leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

Support parallel split K mode for porfiling (#277)

Browse Source 
* Support parallel split K mode for porfiling

Signed-off-by: Peter Han <fujun.han@iluvatar.ai>

* Parallel Split K support

  1. find gemm kernel by preference key
  2. switch m n for redution kernel

Signed-off-by: Peter Han <fujun.han@iluvatar.ai>

* parallel splitk for fp16 gemm

* add one missing file

Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Fujun Han
2022-01-27 10:37:37 -05:00
committed by GitHub
co-authored by Haicheng Wu
parent c3353add63
commit 1e4703cbab
13 changed files with 332 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/library/src/reduction/init_reduction_operations.cu
+1 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
tools/library/src/reduction/reduction_device.cu
+4 -5
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
@@ -46,7 +46,7 @@ void initialize_reduce_add_linear_combination_f32_f32_f16(Manifest &manifest) {
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombination<
ElementOutput,
128 / cutlass::sizeof_bits<ElementOutput>::value,
128 / cutlass::sizeof_bits<ElementWorkspace>::value,
ElementAccumulator,
ElementCompute
>;
@@ -81,7 +81,7 @@ void initialize_reduce_add_linear_combination_f32_f32_f32(Manifest &manifest) {
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombination<
ElementOutput,
128 / cutlass::sizeof_bits<ElementOutput>::value,
128 / cutlass::sizeof_bits<ElementWorkspace>::value,
ElementAccumulator,
ElementCompute
>;
@@ -115,7 +115,7 @@ void initialize_reduce_add_linear_combination_cf32_cf32_cf32(Manifest &manifest)
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombination<
ElementOutput,
128 / cutlass::sizeof_bits<ElementOutput>::value,
128 / cutlass::sizeof_bits<ElementWorkspace>::value,
ElementAccumulator,
ElementCompute
>;
@@ -140,6 +140,5 @@ void initialize_reduce_add_linear_combination_cf32_cf32_cf32(Manifest &manifest)
));
}
}
}
tools/library/src/reduction/reduction_operation.h
+3 -1
View File
@@ -30,6 +30,7 @@
#include <iostream>
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/thread/linear_combination.h"
#include "cutlass/epilogue/thread/linear_combination_clamp.h"
#include "cutlass/reduction/thread/reduction_operators.h"
#include "cutlass/reduction/device/reduce_split_k.h"
@@ -180,7 +181,8 @@ public:
/// Gets the device-side workspace
virtual uint64_t get_device_workspace_size(
void const *configuration_ptr) const {
void const *configuration_ptr,
void const *arguments_ptr = nullptr) const {
OperatorArguments args;