shijin/cuElim
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

使用命名空间进行划分

Browse Source
This commit is contained in:
shijin 2024-09-14 16:15:13 +08:00
parent 2375705792
commit e73b158a37
15 changed files with 826 additions and 822 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
benchmark/bench_gf256_mul.cu
View File

@ -1,6 +1,8 @@
#include <benchmark/benchmark.h>
#include "cuelim.cuh"
using namespace gf256;
template <MatGF256 (*GpuFunc)(const MatGF256 &, const MatGF256 &, const GF256 &)>
void bench_gf256_mul(benchmark::State &state)
{

2
benchmark/bench_gfp_mul.cu
View File

@ -1,6 +1,8 @@
#include <benchmark/benchmark.h>
#include "test_header.cuh"
using namespace gfp;
static void bench_gfp(benchmark::State &state)
{
uint_fast32_t seed = 41921095;

15
include/gf256/gf256_elim.cuh
View File

@ -3,6 +3,8 @@
#include "gf256_mat.cuh"
namespace gf256
{
void MatGF256::cpu_swap_row(size_t r1, size_t r2)
{
if (r1 == r2)
@ -19,7 +21,7 @@ void MatGF256::cpu_swap_row(size_t r1, size_t r2)
}
}
size_t gf256_cpu_elim_base(base_t *base_col, base_t base_col_len, size_t st_r, size_t w, vector<size_t> &p_col, vector<size_t> &p_row, const GF256 &gf)
size_t cpu_elim_base(base_t *base_col, base_t base_col_len, size_t st_r, size_t w, vector<size_t> &p_col, vector<size_t> &p_row, const GF256 &gf)
{
size_t rank = 0;
size_t pivot[gf256_num];
@ -58,7 +60,7 @@ size_t gf256_cpu_elim_base(base_t *base_col, base_t base_col_len, size_t st_r, s
return rank;
}
__global__ void gf256_gpu_mksrc_kernel(base_t *src, size_t s_rowstride, base_t *spL, size_t src_rank, size_t width)
__global__ void gpu_mksrc_kernel(base_t *src, size_t s_rowstride, base_t *spL, size_t src_rank, size_t width)
{
size_t w = blockIdx.x * blockDim.x + threadIdx.x;
if (w >= width)
@ -92,7 +94,7 @@ __global__ void gf256_gpu_mksrc_kernel(base_t *src, size_t s_rowstride, base_t *
}
}
__global__ void gf256_gpu_elim_kernel(base_t *idx, base_t *tb, size_t tb_rowstride, base_t *data, size_t rowstride, size_t rank, base_t pivot_base, size_t st_skip, size_t width, size_t nrows)
__global__ void gpu_elim_kernel(base_t *idx, base_t *tb, size_t tb_rowstride, base_t *data, size_t rowstride, size_t rank, base_t pivot_base, size_t st_skip, size_t width, size_t nrows)
{
size_t w = blockIdx.x * blockDim.x + threadIdx.x;
size_t r = blockIdx.y * blockDim.y + threadIdx.y;
@ -131,7 +133,7 @@ __host__ ElimResult MatGF256::gpu_elim(const GF256 &gf)
{
CUDA_CHECK(cudaMemcpy2D(base_col + rank, sizeof(base_t), at_base(rank, w), rowstride * sizeof(base_t), sizeof(base_t), nrows - rank, cudaMemcpyDefault));
size_t src_rank = gf256_cpu_elim_base(base_col + rank, nrows - rank, rank, w, p_col, p_row, gf);
size_t src_rank = cpu_elim_base(base_col + rank, nrows - rank, rank, w, p_col, p_row, gf);
if (src_rank == 0)
{
@ -162,7 +164,7 @@ __host__ ElimResult MatGF256::gpu_elim(const GF256 &gf)
dim3 block_src(THREAD_X);
dim3 grid_src((width - w - 1) / block_src.x + 1);
gf256_gpu_mksrc_kernel<<<grid_src, block_src>>>(at_base(rank, w), rowstride, spL, src_rank, width);
gpu_mksrc_kernel<<<grid_src, block_src>>>(at_base(rank, w), rowstride, spL, src_rank, width);
cudaDeviceSynchronize();
dim3 block_tb(THREAD_X, THREAD_Y);
@ -174,7 +176,7 @@ __host__ ElimResult MatGF256::gpu_elim(const GF256 &gf)
dim3 block(THREAD_X, THREAD_Y);
dim3 grid((width - w - 1) / block.x + 1, (nrows - 1) / block.y + 1);
gf256_gpu_elim_kernel<<<grid, block>>>(idx, tb.data, tb.rowstride, at_base(0, w), rowstride, src_rank, pivot_base, rank, width - w, nrows);
gpu_elim_kernel<<<grid, block>>>(idx, tb.data, tb.rowstride, at_base(0, w), rowstride, src_rank, pivot_base, rank, width - w, nrows);
cudaDeviceSynchronize();
rank += src_rank;
@ -188,5 +190,6 @@ __host__ ElimResult MatGF256::gpu_elim(const GF256 &gf)
cudaFree(idx);
return {rank, p_col, p_row};
}
}
#endif

