leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

CUTLASS 3.7 (#2045)

Browse Source 
* CUTLASS 3.7

* clean up changelog

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2025-01-18 09:53:07 -05:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent 902dff3663
commit b78588d163
2030 changed files with 8947 additions and 3475 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/profiler/src/conv2d_operation_profiler.cu
+7 -87
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* Copyright (c) 2017 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
@@ -1281,8 +1281,6 @@ Status Conv2dOperationProfiler::profile_cutlass_(
void *host_workspace,
void *device_workspace) {
GpuTimer timer;
// initialize conv2d underlying operation to handle parallel reduction
library::Operation const* underlying_operation = operation;
@@ -1294,23 +1292,9 @@ Status Conv2dOperationProfiler::profile_cutlass_(
}
}
//
// Optional sleep to limit power consumption and thermals
//
sleep(options.profiling.sleep_duration);
//
// Warmup loop
//
Status status;
for (int iteration = 0; iteration < options.profiling.warmup_iterations; ++iteration) {
auto func = [&](cudaStream_t, int iteration) {
// Setup rotating workspace
int workspace_idx = options.profiling.warmup_iterations + iteration;
int problem_idx = (workspace_idx % conv_workspace_.problem_count);
int problem_idx = iteration % conv_workspace_.problem_count;
conv_arguments->A = conv_workspace_.A->batch_data(problem_idx);
conv_arguments->B = conv_workspace_.B->batch_data(problem_idx);
@@ -1328,7 +1312,7 @@ Status Conv2dOperationProfiler::profile_cutlass_(
}
// Run underlying conv2d operation
status = underlying_operation->run(
Status status = underlying_operation->run(
arguments,
host_workspace,
device_workspace);
@@ -1345,74 +1329,10 @@ Status Conv2dOperationProfiler::profile_cutlass_(
if (status != Status::kSuccess) {
return status;
}
}
return status;
};
//
// Initialize GPU timer
//
timer.start();
//
// Profiling loop
//
int Iterations = options.profiling.iterations;
int iteration = 0;
for (; iteration < Iterations; ++iteration) {
// Setup rotating workspace
int problem_idx = (iteration % conv_workspace_.problem_count);
conv_arguments->A = conv_workspace_.A->batch_data(problem_idx);
conv_arguments->B = conv_workspace_.B->batch_data(problem_idx);
conv_arguments->C = conv_workspace_.C->batch_data(problem_idx);
conv_arguments->D = conv_workspace_.Computed->batch_data(problem_idx);
if (conv_workspace_.configuration.split_k_mode == conv::SplitKMode::kParallel) {
// update library::ConvArguments for parallel split-k reduction
conv_arguments->D = conv_workspace_.device_workspace.data();
/// initialize library::ReductionArguments
conv_workspace_.reduction_arguments.workspace = conv_workspace_.device_workspace.data();
conv_workspace_.reduction_arguments.source = conv_workspace_.C->batch_data(problem_idx);
conv_workspace_.reduction_arguments.destination = conv_workspace_.Computed->batch_data(problem_idx);
}
// Run underlying conv2d operation
status = underlying_operation->run(
arguments,
host_workspace,
device_workspace);
// Run parallel reduction kernel for parallel split_k_mode
if (conv_workspace_.configuration.split_k_mode == conv::SplitKMode::kParallel) {
status = reduction_op_->run(
&conv_workspace_.reduction_arguments,
conv_workspace_.reduction_host_workspace.data(),
nullptr);
}
if (status != Status::kSuccess) {
return status;
}
}
//
// Wait for completion
//
timer.stop_and_wait();
//
// Update performance result
//
result.runtime = timer.duration(iteration);
return status;
return profile_kernel_(result, options, func);
}
/////////////////////////////////////////////////////////////////////////////////////////////////