Skip to content

GitLab

Commit 289f15de authored by aska-0096's avatar aska-0096
Browse files

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

parents 9bd44685 d58b7f51
Hide whitespace changes
Inline Side-by-side
Showing with 251 additions and 210 deletions
profiler/src/profile_conv_fwd_bias_relu.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_conv_fwd_bias_relu_impl.hpp"
#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct ConvDataType
{
......@@ -32,11 +33,14 @@ enum struct ConvOutputLayout
NHWK, // 1
};
#define OP_NAME "conv_fwd_bias_relu"
#define OP_DESC "Convolution Forward+Bias+ReLU"
int profile_conv_fwd_bias_relu(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_fwd_bias_relu: ForwardConvolution+Bias+ReLu)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
......@@ -114,3 +118,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu);
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp"
#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct ConvDataType
{
......@@ -32,12 +33,14 @@ enum struct ConvOutputLayout
NHWK, // 1
};
#define OP_NAME "conv_fwd_bias_relu_add"
#define OP_DESC "Convolution Forward+Bias+ReLU+Add"
int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
{
if(argc != 25)
{
printf(
"arg1: tensor operation (conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLu+Add)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
......@@ -115,3 +118,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu_add);
profiler/src/profile_gemm.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_impl.hpp"
#include "profiler/profile_gemm_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
......@@ -24,9 +25,12 @@ enum struct GemmDataType
INT8_INT8_INT8, // 3
};
#define OP_NAME "gemm"
#define OP_DESC "GEMM"
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (gemm: GEMM)\n"
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
<< " 1: A[m, k] * B[n, k] = C[m, n];\n"
......@@ -184,3 +188,5 @@ int profile_gemm(int argc, char* argv[])
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm);
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ 289f15de
......@@ -6,7 +6,11 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_add_add_fastgelu_impl.hpp"
#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add_add_fastgelu"
#define OP_DESC "GEMM+Add+Add+FastGeLU"
int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{
......@@ -29,7 +33,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
......@@ -150,3 +154,5 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_add_fastgelu);
profiler/src/profile_gemm_bias_add_reduce.cpp
View file @ 289f15de
......@@ -6,7 +6,11 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_add_reduce_impl.hpp"
#include "profiler/profile_gemm_bias_add_reduce_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_bias_add_reduce"
#define OP_DESC "GEMM+Bias+Add+Reduce"
int profile_gemm_bias_add_reduce(int argc, char* argv[])
{
......@@ -26,7 +30,7 @@ int profile_gemm_bias_add_reduce(int argc, char* argv[])
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+bias+add+Reduce)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
......@@ -159,3 +163,5 @@ int profile_gemm_bias_add_reduce(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bias_add_reduce);
profiler/src/profile_gemm_bilinear.cpp
View file @ 289f15de
......@@ -6,7 +6,11 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bilinear_impl.hpp"
#include "profiler/profile_gemm_bilinear_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_bilinear"
#define OP_DESC "GEMM+Bilinear"
int profile_gemm_bilinear(int argc, char* argv[])
{
......@@ -29,7 +33,7 @@ int profile_gemm_bilinear(int argc, char* argv[])
if(argc != 17)
{
// clang-format off
printf("arg1: tensor operation (gemm_bilinear: GEMM+Bilinear)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
printf(" 1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
......@@ -144,3 +148,5 @@ int profile_gemm_bilinear(int argc, char* argv[])
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bilinear);
profiler/src/profile_gemm_reduce.cpp
View file @ 289f15de
......@@ -6,7 +6,11 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_reduce_impl.hpp"
#include "profiler/profile_gemm_reduce_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_reduce"
#define OP_DESC "GEMM+Reduce"
int profile_gemm_reduce(int argc, char* argv[])
{
......@@ -26,7 +30,7 @@ int profile_gemm_reduce(int argc, char* argv[])
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+Reduce)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
......@@ -146,3 +150,5 @@ int profile_gemm_reduce(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_reduce);
profiler/src/profile_gemm_splitk.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_splitk_impl.hpp"
#include "profiler/profile_gemm_splitk_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