4
include/gf256/gf256_header.cuh
View File

@ -4,6 +4,8 @@
#include "../header.cuh"
#include <set>
namespace gf256
{
using gf256_t = uint8_t;
static const size_t gf256_len = sizeof(gf256_t) * 8;
@ -188,5 +190,5 @@ ostream &operator<<(ostream &out, const GF256 &gf)
}
return out;
}
}
#endif

3
include/gf256/gf256_mat.cuh
View File

@ -7,6 +7,8 @@
#include <vector>
#include <algorithm>
namespace gf256
{
struct ElimResult
{
size_t rank;
@ -222,5 +224,6 @@ ostream &operator<<(ostream &out, const MatGF256 &m)
}
return out;
}
}
#endif

7
include/gf256/gf256_mul.cuh
View File

@ -3,7 +3,9 @@
#include "gf256_mat.cuh"
__global__ void gf256_gpu_addmul_kernel(base_t *a, size_t a_rowstride, base_t *tb, size_t tb_rowstride, base_t *c, size_t c_rowstride, size_t tb_num, size_t width, size_t nrows)
namespace gf256
{
__global__ void gpu_addmul_kernel(base_t *a, size_t a_rowstride, base_t *tb, size_t tb_rowstride, base_t *c, size_t c_rowstride, size_t tb_num, size_t width, size_t nrows)
{
size_t w = blockIdx.x * blockDim.x + threadIdx.x;
size_t r = blockIdx.y * blockDim.y + threadIdx.y;
@ -40,7 +42,7 @@ __host__ void MatGF256::gpu_addmul(const MatGF256 &a, const MatGF256 &b, const G
dim3 block(THREAD_X, THREAD_Y);
dim3 grid((b.width - 1) / block.x + 1, (nrows - 1) / block.y + 1);
gf256_gpu_addmul_kernel<<<grid, block>>>(a.at_base(0, w), a.rowstride, tb.data, tb.rowstride, data, rowstride, tb_num, width, nrows);
gpu_addmul_kernel<<<grid, block>>>(a.at_base(0, w), a.rowstride, tb.data, tb.rowstride, data, rowstride, tb_num, width, nrows);
cudaDeviceSynchronize();
}
}
@ -52,5 +54,6 @@ __host__ MatGF256 gpu_mul(const MatGF256 &a, const MatGF256 &b, const GF256 &gf)
c.gpu_addmul(a, b, gf);
return c;
}
}
#endif

14
include/gfp/gfp_header.cuh
View File

@ -3,25 +3,25 @@
#include "../header.cuh"
namespace gfp
{
using gfp_t = uint32_t;
#define gfp_bits 32
static_assert(sizeof(gfp_t) * 8 == gfp_bits);
static const gfp_t gfp = 65521;
static const gfp_t gfprime = 65521;
static const gfp_t gfp_zero = (gfp_t)0;
static const gfp_t gfp_one = (gfp_t)1;
static const gfp_t gfp_fullmask = (gfp_t)0xFF'FF;
__managed__ gfp_t gfp_inv_table[gfp];
__managed__ gfp_t gfp_inv_table[gfprime];
void init_inv_table()
{
gfp_inv_table[0] = 0;
gfp_inv_table[1] = 1;
for (int i = 2; i < gfp; ++i)
gfp_inv_table[i] = (gfp - gfp / i) * gfp_inv_table[gfp % i] % gfp;
for (int i = 2; i < gfprime; ++i)
gfp_inv_table[i] = (gfprime - gfprime / i) * gfp_inv_table[gfprime % i] % gfprime;
}
}
#endif

