Skip to content

GitLab

Unverified Commit a93d07c7 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into ck_codegen_build

parents 9d9ad510 afbf6350
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1065 additions and 272 deletions
include/ck/utility/f8_utils.hpp
View file @ a93d07c7
...@@ -165,7 +165,7 @@ In this case, the fp16 mantissa should be shift left by 1 */ ...@@ -165,7 +165,7 @@ In this case, the fp16 mantissa should be shift left by 1 */
if(out_exponent > max_exp) if(out_exponent > max_exp)
{ {
if(clip) if constexpr(clip)
{ {
mantissa = (1 << out_mant) - 1; mantissa = (1 << out_mant) - 1;
out_exponent = max_exp; out_exponent = max_exp;
...@@ -235,7 +235,8 @@ __host__ __device__ Y run_cast_from_f8(X x) ...@@ -235,7 +235,8 @@ __host__ __device__ Y run_cast_from_f8(X x)
return (mantissa == 0) ? (sign ? NegInf : Inf) : NaN; return (mantissa == 0) ? (sign ? NegInf : Inf) : NaN;
} }
if((NumericUtils<Y>::mant == 10) && (NumericUtils<X>::mant == 2) && !negative_zero_nan) if constexpr((NumericUtils<Y>::mant == 10) && (NumericUtils<X>::mant == 2) &&
!negative_zero_nan)
{ {
retval = x; retval = x;
retval <<= 8; retval <<= 8;
......
include/ck/utility/type_convert.hpp
View file @ a93d07c7
// SPDX-License-Identifier: MIT // 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 #pragma once
#include "ck/utility/data_type.hpp" #include "ck/utility/data_type.hpp"
#include "ck/utility/f8_utils.hpp" #include "ck/utility/f8_utils.hpp"
#include "ck/utility/random_gen.hpp" #include "ck/utility/random_gen.hpp"
#include "ck/utility/array.hpp"
namespace ck { namespace ck {
// Define the common macro for gfx94x models // Define the common macro for gfx94x models
...@@ -500,6 +501,25 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x) ...@@ -500,6 +501,25 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
#endif #endif
} }
template <typename Y, typename X, std::size_t NumElems>
inline __host__ __device__ void array_convert(std::array<Y, NumElems>& y,
const std::array<X, NumElems>& x)
{
for(std::size_t i = 0; i < NumElems; i++)
{
y[i] = type_convert<Y>(x[i]);
}
}
template <typename Y, typename X, index_t NumElems>
inline __host__ __device__ void array_convert(Array<Y, NumElems>& y, const Array<X, NumElems>& x)
{
for(std::size_t i = 0; i < NumElems; i++)
{
y[i] = type_convert<Y>(x[i]);
}
}
// Declare a template function for bf16 conversion using RTN // Declare a template function for bf16 conversion using RTN
template <typename Y, typename X> template <typename Y, typename X>
__host__ __device__ constexpr Y bf16_convert_rtn(X x); __host__ __device__ constexpr Y bf16_convert_rtn(X x);
......
library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
View file @ a93d07c7
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -39,11 +39,11 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -39,11 +39,11 @@ struct ReferenceColumnToImage : public device::BaseOperator
public: public:
Argument(const Tensor<InDataType>& input, Argument(const Tensor<InDataType>& input,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> filter_spatial_lengths, std::vector<ck::long_index_t> filter_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads) std::vector<ck::long_index_t> input_right_pads)
: input_{input}, : input_{input},
output_{output}, output_{output},
conv_strides_{conv_filter_strides}, conv_strides_{conv_filter_strides},
...@@ -58,24 +58,25 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -58,24 +58,25 @@ struct ReferenceColumnToImage : public device::BaseOperator
const Tensor<InDataType>& input_; const Tensor<InDataType>& input_;
Tensor<OutDataType>& output_; Tensor<OutDataType>& output_;
std::vector<index_t> conv_strides_; std::vector<long_index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<long_index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<long_index_t> in_left_pads_;
std::vector<index_t> in_right_pads_; std::vector<long_index_t> in_right_pads_;
std::vector<index_t> filter_spatial_lengths_; std::vector<long_index_t> filter_spatial_lengths_;
std::vector<index_t> output_spatial_lengths_; std::vector<long_index_t> output_spatial_lengths_;
private: private:
void initOutputSpatialLengths() void initOutputSpatialLengths()
{ {
constexpr auto input_offset_to_spatial = 3; constexpr auto input_offset_to_spatial = 3;
for(ck::index_t i = 0; i < NDimSpatial; ++i) for(ck::long_index_t i = 0; i < NDimSpatial; ++i)
{ {
// XEff = (X - 1) * conv_dilation_w + 1; // XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1; // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1; const ck::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1;
output_spatial_lengths_.push_back( output_spatial_lengths_.push_back(
(output_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] + (output_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] +
...@@ -98,26 +99,26 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -98,26 +99,26 @@ struct ReferenceColumnToImage : public device::BaseOperator
