xgiovio/MultiPar
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fixed some bugs in OpenCL code for GPU

Browse Source
This commit is contained in:
Yutaka Sawada
2023-05-29 10:16:41 +09:00
committed by GitHub
parent d1de5c433f
commit e42e2b0710
3 changed files with 43 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
source/par2j/lib_opencl.c
View File

@@ -1,5 +1,5 @@
// lib_opencl.c
// Copyright : 2022-01-15 Yutaka Sawada
// Copyright : 2023-05-29 Yutaka Sawada
// License : GPL
#ifndef _WIN32_WINNT
@@ -317,7 +317,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
#ifdef DEBUG_OUTPUT
printf("MAX_COMPUTE_UNITS = %d\n", num_groups);
printf("MAX_WORK_GROUP_SIZE = %d\n", data_size);
printf("MAX_WORK_GROUP_SIZE = %zd\n", data_size);
printf("HOST_UNIFIED_MEMORY = %d\n", param_value);
#endif
// MAX_COMPUTE_UNITS * MAX_WORK_GROUP_SIZE で計算力を測る、外付けGPUなら値を倍にする
@@ -436,7 +436,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
// printf("4 KB cache (16-bytes * 256 work items), use if\n");
#endif
} else {
OpenCL_method = 0; // MMX用のコードは遅いので、キャッシュ最適化する必要が無い
OpenCL_method = 1; // MMX用のコードは遅いので、キャッシュ最適化する必要が無い
}
// work group 数が必要以上に多い場合は減らす
@@ -466,7 +466,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
#ifdef DEBUG_OUTPUT
data_size = (size_t)unit_size * count;
printf("src buf : %d KB (%d blocks), possible\n", data_size >> 10, count);
printf("src buf : %zd KB (%d blocks), possible\n", data_size >> 10, count);
#endif
// 出力先は1ブロック分だけあればいい
@@ -476,7 +476,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
if (ret != CL_SUCCESS)
return (ret << 8) | 13;
#ifdef DEBUG_OUTPUT
printf("dst buf : %d KB (%d Bytes), OK\n", data_size >> 10, data_size);
printf("dst buf : %zd KB (%zd Bytes), OK\n", data_size >> 10, data_size);
#endif
// factor は最大個数分 (src_max個)
@@ -485,7 +485,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
if (ret != CL_SUCCESS)
return (ret << 8) | 14;
#ifdef DEBUG_OUTPUT
printf("factor buf : %d Bytes (%d factors), OK\n", data_size, (*src_max));
printf("factor buf : %zd Bytes (%d factors), OK\n", data_size, (*src_max));
#endif
/*
@@ -539,25 +539,33 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
// Referred to "Embedding OpenCL Kernel Files in the Application on Windows"
res = FindResource(NULL, L"#1", L"RT_STRING"); // find the resource
if (res == NULL){
//printf("cannot find resource\n");
#ifdef DEBUG_OUTPUT
printf("cannot find resource\n");
#endif
return err;
}
glob = LoadResource(NULL, res); // load the resource.
if (glob == NULL){
//printf("cannot load resource\n");
#ifdef DEBUG_OUTPUT
printf("cannot load resource\n");
#endif
return err;
}
p_source = (char *)LockResource(glob); // lock the resource to get a char*
if (res == NULL){
//printf("cannot lock resource\n");
if (p_source == NULL){
#ifdef DEBUG_OUTPUT
printf("cannot lock resource\n");
#endif
return err;
}
data_size = SizeofResource(NULL, res);
if (data_size == 0){
//printf("cannot get size of resource\n");
#ifdef DEBUG_OUTPUT
printf("cannot get size of resource\n");
#endif
return err;
}
//printf("Source code length = %d characters\n", data_size);
//printf("OpenCL source code length = %zd characters\n", data_size);
// プログラムを作成する
program = fn_clCreateProgramWithSource(OpenCL_context, 1, (char **)&p_source, &data_size, &ret);
@@ -593,7 +601,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
ret = fn_clGetKernelWorkGroupInfo(OpenCL_kernel, NULL, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &data_size, NULL);
if ((ret == CL_SUCCESS) && (data_size < 256)){ // 最低でも 256以上は必要
#ifdef DEBUG_OUTPUT
printf("KERNEL_WORK_GROUP_SIZE = %d\n", data_size);
printf("KERNEL_WORK_GROUP_SIZE = %zd\n", data_size);
#endif
return (ret << 8) | 23;
}
@@ -622,7 +630,7 @@ int init_OpenCL(int unit_size, int *src_max, int *chunk_size)
#ifdef DEBUG_OUTPUT
// ワークアイテム数
printf("\nMax number of work items = %d (256 * %d)\n", OpenCL_group_num * 256, OpenCL_group_num);
printf("\nMax number of work items = %zd (256 * %zd)\n", OpenCL_group_num * 256, OpenCL_group_num);
#endif
return 0;

