044a77352f
NOTE: this feature is not ready for user testing, and not yet enabled in daily builds. It is being merged now for easier collaboration on development. HIP is a heterogenous compute interface allowing C++ code to be executed on GPUs similar to CUDA. It is intended to bring back AMD GPU rendering support on Windows and Linux. https://github.com/ROCm-Developer-Tools/HIP. As of the time of writing, it should compile and run on Linux with existing HIP compilers and driver runtimes. Publicly available compilers and drivers for Windows will come later. See task T91571 for more details on the current status and work remaining to be done. Credits: Sayak Biswas (AMD) Arya Rafii (AMD) Brian Savery (AMD) Differential Revision: https://developer.blender.org/D12578
88 lines
2.4 KiB
C++
88 lines
2.4 KiB
C++
/*
|
|
* Copyright 2021 Blender Foundation
|
|
*
|
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
* you may not use this file except in compliance with the License.
|
|
* You may obtain a copy of the License at
|
|
*
|
|
* http://www.apache.org/licenses/LICENSE-2.0
|
|
*
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
* See the License for the specific language governing permissions and
|
|
* limitations under the License.
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
CCL_NAMESPACE_BEGIN
|
|
|
|
/* Parallel sum of array input_data with size n into output_sum.
|
|
*
|
|
* Adapted from "Optimizing Parallel Reduction in GPU", Mark Harris.
|
|
*
|
|
* This version adds multiple elements per thread sequentially. This reduces
|
|
* the overall cost of the algorithm while keeping the work complexity O(n) and
|
|
* the step complexity O(log n). (Brent's Theorem optimization) */
|
|
|
|
#ifdef __HIP__
|
|
# define GPU_PARALLEL_SUM_DEFAULT_BLOCK_SIZE 1024
|
|
#else
|
|
# define GPU_PARALLEL_SUM_DEFAULT_BLOCK_SIZE 512
|
|
#endif
|
|
|
|
template<uint blocksize, typename InputT, typename OutputT, typename ConvertOp>
|
|
__device__ void gpu_parallel_sum(
|
|
const InputT *input_data, const uint n, OutputT *output_sum, OutputT zero, ConvertOp convert)
|
|
{
|
|
extern ccl_gpu_shared OutputT shared_data[];
|
|
|
|
const uint tid = ccl_gpu_thread_idx_x;
|
|
const uint gridsize = blocksize * ccl_gpu_grid_dim_x();
|
|
|
|
OutputT sum = zero;
|
|
for (uint i = ccl_gpu_block_idx_x * blocksize + tid; i < n; i += gridsize) {
|
|
sum += convert(input_data[i]);
|
|
}
|
|
shared_data[tid] = sum;
|
|
|
|
ccl_gpu_syncthreads();
|
|
|
|
if (blocksize >= 512 && tid < 256) {
|
|
shared_data[tid] = sum = sum + shared_data[tid + 256];
|
|
}
|
|
|
|
ccl_gpu_syncthreads();
|
|
|
|
if (blocksize >= 256 && tid < 128) {
|
|
shared_data[tid] = sum = sum + shared_data[tid + 128];
|
|
}
|
|
|
|
ccl_gpu_syncthreads();
|
|
|
|
if (blocksize >= 128 && tid < 64) {
|
|
shared_data[tid] = sum = sum + shared_data[tid + 64];
|
|
}
|
|
|
|
ccl_gpu_syncthreads();
|
|
|
|
if (blocksize >= 64 && tid < 32) {
|
|
shared_data[tid] = sum = sum + shared_data[tid + 32];
|
|
}
|
|
|
|
ccl_gpu_syncthreads();
|
|
|
|
if (tid < 32) {
|
|
for (int offset = ccl_gpu_warp_size / 2; offset > 0; offset /= 2) {
|
|
sum += ccl_shfl_down_sync(0xFFFFFFFF, sum, offset);
|
|
}
|
|
}
|
|
|
|
if (tid == 0) {
|
|
output_sum[ccl_gpu_block_idx_x] = sum;
|
|
}
|
|
}
|
|
|
|
CCL_NAMESPACE_END
|