throw std::runtime_error("wrong! inconsistent dimension"); throw std::runtime_error("wrong! inconsistent dimension");
} }
const index_t G = arg.output_.GetLengths()[0]; const long_index_t G = arg.output_.GetLengths()[0];
const index_t N = arg.output_.GetLengths()[1]; const long_index_t N = arg.output_.GetLengths()[1];
const index_t C = arg.output_.GetLengths()[2]; const long_index_t C = arg.output_.GetLengths()[2];
if constexpr(NDimSpatial == 1) if constexpr(NDimSpatial == 1)
{ {
const index_t Wo = arg.output_spatial_lengths_[0]; const long_index_t Wo = arg.output_spatial_lengths_[0];
auto func = [&](auto g, auto n) { auto func = [&](auto g, auto n) {
for(index_t wo = 0; wo < Wo; ++wo) for(long_index_t wo = 0; wo < Wo; ++wo)
{ {
index_t row = n * Wo + wo; long_index_t row = n * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x)
{ {
auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) + auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.output_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.output_.GetLengths()[3])
...@@ -140,32 +141,32 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -140,32 +141,32 @@ struct ReferenceColumnToImage : public device::BaseOperator
} }
else if constexpr(NDimSpatial == 2) else if constexpr(NDimSpatial == 2)
{ {
const index_t Ho = arg.output_spatial_lengths_[0]; const long_index_t Ho = arg.output_spatial_lengths_[0];
const index_t Wo = arg.output_spatial_lengths_[1]; const long_index_t Wo = arg.output_spatial_lengths_[1];
auto func = [&](auto g, auto n) { auto func = [&](auto g, auto n) {
for(index_t ho = 0; ho < Ho; ++ho) for(long_index_t ho = 0; ho < Ho; ++ho)
{ {
for(index_t wo = 0; wo < Wo; ++wo) for(long_index_t wo = 0; wo < Wo; ++wo)
{ {
index_t row = n * Ho * Wo + ho * Wo + wo; long_index_t row = n * Ho * Wo + ho * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y) for(long_index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y)
{ {
auto hi = auto hi =
static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) + static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x)
{ {
auto wi = auto wi =
static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) + static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) - static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]); static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(hi >= 0 && if(hi >= 0 &&
...@@ -196,27 +197,27 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -196,27 +197,27 @@ struct ReferenceColumnToImage : public device::BaseOperator
} }
else if constexpr(NDimSpatial == 3) else if constexpr(NDimSpatial == 3)
{ {
const index_t Do = arg.output_spatial_lengths_[0]; const long_index_t Do = arg.output_spatial_lengths_[0];
const index_t Ho = arg.output_spatial_lengths_[1]; const long_index_t Ho = arg.output_spatial_lengths_[1];
const index_t Wo = arg.output_spatial_lengths_[2]; const long_index_t Wo = arg.output_spatial_lengths_[2];
auto func = [&](auto g, auto n) { auto func = [&](auto g, auto n) {
for(index_t d_o = 0; d_o < Do; ++d_o) for(long_index_t d_o = 0; d_o < Do; ++d_o)
{ {
for(index_t ho = 0; ho < Ho; ++ho) for(long_index_t ho = 0; ho < Ho; ++ho)
{ {
for(index_t wo = 0; wo < Wo; ++wo) for(long_index_t wo = 0; wo < Wo; ++wo)
{ {
index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo; long_index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z) for(long_index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z)
{ {
auto di = auto di =
static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) + static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y) for(long_index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y)
{ {
auto hi = auto hi =
static_cast<ck::long_index_t>(ho * static_cast<ck::long_index_t>(ho *
...@@ -224,7 +225,8 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -224,7 +225,8 @@ struct ReferenceColumnToImage : public device::BaseOperator
static_cast<ck::long_index_t>(y * static_cast<ck::long_index_t>(y *
arg.conv_dilations_[1]) - arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]); static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(index_t x = 0; x < arg.filter_spatial_lengths_[2]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[2];
++x)
{ {
auto wi = auto wi =
static_cast<ck::long_index_t>( static_cast<ck::long_index_t>(
...@@ -232,7 +234,7 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -232,7 +234,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
static_cast<ck::long_index_t>( static_cast<ck::long_index_t>(
x * arg.conv_dilations_[2]) - x * arg.conv_dilations_[2]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[2]); static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(di >= 0 && if(di >= 0 &&
ck::type_convert<std::size_t>(di) < ck::type_convert<std::size_t>(di) <
...@@ -294,15 +296,15 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -294,15 +296,15 @@ struct ReferenceColumnToImage : public device::BaseOperator
bool IsSupportedArgument(const Argument& arg) bool IsSupportedArgument(const Argument& arg)
{ {
const ck::index_t G = arg.output_.GetLengths()[0]; const ck::long_index_t G = arg.output_.GetLengths()[0];