27
source/par2j/reedsolomon.c
View File

@@ -1,5 +1,5 @@
// reedsolomon.c
// Copyright : 2022-10-08 Yutaka Sawada
// Copyright : 2023-05-29 Yutaka Sawada
// License : GPL
#ifndef _UNICODE
@@ -28,6 +28,16 @@
#include "reedsolomon.h"
// GPU を使う最小データサイズ (MB 単位)
// GPU の起動には時間がかかるので、データが小さすぎると逆に遅くなる
#define GPU_DATA_LIMIT 512
// GPU を使う最小ブロックサイズとブロック数
// CPU と GPU で処理を割り振る為には、ある程度のブロック数を必要とする
#define GPU_BLOCK_SIZE_LIMIT 65536
#define GPU_SOURCE_COUNT_LIMIT 256
#define GPU_PARITY_COUNT_LIMIT 32
/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
// chunk がキャッシュに収まるようにすれば速くなる！ (Cache Blocking という最適化手法)
@@ -554,8 +564,9 @@ unsigned int time_total = GetTickCount();
#endif
// HDD なら 1-pass & Read some 方式を使う
// メモリー不足や SSD なら、Read all 方式でブロックを断片化させる
if ((OpenCL_method != 0) && (block_size >= 65536) && (source_num >= 256) && (parity_num >= 32) &&
((source_num + parity_num) * (__int64)block_size > 1048576 * 512)){
if ((OpenCL_method != 0) && (block_size >= GPU_BLOCK_SIZE_LIMIT) &&
(source_num >= GPU_SOURCE_COUNT_LIMIT) && (parity_num >= GPU_PARITY_COUNT_LIMIT) &&
((source_num + parity_num) * (__int64)block_size > 1048576 * GPU_DATA_LIMIT)){
// ブロック数が多いなら、ブロックごとにスレッドを割り当てる (GPU を使う)
err = -4; // 2-pass & GPU read all
} else {
@@ -636,8 +647,9 @@ unsigned int time_total = GetTickCount();
if (err == 0){
#endif
// メモリーが足りてる場合だけ 1-pass方式を使う
if ((OpenCL_method != 0) && (block_size >= 65536) && (source_num >= 256) && (parity_num >= 32) &&
((source_num + parity_num) * (__int64)block_size > 1048576 * 512)){
if ((OpenCL_method != 0) && (block_size >= GPU_BLOCK_SIZE_LIMIT) &&
(source_num >= GPU_SOURCE_COUNT_LIMIT) && (parity_num >= GPU_PARITY_COUNT_LIMIT) &&
((source_num + parity_num) * (__int64)block_size > 1048576 * GPU_DATA_LIMIT)){
err = -5; // 1-pass & GPU read some
} else {
err = -3; // 1-pass & Read some
@@ -764,8 +776,9 @@ time_matrix = GetTickCount() - time_matrix;
err = 0; // IO method : 0=Auto, -2=Read all, -3=Read some, -4=GPU all, -5=GPU some
if (err == 0){
#endif
if ((OpenCL_method != 0) && (block_size >= 65536) && (source_num >= 256) && (block_lost >= 32) &&
((source_num + block_lost) * (__int64)block_size > 1048576 * 512)){
if ((OpenCL_method != 0) && (block_size >= GPU_BLOCK_SIZE_LIMIT) &&
(source_num >= GPU_SOURCE_COUNT_LIMIT) && (block_lost >= GPU_PARITY_COUNT_LIMIT) &&
((source_num + block_lost) * (__int64)block_size > 1048576 * GPU_DATA_LIMIT)){
// ブロック数が多いなら、ブロックごとにスレッドを割り当てる (GPU を使う)
if (memory_use & 16){
err = -4; // SSD なら Read all 方式でブロックが断片化しても速い

2
source/par2j/source.cl
View File

@@ -14,7 +14,7 @@ void calc_table(__local uint *mtab, int id, int factor)
mtab[id + 256] = sum;
}
__kernel void method0(
__kernel void method1(
__global uint *src,
__global uint *dst,
__global ushort *factors,