......@@ -24,11 +25,14 @@ enum struct GemmDataType
INT8_INT8_INT8, // 3
};
#define OP_NAME "gemm_splitk"
#define OP_DESC "Split-K GEMM"
int profile_gemm_splitk(int argc, char* argv[])
{
if(argc != 15)
{
printf("arg1: tensor operation (gemm_splitk: Split-K GEMM)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
......@@ -146,3 +150,5 @@ int profile_gemm_splitk(int argc, char* argv[])
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_splitk);
profiler/src/profile_conv_bwd_weight.cpp → profiler/src/profile_grouped_conv_bwd_weight.cpp
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
#include "profiler_operation_registry.hpp"
namespace {
enum struct ConvLayout
{
NCHW_KCYX_NKHW, // 0
NHWC_KYXC_NHWK, // 1
GNCHW_GKCYX_GNKHW, // 0
GNHWC_GKYXC_GNHWK, // 1
};
enum struct ConvDataType
......@@ -23,26 +24,30 @@ enum struct ConvDataType
BF16_F32_BF16, // 2
};
#define OP_NAME "grouped_conv_bwd_weight"
#define OP_DESC "Grouped Convolution Backward Weight"
static void print_helper_msg()
{
std::cout
<< "arg1: tensor operation (conv_bwd_weight: Convolution Backward Weight\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight fp32, Output bf16)\n"
<< "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
<< " 1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, K]\n"
<< "arg4: verification (0: no, 1: yes)\n"
<< "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << " SplitK\n"
<< std::endl;
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight fp32, Output bf16)\n"
<< "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
"N, K, Ho, Wo]\n"
<< " 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
"N, Ho, Wo, K]\n"
<< "arg4: verification (0: no, 1: yes)\n"
<< "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << " SplitK\n"
<< std::endl;
}
} // namespace
int profile_conv_bwd_weight(int argc, char* argv[])
int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
// 8 for control, 1 for num_dim_spatial
if(argc < 9)
......@@ -75,17 +80,17 @@ int profile_conv_bwd_weight(int argc, char* argv[])
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using NWC = ck::tensor_layout::convolution::NWC;
using NHWC = ck::tensor_layout::convolution::NHWC;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
using GNWC = ck::tensor_layout::convolution::GNWC;
using GNHWC = ck::tensor_layout::convolution::GNHWC;
using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
using KXC = ck::tensor_layout::convolution::KXC;
using KYXC = ck::tensor_layout::convolution::KYXC;
using KZYXC = ck::tensor_layout::convolution::KZYXC;
using GKXC = ck::tensor_layout::convolution::GKXC;
using GKYXC = ck::tensor_layout::convolution::GKYXC;
using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
using NWK = ck::tensor_layout::convolution::NWK;
using NHWK = ck::tensor_layout::convolution::NHWK;
using NDHWK = ck::tensor_layout::convolution::NDHWK;
using GNWK = ck::tensor_layout::convolution::GNWK;
using GNHWK = ck::tensor_layout::convolution::GNHWK;
using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
constexpr auto I1 = ck::Number<1>{};
constexpr auto I2 = ck::Number<2>{};
......@@ -108,64 +113,64 @@ int profile_conv_bwd_weight(int argc, char* argv[])
using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
bool pass = ck::profiler::profile_conv_bwd_weight_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType>(
bool pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType>(
do_verification, init_method, do_log, time_kernel, params, split_k);
return pass ? 0 : 1;
};
if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, F32{}, BF16{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, F32{}, BF16{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, F32{}, BF16{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{});
}
}
......@@ -173,3 +178,5 @@ int profile_conv_bwd_weight(int argc, char* argv[])
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_weight);
profiler/src/profile_grouped_conv_fwd.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_grouped_conv_fwd_impl.hpp"
#include "profiler/profile_grouped_conv_fwd_impl.hpp"
#include "profiler_operation_registry.hpp"
namespace {
......@@ -24,11 +25,14 @@ enum struct ConvDataType
INT8_INT8_INT8, // 3
};
#define OP_NAME "grouped_conv_fwd"
#define OP_DESC "Grouped Convolution Forward"
static void print_helper_msg()
{
std::cout
// clang-format off
<< "arg1: tensor operation (grouped_conv_fwd: Grouped Convolution Forward)\n"
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight bf16, Output bf16\n"
......@@ -252,3 +256,5 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_fwd);
profiler/src/profile_grouped_gemm.cpp
View file @ 289f15de
......@@ -6,7 +6,8 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_grouped_gemm_impl.hpp"
#include "profiler/profile_grouped_gemm_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