const ck::index_t N = arg.output_.GetLengths()[1]; const ck::long_index_t N = arg.output_.GetLengths()[1];
const ck::index_t C = arg.output_.GetLengths()[2]; const ck::long_index_t C = arg.output_.GetLengths()[2];
const index_t NDoHoWo = const long_index_t NDoHoWo =
N * ck::accumulate_n<index_t>( N * ck::accumulate_n<long_index_t>(
arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>()); arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
const index_t CZYX = const long_index_t CZYX =
C * ck::accumulate_n<index_t>( C * ck::accumulate_n<long_index_t>(
arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>()); arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
if(!(arg.input_.GetLengths()[0] == static_cast<std::size_t>(G) && if(!(arg.input_.GetLengths()[0] == static_cast<std::size_t>(G) &&
...@@ -326,11 +328,11 @@ struct ReferenceColumnToImage : public device::BaseOperator ...@@ -326,11 +328,11 @@ struct ReferenceColumnToImage : public device::BaseOperator
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(const Tensor<InDataType>& input,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> filter_spatial_lengths, std::vector<ck::long_index_t> filter_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads) std::vector<ck::long_index_t> input_right_pads)
{ {
return Argument{input, return Argument{input,
output, output,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View file @ a93d07c7
...@@ -38,10 +38,10 @@ struct ReferenceConvBwdData : public device::BaseOperator ...@@ -38,10 +38,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
Tensor<InDataType>& input, Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output, const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -72,10 +72,10 @@ struct ReferenceConvBwdData : public device::BaseOperator ...@@ -72,10 +72,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_; const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_; const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<long_index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<long_index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<long_index_t> in_left_pads_;
std::vector<index_t> in_right_pads_; std::vector<long_index_t> in_right_pads_;
InElementwiseOperation in_element_op_; InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_; WeiElementwiseOperation wei_element_op_;
...@@ -447,10 +447,10 @@ struct ReferenceConvBwdData : public device::BaseOperator ...@@ -447,10 +447,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
Tensor<InDataType>& input, Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output, const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
View file @ a93d07c7
...@@ -40,10 +40,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator ...@@ -40,10 +40,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
const Tensor<InDataType>& in_n_c_hi_wi, const Tensor<InDataType>& in_n_c_hi_wi,
Tensor<WeiDataType>& wei_k_c_y_x, Tensor<WeiDataType>& wei_k_c_y_x,
const Tensor<OutDataType>& out_n_k_ho_wo, const Tensor<OutDataType>& out_n_k_ho_wo,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -74,10 +74,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator ...@@ -74,10 +74,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
const std::array<Tensor<InDataType>, NumBElementwiseTensor>& elementwise_b_tensors_; const std::array<Tensor<InDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<WeiDataType>, NumDElementwiseTensor>& elementwise_d_tensors_; const std::array<Tensor<WeiDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<long_index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<long_index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<long_index_t> in_left_pads_;
std::vector<index_t> in_right_pads_; std::vector<long_index_t> in_right_pads_;
InElementwiseOperation in_element_op_; InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_; WeiElementwiseOperation wei_element_op_;
...@@ -402,10 +402,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator ...@@ -402,10 +402,10 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
const Tensor<InDataType>& in_n_c_hi_wi, const Tensor<InDataType>& in_n_c_hi_wi,
Tensor<WeiDataType>& wei_k_c_y_x, Tensor<WeiDataType>& wei_k_c_y_x,
const Tensor<OutDataType>& out_n_k_ho_wo, const Tensor<OutDataType>& out_n_k_ho_wo,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ a93d07c7
// SPDX-License-Identifier: MIT // 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 #pragma once
...@@ -69,10 +69,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -69,10 +69,10 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<InDataType>& input, const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -103,10 +103,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -103,10 +103,10 @@ struct ReferenceConvFwd : public device::BaseOperator
