::long_index_t> conv_filter_strides_;
std::vector<ck_tile::long_index_t> conv_filter_dilations_;
std::vector<ck_tile::long_index_t> input_left_pads_;
std::vector<ck_tile::long_index_t> input_right_pads_;
std::vector<ck_tile::long_index_t> GetOutputSpatialLengths() const
{
return output_spatial_lengths_;
}
std::size_t GetFlops() const
{
// 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
return static_cast<std::size_t>(2) * G_ * N_ * K_ * C_ *
std::accumulate(std::begin(output_spatial_lengths_),
std::next(std::begin(output_spatial_lengths_), num_dim_spatial_),
1,
std::multiplies<>()) *
std::accumulate(std::begin(filter_spatial_lengths_),
std::next(std::begin(filter_spatial_lengths_), num_dim_spatial_),
1,
std::multiplies<>());
}
template <typename InDataType>
std::size_t GetInputByte() const
{
// sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
return sizeof(InDataType) *
(G_ * N_ * C_ *
std::accumulate(std::begin(input_spatial_lengths_),
std::next(std::begin(input_spatial_lengths_), num_dim_spatial_),
1,
std::multiplies<>()));
}
template <typename WeiDataType>
std::size_t GetWeightByte() const
{
// sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
return sizeof(WeiDataType) *
(G_ * K_ * C_ *
std::accumulate(std::begin(filter_spatial_lengths_),
std::next(std::begin(filter_spatial_lengths_), num_dim_spatial_),
1,
std::multiplies<>()));
}
template <typename OutDataType>
std::size_t GetOutputByte() const
{
// sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
return sizeof(OutDataType) * (G_ * 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>()));
}
template <typename InDataType, typename WeiDataType, typename OutDataType>
std::size_t GetByte() const
{
return GetInputByte<InDataType>() + GetWeightByte<WeiDataType>() +
GetOutputByte<OutDataType>();
}
};
ConvParam::ConvParam()
: ConvParam::ConvParam(2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1})
{
}
CK_TILE_HOST std::string get_conv_param_parser_helper_msg()
{
std::string msg;
msg += "Following arguments (depending on number of spatial dims):\n"
" Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)\n"
" G, 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";
return msg;
}
CK_TILE_HOST ck_tile::conv::ConvParam
parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck_tile::long_index_t G = std::stol(argv[arg_idx++]);
const ck_tile::long_index_t N = std::stol(argv[arg_idx++]);
const ck_tile::long_index_t K = std::stol(argv[arg_idx++]);
const ck_tile::long_index_t C = std::stol(argv[arg_idx++]);
std::vector<ck_tile::long_index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck_tile::long_index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck_tile::long_index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck_tile::long_index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck_tile::long_index_t> input_left_pads(num_dim_spatial);
std::vector<ck_tile::long_index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stol(argv[arg_idx++]);
}
return ck_tile::conv::ConvParam{num_dim_spatial,
G,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
} // namespace conv
} // namespace ck_tile
include/ck_tile/host/host_tensor.hpp
View file @ 7ffb0921
......@@ -176,7 +176,20 @@ struct HostTensorDescriptor
return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
}
friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc)
{
os << "dim " << desc.get_num_of_dimension() << ", ";
os << "lengths {";
LogRange(os, desc.get_lengths(), ", ");
os << "}, ";
os << "strides {";
LogRange(os, desc.get_strides(), ", ");
os << "}";
return os;
}
private:
std::vector<std::size_t> mLens;
......
