Skip to content

GitLab

Unverified Commit 08d5c02c authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

OCP FP8 support for gfx12. (#1710) 

* (2/5) bilinear gemm pass, perf bug: skip a lds has lower performance than skip b lds

* (3/5) batched gemm pass, perf bug: skip a lds has lower performance than skip b lds

* (4/5) grouped conv pass

* (5/5) attention pass, todo: debug lds perf bug

* AIT Attention API refactor (#8)

* sanity pass

* sanity pass 2

* confirm significant performance regression.

* turn on all instances

* turn off instance format

* Fix bug & tunning & format

* DML meta, self_attn+cross_attn

* sanity pass

* remove useless flag

* update tile and problem size used in AIT attention

* bug fix in grouped conv supporting check

* deprecate inline asm wmma

* Bug fix: double lds skip

* clang-format

* Fix errors in
1. example, fmha
2. gridwise pipeline
3. deviceop, fmha, change some containers from vector to array

* part2 of previous commit

* clang format

* API fix of gridwisegemmpipeline

* separate array base and vector base attention...
parent 50ee4267
Hide whitespace changes
Inline Side-by-side
Showing with 62 additions and 42 deletions
CMakeLists.txt
View file @ 08d5c02c
......@@ -185,13 +185,22 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx9")
add_definitions(-DCK_USE_XDL)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx94")
message("Enabling FP8 gemms in ckProfiler")
message("Enabling FP8 gemms on native architectures")
add_definitions(-DCK_USE_GFX94)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
message("Enabling WMMA instances")
add_definitions(-DCK_USE_WMMA)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx12")
add_definitions(-DCK_USE_OCP_FP8)
set(CK_USE_OCP_FP8 "ON")
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_USE_FNUZ_FP8)
set(CK_USE_FNUZ_FP8 "ON")
endif()
option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF)
if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908"))
add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH)
......
client_example/CMakeLists.txt
View file @ 08d5c02c
......@@ -56,6 +56,14 @@ if (GPU_TARGETS)
add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON")
endif()
if (GPU_TARGETS MATCHES "gfx12")
add_definitions(-DCK_USE_OCP_FP8)
set(CK_USE_OCP_FP8 "ON")
endif()
if (GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx94")
add_definitions(-DCK_USE_FNUZ_FP8)
set(CK_USE_FNUZ_FP8 "ON")
endif()
else()
add_definitions(-DCK_USE_WMMA -DCK_USE_XDL)
set(CK_USE_XDL "ON")
......
example/01_gemm/common.hpp
View file @ 08d5c02c
......@@ -76,7 +76,7 @@ struct ProblemSizeSplitK final
struct ExecutionConfig final
{
// 0 - no verification, 1 - CPU, 2 - GPU, 3 - CPU + GPU
int do_verification = 3;
int do_verification = 1;
int init_method = 2;
bool time_kernel = false;
};
......
example/01_gemm/run_gemm_example.inc
View file @ 08d5c02c
......@@ -143,8 +143,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
switch(config.init_method)
{
case 0:
ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
ck::utils::FillConstant<ADataType>{ck::type_convert<ADataType>(1.f)}(a_m_k);
ck::utils::FillConstant<BDataType>{ck::type_convert<BDataType>(1.f)}(b_k_n);
break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
......
example/15_grouped_gemm/grouped_gemm_multiple_d_splitk_xdl_fp16.cpp
View file @ 08d5c02c
......@@ -186,15 +186,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
for(int j = 0; j < NumDMatrices; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
for(int j = 0; j < NumDMatrices; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<DDataType, 0>{});
}
}
}
......
example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp
View file @ 08d5c02c
......@@ -190,15 +190,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
for(int j = 0; j < NumDs; ++j)
{
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<DDataType, 0>{});
}
}
}
......
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
View file @ 08d5c02c
......@@ -167,11 +167,11 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<D0DataType, 1>{});
}
using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<1>;
......
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
View file @ 08d5c02c
......@@ -157,8 +157,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
......
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16_fp8.cpp
View file @ 08d5c02c
......@@ -158,8 +158,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
......
example/15_grouped_gemm/run_grouped_gemm_example.inc
View file @ 08d5c02c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
struct ProblemSize final
......@@ -124,8 +127,8 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
}
......
example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
View file @ 08d5c02c
// 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.
#include <iostream>
#include <numeric>
......@@ -175,8 +175,8 @@ int main(int argc, char* argv[])
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
c0_n_bias.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});
......
example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
View file @ 08d5c02c
// 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
......@@ -150,7 +150,7 @@ bool run_batched_gemm_gemm_example(int argc, char* argv[])
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -157,7 +157,7 @@ int run(int argc, char* argv[])
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -118,7 +118,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -153,7 +153,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc
View file @ 08d5c02c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
......@@ -178,7 +178,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -152,7 +152,7 @@ int run(int argc, char* argv[])
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -156,7 +156,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc
View file @ 08d5c02c
// 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.
int run(int argc, char* argv[])
{
......@@ -156,7 +156,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc
View file @ 08d5c02c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
......@@ -173,7 +173,7 @@ int run(int argc, char* argv[])
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
......
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