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_; const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_; const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<ck::long_index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<ck::long_index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<ck::long_index_t> in_left_pads_;
std::vector<index_t> in_right_pads_; std::vector<ck::long_index_t> in_right_pads_;
InElementwiseOperation in_element_op_; InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_; WeiElementwiseOperation wei_element_op_;
...@@ -416,10 +416,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -416,10 +416,10 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<InDataType>& input, const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads, std::vector<ck::long_index_t> input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
View file @ a93d07c7
...@@ -40,11 +40,11 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -40,11 +40,11 @@ struct ReferenceImageToColumn : public device::BaseOperator
public: public:
Argument(const Tensor<InDataType>& input, Argument(const Tensor<InDataType>& input,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> filter_spatial_lengths, std::vector<ck::long_index_t> filter_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads) std::vector<ck::long_index_t> input_right_pads)
: input_{input}, : input_{input},
output_{output}, output_{output},
conv_strides_{conv_filter_strides}, conv_strides_{conv_filter_strides},
...@@ -59,13 +59,13 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -59,13 +59,13 @@ struct ReferenceImageToColumn : public device::BaseOperator
const Tensor<InDataType>& input_; const Tensor<InDataType>& input_;
Tensor<OutDataType>& output_; Tensor<OutDataType>& output_;
std::vector<index_t> conv_strides_; std::vector<long_index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<long_index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<long_index_t> in_left_pads_;
std::vector<index_t> in_right_pads_; std::vector<long_index_t> in_right_pads_;
std::vector<index_t> filter_spatial_lengths_; std::vector<long_index_t> filter_spatial_lengths_;
std::vector<index_t> output_spatial_lengths_; std::vector<long_index_t> output_spatial_lengths_;
private: private:
void initOutputSpatialLengths() void initOutputSpatialLengths()
...@@ -76,7 +76,8 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -76,7 +76,8 @@ struct ReferenceImageToColumn : public device::BaseOperator
{ {
// XEff = (X - 1) * conv_dilation_w + 1; // XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1; // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1; const ck::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1;
output_spatial_lengths_.push_back( output_spatial_lengths_.push_back(
(input_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] + (input_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] +
...@@ -99,24 +100,24 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -99,24 +100,24 @@ struct ReferenceImageToColumn : public device::BaseOperator
throw std::runtime_error("wrong! inconsistent dimension"); throw std::runtime_error("wrong! inconsistent dimension");
} }
const index_t G = arg.input_.GetLengths()[0]; const long_index_t G = arg.input_.GetLengths()[0];
const index_t N = arg.input_.GetLengths()[1]; const long_index_t N = arg.input_.GetLengths()[1];
const index_t C = arg.input_.GetLengths()[2]; const long_index_t C = arg.input_.GetLengths()[2];
if constexpr(NDimSpatial == 1) if constexpr(NDimSpatial == 1)
{ {
const index_t Wo = arg.output_spatial_lengths_[0]; const long_index_t Wo = arg.output_spatial_lengths_[0];
auto func = [&](auto g, auto n, auto wo) { auto func = [&](auto g, auto n, auto wo) {
index_t row = n * Wo + wo; long_index_t row = n * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x)
{ {
auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) + auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
...@@ -135,26 +136,26 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -135,26 +136,26 @@ struct ReferenceImageToColumn : public device::BaseOperator
} }
else if constexpr(NDimSpatial == 2) else if constexpr(NDimSpatial == 2)
{ {
const index_t Ho = arg.output_spatial_lengths_[0]; const long_index_t Ho = arg.output_spatial_lengths_[0];
const index_t Wo = arg.output_spatial_lengths_[1]; const long_index_t Wo = arg.output_spatial_lengths_[1];
auto func = [&](auto g, auto n, auto ho, auto wo) { auto func = [&](auto g, auto n, auto ho, auto wo) {
index_t row = n * Ho * Wo + ho * Wo + wo; long_index_t row = n * Ho * Wo + ho * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y) for(long_index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y)
{ {
auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) + auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x)
{ {
auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) + auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) - static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]); static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(hi >= 0 && if(hi >= 0 &&
