mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Rotating buffer PR CI fix (#2257)
* Revert "Revert "[CK_tile] Add rotating buffer feature for universal gemm (#2200)" (#2256)"
This reverts commit bbdaf79a52.
* fix regression
This commit is contained in:
@@ -1,5 +1,5 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
@@ -11,6 +11,13 @@
|
||||
#include <cstddef>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
#define LOW_CU_PROCESSORS 80
|
||||
#define HIGH_CU_PROCESSORS 228
|
||||
#define OPTIMAL_LATENCY_LOW_CU_PROCESSORS 0.005
|
||||
#define OPTIMAL_LATENCY_HIGH_CU_PROCESSORS 0.0015
|
||||
#define OPTIMAL_LATENCY_SAFE_MARGIN 0.01
|
||||
|
||||
template <int MaxThreadPerBlock, int MinBlockPerCu, typename Kernel, typename... Args>
|
||||
#if CK_TILE_USE_LAUNCH_BOUNDS
|
||||
__launch_bounds__(MaxThreadPerBlock, MinBlockPerCu)
|
||||
@@ -81,6 +88,8 @@ CK_TILE_HOST void launch_and_check(const stream_config& sc, Callables&&... calla
|
||||
template <typename... Callables>
|
||||
CK_TILE_HOST float launch_kernel(const stream_config& s, Callables&&... callables)
|
||||
{
|
||||
static_assert(sizeof...(callables) > 0, "At least one callable is required!");
|
||||
|
||||
if(!s.time_kernel_)
|
||||
{
|
||||
launch_and_check(s, std::forward<Callables>(callables)...);
|
||||
@@ -88,7 +97,7 @@ CK_TILE_HOST float launch_kernel(const stream_config& s, Callables&&... callable
|
||||
}
|
||||
|
||||
auto time_launches = [&](auto timer) {
|
||||
// warmup
|
||||
// Warmup
|
||||
for(int i = 0; i < s.cold_niters_; i++)
|
||||
{
|
||||
launch_and_check(s, std::forward<Callables>(callables)...);
|
||||
@@ -114,4 +123,53 @@ CK_TILE_HOST float launch_kernel(const stream_config& s, Callables&&... callable
|
||||
}
|
||||
}
|
||||
|
||||
template <typename PreprocessFunc, typename... Callables>
|
||||
CK_TILE_HOST float launch_kernel_preprocess(const stream_config& s,
|
||||
PreprocessFunc preprocess,
|
||||
Callables&&... callables)
|
||||
{
|
||||
static_assert(sizeof...(callables) > 0, "At least one callable is required!");
|
||||
|
||||
if(!s.time_kernel_)
|
||||
{
|
||||
preprocess();
|
||||
launch_and_check(s, std::forward<Callables>(callables)...);
|
||||
return 0;
|
||||
}
|
||||
|
||||
auto time_launches = [&](auto timer) {
|
||||
// Warmup
|
||||
for(int i = 0; i < s.cold_niters_; i++)
|
||||
{
|
||||
launch_and_check(s, std::forward<Callables>(callables)...);
|
||||
}
|
||||
|
||||
timer.start(s.stream_id_);
|
||||
for(int i = 0; i < s.nrepeat_; i++)
|
||||
{
|
||||
preprocess();
|
||||
launch_and_check(s, std::forward<Callables>(callables)...);
|
||||
}
|
||||
timer.stop(s.stream_id_);
|
||||
|
||||
hipDeviceProp_t deviceProps;
|
||||
HIP_CHECK_ERROR(hipGetDeviceProperties(&deviceProps, 0));
|
||||
|
||||
float preprocess_offset = (deviceProps.multiProcessorCount >= HIGH_CU_PROCESSORS)
|
||||
? OPTIMAL_LATENCY_HIGH_CU_PROCESSORS
|
||||
: (deviceProps.multiProcessorCount == LOW_CU_PROCESSORS)
|
||||
? OPTIMAL_LATENCY_LOW_CU_PROCESSORS
|
||||
: OPTIMAL_LATENCY_SAFE_MARGIN;
|
||||
return (timer.duration() - preprocess_offset * s.nrepeat_) / s.nrepeat_;
|
||||
};
|
||||
|
||||
if(s.is_gpu_timer_)
|
||||
{
|
||||
return time_launches(gpu_timer{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return time_launches(cpu_timer{});
|
||||
}
|
||||
}
|
||||
} // namespace ck_tile
|
||||
|
||||
Reference in New Issue
Block a user