......@@ -44,11 +45,14 @@ std::vector<int> argToIntArray(char* input)
return out;
}
#define OP_NAME "grouped_gemm"
#define OP_DESC "Grouped GEMM"
int profile_grouped_gemm(int argc, char* argv[])
{
if(!(argc == 14))
{
printf("arg1: tensor operation (grouped_gemm: Grouped GEMM)\n");
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
......@@ -161,3 +165,5 @@ int profile_grouped_gemm(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm);
profiler/src/profile_groupnorm.cpp
View file @ 289f15de
......@@ -5,8 +5,9 @@
#include <vector>
#include <unordered_map>
#include "profiler/include/data_type_enum.hpp"
#include "profiler/include/profile_groupnorm_impl.hpp"
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_groupnorm_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
......@@ -43,9 +44,12 @@ struct GroupnormArgParser
}
};
#define OP_NAME "groupnorm"
#define OP_DESC "Group Normalization"
void print_help_groupnorm()
{
std::cout << "arg1: tensor operation (groupnorm: Group normalization)\n"
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp16; 1: fp32)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
......@@ -104,3 +108,5 @@ int profile_groupnorm(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_groupnorm);
profiler/src/profile_layernorm.cpp
View file @ 289f15de
......@@ -5,8 +5,9 @@
#include <vector>
#include <unordered_map>
#include "profiler/include/data_type_enum.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_layernorm_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
......@@ -96,3 +97,5 @@ int profile_layernorm(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);
profiler/src/profile_reduce.cpp
View file @ 289f15de
......@@ -13,8 +13,9 @@
#include "ck/library/utility/host_common_util.hpp"
#include "profiler/include/profile_reduce_impl.hpp"
#include "profiler/include/data_type_enum.hpp"
#include "profiler/profile_reduce_impl.hpp"
#include "profiler/data_type_enum.hpp"
#include "profiler_operation_registry.hpp"
using namespace std;
......@@ -429,3 +430,5 @@ int profile_reduce(int argc, char* argv[])
return (0);
};
REGISTER_PROFILER_OPERATION("reduce", "Reduce", profile_reduce);
profiler/src/profile_softmax.cpp
View file @ 289f15de
......@@ -5,7 +5,8 @@
#include <vector>
#include <unordered_map>
#include "profiler/include/profile_softmax_impl.hpp"
#include "profiler/profile_softmax_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
using ck::profiler::SoftmaxDataType;
......@@ -164,3 +165,5 @@ int profile_softmax(int argc, char* argv[])
// profile_normalization(argc, argv);
// return 0;
// }
REGISTER_PROFILER_OPERATION("softmax", "Softmax", profile_softmax);
profiler/src/profiler.cpp
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstring>
#include <cstdlib>
#include <iostream>
int profile_gemm(int, char*[]);
int profile_gemm_splitk(int, char*[]);
int profile_gemm_bilinear(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_gemm(int, char*[]);
int profile_batched_gemm_add_relu_gemm_add(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_bwd_data(int, char*[]);
int profile_conv_bwd_weight(int, char*[]);
int profile_grouped_conv_fwd(int, char*[]);
int profile_softmax(int, char*[]);
int profile_layernorm(int, char*[]);
int profile_groupnorm(int, char*[]);
int profile_reduce(int, char*[]);
#include "profiler_operation_registry.hpp"
static void print_helper_message()
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_splitk: Split-K GEMM\n"
" gemm_bilinear: GEMM+Bilinear\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" gemm_reduce: GEMM+Reduce\n"
" gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
" batched_gemm: Batched GEMM\n"
" batched_gemm_gemm: Batched+GEMM+GEMM\n"
" batched_gemm_add_relu_gemm_add: Batched+GEMM+bias+gelu+GEMM+bias\n"
" batched_gemm_reduce: Batched GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: Convolution Forward\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_bwd_data: Convolution Backward Data\n"
" conv_bwd_weight: Convolution Backward Weight\n"
" grouped_conv_fwd: Grouped Convolution Forward\n"
" softmax: Softmax\n"
" reduce: Reduce\n");
// clang-format on
std::cout << "arg1: tensor operation " << ProfilerOperationRegistry::GetInstance() << std::endl;
}
int main(int argc, char* argv[])
......@@ -55,97 +16,15 @@ int main(int argc, char* argv[])
if(argc == 1)
{
print_helper_message();
return 0;
}
else if(strcmp(argv[1], "gemm") == 0)
{
return profile_gemm(argc, argv);
}
else if(strcmp(argv[1], "gemm_splitk") == 0)
{
return profile_gemm_splitk(argc, argv);
}
else if(strcmp(argv[1], "gemm_bilinear") == 0)
{
return profile_gemm_bilinear(argc, argv);
}
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
return profile_gemm_add_add_fastgelu(argc, argv);
}
else if(strcmp(argv[1], "gemm_reduce") == 0)
{