...@@ -178,31 +179,31 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -178,31 +179,31 @@ struct ReferenceImageToColumn : public device::BaseOperator
} }
else if constexpr(NDimSpatial == 3) else if constexpr(NDimSpatial == 3)
{ {
const index_t Do = arg.output_spatial_lengths_[0]; const long_index_t Do = arg.output_spatial_lengths_[0];
const index_t Ho = arg.output_spatial_lengths_[1]; const long_index_t Ho = arg.output_spatial_lengths_[1];
const index_t Wo = arg.output_spatial_lengths_[2]; const long_index_t Wo = arg.output_spatial_lengths_[2];
auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) { auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) {
index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo; long_index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
index_t column = 0; long_index_t column = 0;
for(index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z) for(long_index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z)
{ {
auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) + auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) - static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[0]); static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
for(index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y) for(long_index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y)
{ {
auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) + auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) - static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[1]); static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
for(index_t x = 0; x < arg.filter_spatial_lengths_[2]; ++x) for(long_index_t x = 0; x < arg.filter_spatial_lengths_[2]; ++x)
{ {
auto wi = auto wi =
static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) + static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) - static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
static_cast<ck::long_index_t>(arg.in_left_pads_[2]); static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
for(index_t c = 0; c < C; ++c) for(long_index_t c = 0; c < C; ++c)
{ {
if(di >= 0 && if(di >= 0 &&
ck::type_convert<std::size_t>(di) < ck::type_convert<std::size_t>(di) <
...@@ -259,15 +260,15 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -259,15 +260,15 @@ struct ReferenceImageToColumn : public device::BaseOperator
bool IsSupportedArgument(const Argument& arg) bool IsSupportedArgument(const Argument& arg)
{ {
const ck::index_t G = arg.input_.GetLengths()[0]; const ck::long_index_t G = arg.input_.GetLengths()[0];
const ck::index_t N = arg.input_.GetLengths()[1]; const ck::long_index_t N = arg.input_.GetLengths()[1];
const ck::index_t C = arg.input_.GetLengths()[2]; const ck::long_index_t C = arg.input_.GetLengths()[2];
const index_t NDoHoWo = const long_index_t NDoHoWo =
N * ck::accumulate_n<index_t>( N * ck::accumulate_n<long_index_t>(
arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>()); arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
const index_t CZYX = const long_index_t CZYX =
C * ck::accumulate_n<index_t>( C * ck::accumulate_n<long_index_t>(
arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>()); arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(G) && if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(G) &&
...@@ -291,11 +292,11 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -291,11 +292,11 @@ struct ReferenceImageToColumn : public device::BaseOperator
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(const Tensor<InDataType>& input,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> filter_spatial_lengths, std::vector<ck::long_index_t> filter_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::long_index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::long_index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads, std::vector<ck::long_index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads) std::vector<ck::long_index_t> input_right_pads)
{ {
return Argument{input, return Argument{input,
output, output,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_bf16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_f32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 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>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ a93d07c7
...@@ -17,6 +17,7 @@ ...@@ -17,6 +17,7 @@
#endif #endif
#ifdef CK_USE_XDL #ifdef CK_USE_XDL
#include "grouped_convolution_forward_xdl.inc" #include "grouped_convolution_forward_xdl.inc"
#include "grouped_convolution_forward_xdl_large_tensor.inc"
#include "grouped_convolution_forward_xdl_merged_groups.inc" #include "grouped_convolution_forward_xdl_merged_groups.inc"
#include "grouped_convolution_forward_comp_xdl.inc" #include "grouped_convolution_forward_comp_xdl.inc"
#include "grouped_convolution_forward_mem_inter_xdl.inc" #include "grouped_convolution_forward_mem_inter_xdl.inc"
...@@ -200,6 +201,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -200,6 +201,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>) is_same_v<BComputeType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances( add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs);