include/ck_tile/host/reference/reference_im2col.hpp
View file @ 7ffb0921
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -9,53 +9,125 @@
namespace ck_tile {
template <typename T>
CK_TILE_HOST void reference_im2col(HostTensor<T>& in_mtx_host_ref,
const HostTensor<T>& in_host,
int /*N*/,
int /*K*/,
int C,
int /*Y*/,
int X,
int Hi,
int Wi,
int Ho,
int Wo,
int ConvStrideH,
int ConvStrideW,
int ConvDilationH,
int ConvDilationW,
int InLeftPadH,
int InLeftPadW,
int /*InRightPadH*/,
int /*InRightPadW*/)
template <typename InDataType, typename OutDataType, index_t NDimSpatial>
CK_TILE_HOST void reference_im2col(const HostTensor<InDataType>& in_host,
HostTensor<OutDataType>& out_host,
const ck_tile::conv::ConvParam& conv_params)
{
int GemmM = in_mtx_host_ref.get_lengths()[0];
int GemmK = in_mtx_host_ref.get_lengths()[1];
const long_index_t G = in_host.get_lengths()[0];
const long_index_t N = in_host.get_lengths()[1];
const long_index_t C = in_host.get_lengths()[2];
for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
if constexpr(NDimSpatial == 1)
{
int mtmp = gemm_m;
int n = mtmp / (Ho * Wo);
mtmp -= n * Ho * Wo;
int ho = mtmp / Wo;
int wo = mtmp - ho * Wo;
for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
{
int ktmp = gemm_k;
int y = ktmp / (X * C);
ktmp -= y * X * C;
int x = ktmp / C;
int c = ktmp - x * C;
int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
}
const long_index_t Wo = conv_params.output_spatial_lengths_[0];
auto func = [&](auto g, auto n, auto wo) {
long_index_t row = n * Wo + wo;
long_index_t column = 0;
for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[0]; ++x)
{
auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[0]) +
static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[0]) -
static_cast<long_index_t>(conv_params.input_left_pads_[0]);
for(long_index_t c = 0; c < C; ++c)
{
if(wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[3])
{
InDataType v_in = in_host(g, n, c, wi);
out_host(g, row, column) = type_convert<OutDataType>(v_in);
}
column++;
}
}
};
make_ParallelTensorFunctor(func, G, N, Wo)(std::thread::hardware_concurrency());
}
else if constexpr(NDimSpatial == 2)
{
const long_index_t Ho = conv_params.output_spatial_lengths_[0];
const long_index_t Wo = conv_params.output_spatial_lengths_[1];
auto func = [&](auto g, auto n, auto ho, auto wo) {
long_index_t row = n * Ho * Wo + ho * Wo + wo;
long_index_t column = 0;
for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[0]; ++y)
{
auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[0]) +
static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[0]) -
static_cast<long_index_t>(conv_params.input_left_pads_[0]);
for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[1]; ++x)
{
auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[1]) +
static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[1]) -
static_cast<long_index_t>(conv_params.input_left_pads_[1]);
for(long_index_t c = 0; c < C; ++c)
{
if(hi >= 0 && type_convert<std::size_t>(hi) < in_host.get_lengths()[3] &&
wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[4])
{
InDataType v_in = in_host(g, n, c, hi, wi);
out_host(g, row, column) = type_convert<OutDataType>(v_in);
}
column++;
}
}
}
};
make_ParallelTensorFunctor(func, G, N, Ho, Wo)(std::thread::hardware_concurrency());
}
else if constexpr(NDimSpatial == 3)
{
const long_index_t Do = conv_params.output_spatial_lengths_[0];
const long_index_t Ho = conv_params.output_spatial_lengths_[1];
const long_index_t Wo = conv_params.output_spatial_lengths_[2];
auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) {
long_index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
long_index_t column = 0;
for(long_index_t z = 0; z < conv_params.filter_spatial_lengths_[0]; ++z)
{
auto di = static_cast<long_index_t>(d_o * conv_params.conv_filter_strides_[0]) +
static_cast<long_index_t>(z * conv_params.conv_filter_dilations_[0]) -
static_cast<long_index_t>(conv_params.input_left_pads_[0]);
for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[1]; ++y)
{
auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[1]) +
static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[1]) -
static_cast<long_index_t>(conv_params.input_left_pads_[1]);
for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[2]; ++x)
{
auto wi =
static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[2]) +
static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[2]) -
static_cast<long_index_t>(conv_params.input_left_pads_[2]);
for(long_index_t c = 0; c < C; ++c)
{
if(di >= 0 &&
type_convert<std::size_t>(di) < in_host.get_lengths()[3] &&
hi >= 0 &&
type_convert<std::size_t>(hi) < in_host.get_lengths()[4] &&
wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[5])
{
InDataType v_in = in_host(g, n, c, di, hi, wi);
out_host(g, row, column) = type_convert<OutDataType>(v_in);
}
column++;
}
}
}
}
};
make_ParallelTensorFunctor(func, G, N, Do, Ho, Wo)(std::thread::hardware_concurrency());
}
}
} // namespace ck_tile
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