return profile_gemm_reduce(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_add_reduce") == 0)
else if(const auto operation = ProfilerOperationRegistry::GetInstance().Get(argv[1]);
operation.has_value())
{
return profile_gemm_bias_add_reduce(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm") == 0)
{
return profile_batched_gemm(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm_gemm") == 0)
{
return profile_batched_gemm_gemm(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm_add_relu_gemm_add") == 0)
{
return profile_batched_gemm_add_relu_gemm_add(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm_reduce") == 0)
{
return profile_batched_gemm_reduce(argc, argv);
}
else if(strcmp(argv[1], "grouped_gemm") == 0)
{
return profile_grouped_gemm(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd") == 0)
{
return profile_conv_fwd(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0)
{
return profile_conv_fwd_bias_relu(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd_bias_relu_add") == 0)
{
return profile_conv_fwd_bias_relu_add(argc, argv);
}
else if(strcmp(argv[1], "conv_bwd_data") == 0)
{
return profile_conv_bwd_data(argc, argv);
}
else if(strcmp(argv[1], "conv_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "grouped_conv_fwd") == 0)
{
return profile_grouped_conv_fwd(argc, argv);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "softmax") == 0)
{
return profile_softmax(argc, argv);
}
else if(strcmp(argv[1], "layernorm") == 0)
{
return profile_layernorm(argc, argv);
}
else if(strcmp(argv[1], "groupnorm") == 0)
{
return profile_groupnorm(argc, argv);
return (*operation)(argc, argv);
}
else
{
print_helper_message();
return 0;
std::cerr << "cannot find operation: " << argv[1] << std::endl;
return EXIT_FAILURE;
}
}
profiler/src/profiler_operation_registry.hpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <functional>
#include <iostream>
#include <iterator>
#include <map>
#include <optional>
#include <string_view>
#include <utility>
class ProfilerOperationRegistry final
{
ProfilerOperationRegistry() = default;
~ProfilerOperationRegistry() = default;
public:
using Operation = std::function<int(int, char*[])>;
private:
struct Entry final
{
explicit Entry(std::string_view description, Operation operation) noexcept
: description_(description), operation_(std::move(operation))
{
}
std::string_view description_;
Operation operation_;
};
std::map<std::string_view, Entry> entries_;
friend std::ostream& operator<<(std::ostream& stream, const ProfilerOperationRegistry& registry)
{
stream << "{\n";
for(auto& [name, entry] : registry.entries_)
{
stream << "\t" << name << ": " << entry.description_ << "\n";
}
stream << "}";
return stream;
}
public:
static ProfilerOperationRegistry& GetInstance()
{
static ProfilerOperationRegistry registry;
return registry;
}
std::optional<Operation> Get(std::string_view name) const
{
const auto found = entries_.find(name);
if(found == end(entries_))
{
return std::nullopt;
}
return (found->second).operation_;
}
bool Add(std::string_view name, std::string_view description, Operation operation)
{
return entries_
.emplace(std::piecewise_construct,
std::forward_as_tuple(name),
std::forward_as_tuple(description, std::move(operation)))
.second;
}
};
#define PP_CONCAT(x, y) PP_CONCAT_IMPL(x, y)
#define PP_CONCAT_IMPL(x, y) x##y
#define REGISTER_PROFILER_OPERATION(name, description, operation) \
static const bool PP_CONCAT(operation_registration_result_, __COUNTER__) = \
::ProfilerOperationRegistry::GetInstance().Add(name, description, operation)
test/CMakeLists.txt
View file @ 289f15de
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/
${PROJECT_SOURCE_DIR}/profiler/include
)
include(googletest)
......@@ -45,14 +46,15 @@ add_subdirectory(batched_gemm_softmax_gemm_permute)
add_subdirectory(grouped_gemm)
add_subdirectory(reduce)
add_subdirectory(convnd_fwd)
add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_convnd_fwd)
add_subdirectory(grouped_convnd_bwd_weight)
add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
add_subdirectory(normalization)
add_subdirectory(data_type)
add_subdirectory(elementwise_normalization)
add_subdirectory(batchnorm)
if(GPU_TARGETS MATCHES "gfx1100")
add_subdirectory(wmma_op)
endif()
\ No newline at end of file
endif()
test/batched_gemm/batched_gemm_bf16.cpp
View file @ 289f15de
......@@ -3,7 +3,7 @@
#include <iostream>
#include "profiler/include/profile_batched_gemm_impl.hpp"
#include "profiler/profile_batched_gemm_impl.hpp"
namespace {
using ADataType = ck::bhalf_t;
......
test/batched_gemm/batched_gemm_fp16.cpp
View file @ 289f15de
......@@ -3,7 +3,7 @@
#include <iostream>
#include "profiler/include/profile_batched_gemm_impl.hpp"
#include "profiler/profile_batched_gemm_impl.hpp"
namespace {
using ADataType = ck::half_t;
......
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