...@@ -215,6 +218,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -215,6 +218,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>) is_same_v<BComputeType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances( add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs);
...@@ -232,6 +237,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -232,6 +237,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>) is_same_v<BComputeType, ck::bhalf_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances( add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs);
...@@ -291,6 +298,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -291,6 +298,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>) is_same_v<BComputeType, float>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances( add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs);
...@@ -347,6 +356,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -347,6 +356,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>) is_same_v<BComputeType, half_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances( add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs);
...@@ -364,6 +375,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -364,6 +375,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>) is_same_v<BComputeType, ck::bhalf_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances( add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl_large_tensor.inc 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/utility/convolution_parameter.hpp
View file @ a93d07c7
// SPDX-License-Identifier: MIT // 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 #pragma once
...@@ -31,23 +31,35 @@ struct ConvParam ...@@ -31,23 +31,35 @@ struct ConvParam
const std::vector<ck::index_t>& left_pads, const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads); const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial_; ConvParam(ck::long_index_t n_dim,
ck::index_t G_; ck::long_index_t group_count,
ck::index_t N_; ck::long_index_t n_batch,
ck::index_t K_; ck::long_index_t n_out_channels,
ck::index_t C_; ck::long_index_t n_in_channels,
const std::vector<ck::long_index_t>& filters_len,
std::vector<ck::index_t> filter_spatial_lengths_; const std::vector<ck::long_index_t>& input_len,
std::vector<ck::index_t> input_spatial_lengths_; const std::vector<ck::long_index_t>& strides,
std::vector<ck::index_t> output_spatial_lengths_; const std::vector<ck::long_index_t>& dilations,
const std::vector<ck::long_index_t>& left_pads,
std::vector<ck::index_t> conv_filter_strides_; const std::vector<ck::long_index_t>& right_pads);
std::vector<ck::index_t> conv_filter_dilations_;
ck::long_index_t num_dim_spatial_;
std::vector<ck::index_t> input_left_pads_; ck::long_index_t G_;
std::vector<ck::index_t> input_right_pads_; ck::long_index_t N_;
ck::long_index_t K_;
std::vector<ck::index_t> GetOutputSpatialLengths() const; ck::long_index_t C_;
std::vector<ck::long_index_t> filter_spatial_lengths_;
std::vector<ck::long_index_t> input_spatial_lengths_;
std::vector<ck::long_index_t> output_spatial_lengths_;
std::vector<ck::long_index_t> conv_filter_strides_;
std::vector<ck::long_index_t> conv_filter_dilations_;
std::vector<ck::long_index_t> input_left_pads_;
std::vector<ck::long_index_t> input_right_pads_;
std::vector<ck::long_index_t> GetOutputSpatialLengths() const;
std::size_t GetFlops() const; std::size_t GetFlops() const;
......
library/include/ck/library/utility/host_tensor.hpp
View file @ a93d07c7
...@@ -96,9 +96,16 @@ struct HostTensorDescriptor ...@@ -96,9 +96,16 @@ struct HostTensorDescriptor
this->CalculateStrides(); this->CalculateStrides();
} }
HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens)
: mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename Lengths, template <typename Lengths,
typename = std::enable_if_t< typename = std::enable_if_t<
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t>>> std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> ||
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t>>>
HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end()) HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end())
{ {
this->CalculateStrides(); this->CalculateStrides();
...@@ -114,11 +121,19 @@ struct HostTensorDescriptor ...@@ -114,11 +121,19 @@ struct HostTensorDescriptor
{ {
} }
HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens,
const std::initializer_list<ck::long_index_t>& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
template <typename Lengths, template <typename Lengths,
typename Strides, typename Strides,
typename = std::enable_if_t< typename = std::enable_if_t<
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> && (std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>>> std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>) ||
(std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t> &&
std::is_convertible_v<ck::ranges::range_value_t<Strides>, ck::long_index_t>)>>
HostTensorDescriptor(const Lengths& lens, const Strides& strides) HostTensorDescriptor(const Lengths& lens, const Strides& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end()) : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{ {
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ a93d07c7
...@@ -64,6 +64,13 @@ function(add_instance_library INSTANCE_NAME) ...@@ -64,6 +64,13 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}") list(REMOVE_ITEM ARGN "${source}")
endif() endif()
endforeach() endforeach()
# Do not build mha instances if gfx94 targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "mha")
message("removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list #only continue if there are some source files left on the list
if(ARGN) if(ARGN)
set(INST_OBJ) set(INST_OBJ)
...@@ -77,6 +84,8 @@ function(add_instance_library INSTANCE_NAME) ...@@ -77,6 +84,8 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103)
elseif(ARGN MATCHES "_wmma") elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103)
endif() endif()
set(offload_targets) set(offload_targets)
foreach(target IN LISTS INST_TARGETS) foreach(target IN LISTS INST_TARGETS)
...@@ -86,7 +95,29 @@ function(add_instance_library INSTANCE_NAME) ...@@ -86,7 +95,29 @@ function(add_instance_library INSTANCE_NAME)
list(APPEND INST_OBJ ${source}) list(APPEND INST_OBJ ${source})
endforeach() endforeach()
add_library(${INSTANCE_NAME} OBJECT ${INST_OBJ}) add_library(${INSTANCE_NAME} OBJECT ${INST_OBJ})
# Allow comparing floating points directly in order to check sentinel values
if(${INSTANCE_NAME} STREQUAL "device_mha_instance")
if(NOT DEFINED FMHA_FWD_FAST_EXP2)
set(FMHA_FWD_FAST_EXP2 true)
endif()
if(FMHA_FWD_FAST_EXP2)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
endif()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
target_compile_options(device_mha_instance PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})
endif()
target_compile_features(${INSTANCE_NAME} PUBLIC) target_compile_features(${INSTANCE_NAME} PUBLIC)
# flags to compress the library
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding --offload-compress flag for ${INSTANCE_NAME}")
target_compile_options(${INSTANCE_NAME} PRIVATE --offload-compress)
endif()
set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(${INSTANCE_NAME}) clang_tidy_check(${INSTANCE_NAME})
set(result 0) set(result 0)
...@@ -286,20 +317,22 @@ if(CK_DEVICE_CONV_INSTANCES) ...@@ -286,20 +317,22 @@ if(CK_DEVICE_CONV_INSTANCES)
) )
endif() endif()
if(CK_DEVICE_MHA_INSTANCES) if(CK_DEVICE_MHA_INSTANCES)
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES}) set(gpu_list ${INST_TARGETS})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations) list(FILTER gpu_list INCLUDE REGEX "^gfx94")
target_compile_features(device_mha_operations PUBLIC) if(gpu_list)
set_target_properties(device_mha_operations PROPERTIES POSITION_INDEPENDENT_CODE ON) add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
target_include_directories(device_mha_operations PUBLIC add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/mha> target_compile_features(device_mha_operations PUBLIC)
) set_target_properties(device_mha_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
rocm_install(TARGETS device_mha_operations
EXPORT device_mha_operationsTargets) rocm_install(TARGETS device_mha_operations
rocm_install(EXPORT device_mha_operationsTargets EXPORT device_mha_operationsTargets)
FILE composable_kerneldevice_mha_operationsTargets.cmake rocm_install(EXPORT device_mha_operationsTargets
NAMESPACE composable_kernel:: FILE composable_kerneldevice_mha_operationsTargets.cmake
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel NAMESPACE composable_kernel::
) DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
endif()
endif() endif()
if(CK_DEVICE_CONTRACTION_INSTANCES) if(CK_DEVICE_CONTRACTION_INSTANCES)
add_library(device_contraction_operations STATIC ${CK_DEVICE_CONTRACTION_INSTANCES}) add_library(device_contraction_operations STATIC ${CK_DEVICE_CONTRACTION_INSTANCES})
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
View file @ a93d07c7
...@@ -9,6 +9,11 @@ add_instance_library(device_grouped_conv2d_fwd_instance ...@@ -9,6 +9,11 @@ add_instance_library(device_grouped_conv2d_fwd_instance
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
# large tensor
# NHWGC, GKYXC, NHWGK
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp
# merged groups # merged groups
# NHWGC, GKYXC, NHWGK # NHWGC, GKYXC, NHWGK
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_large_tensor_f16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_large_tensor_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
View file @ a93d07c7
...@@ -9,6 +9,10 @@ set(GROUPED_CONV3D_FWD ...@@ -9,6 +9,10 @@ set(GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp 0 → 100644
View file @ a93d07c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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