15
include/gfp/gfp_mat.cuh
View File

@ -7,6 +7,8 @@
#include <vector>
#include <algorithm>
namespace gfp
{
class MatGFP
{
public:
@ -92,7 +94,7 @@ public:
{
for (size_t w = 0; w < width; w++)
{
*at_base(r, w) = d(e) % gfp;
*at_base(r, w) = d(e) % gfprime;
}
}
}
@ -171,7 +173,7 @@ public:
{
for (size_t w = 0; w < width; w++)
{
*at_base(r, w) = (*at_base(r, w) + *m.at_base(r, w)) % gfp;
*at_base(r, w) = (*at_base(r, w) + *m.at_base(r, w)) % gfprime;
}
}
}
@ -192,9 +194,9 @@ public:
{
for (size_t i = 0; i < a.ncols; i++)
{
*at_base(r, w) += (*a.at_base(r, i) * *b.at_base(i, w)) % gfp;
*at_base(r, w) += (*a.at_base(r, i) * *b.at_base(i, w)) % gfprime;
}
*at_base(r, w) %= gfp;
*at_base(r, w) %= gfprime;
}
}
}
@ -228,15 +230,12 @@ ostream &operator<<(ostream &out, const MatGFP &m)
{
for (size_t w = 0; w < m.width; w++)
{
#if gfp_bits == 64
printf("%05lu ", *m.at_base(r, w));
#else
printf("%05u ", *m.at_base(r, w));
#endif
}
printf("\n");
}
return out;
}
}
#endif

25
include/gfp/gfp_mul.cuh
View File

@ -3,12 +3,15 @@
#include "gfp_mat.cuh"
namespace gfp
{
static const int BlockRow = 128, BlockCol = 128; // 每个block处理c矩阵的一个子块
static const int StepSize = 8; // block中一个循环处理的A矩阵的列数（B矩阵的行数）
static_assert(BlockCol % THREAD_X == 0 && BlockRow % THREAD_Y == 0);
__global__ void gfp_gpu_mul_kernel(gfp_t *__restrict__ a, const size_t a_rs, gfp_t *__restrict__ b, const size_t b_rs, gfp_t *__restrict__ c, const size_t c_rs, const size_t nrows, const size_t ncols, const size_t nsteps)
__global__ void gpu_mul_kernel(gfp_t *__restrict__ a, const size_t a_rs, gfp_t *__restrict__ b, const size_t b_rs, gfp_t *__restrict__ c, const size_t c_rs, const size_t nrows, const size_t ncols, const size_t nsteps)
{
const unsigned int bx = blockIdx.x;
@ -17,13 +20,8 @@ __global__ void gfp_gpu_mul_kernel(gfp_t *__restrict__ a, const size_t a_rs, gfp
const unsigned int ty = threadIdx.y;
const unsigned int tid = ty * blockDim.x + tx;
#if gfp_bits == 64
__shared__ alignas(8) gfp_t s_a[StepSize][BlockRow];
__shared__ alignas(8) gfp_t s_b[StepSize][BlockCol];
#else
__shared__ gfp_t s_a[StepSize][BlockRow];
__shared__ gfp_t s_b[StepSize][BlockCol];
#endif
gfp_t tmp_c[BlockRow / THREAD_Y][BlockCol / THREAD_X] = {0};
@ -45,27 +43,21 @@ __global__ void gfp_gpu_mul_kernel(gfp_t *__restrict__ a, const size_t a_rs, gfp
{
for (int i = 0; i < BlockCol / THREAD_X; i++)
{
#if gfp_bits == 64
tmp_c[j][i] += (s_a[k][j * THREAD_Y + ty] * s_b[k][i * THREAD_X + tx]);
#else
tmp_c[j][i] += (s_a[k][j * THREAD_Y + ty] * s_b[k][i * THREAD_X + tx]) % gfp;
#endif
tmp_c[j][i] += (s_a[k][j * THREAD_Y + ty] * s_b[k][i * THREAD_X + tx]) % gfprime;
}
}
}
__syncthreads();
#if gfp_bits != 64
if (s & gfp_fullmask == gfp_fullmask)
{
for (int j = 0; j < BlockRow / THREAD_Y; j++)
{
for (int i = 0; i < BlockCol / THREAD_X; i++)
{
tmp_c[j][i] %= gfp;
tmp_c[j][i] %= gfprime;
}
}
}
#endif
}
for (int j = 0; j < BlockRow / THREAD_Y; j++)
{
@ -73,7 +65,7 @@ __global__ void gfp_gpu_mul_kernel(gfp_t *__restrict__ a, const size_t a_rs, gfp
{
if (by * BlockRow + j * THREAD_Y + ty < nrows && bx * BlockCol + i * THREAD_X + tx < ncols)
{
*at_base(c, c_rs, by * BlockRow + j * THREAD_Y + ty, bx * BlockCol + i * THREAD_X + tx) = tmp_c[j][i] % gfp;
*at_base(c, c_rs, by * BlockRow + j * THREAD_Y + ty, bx * BlockCol + i * THREAD_X + tx) = tmp_c[j][i] % gfprime;
}
}
}
@ -85,8 +77,9 @@ __host__ void MatGFP::gpu_mul(const MatGFP &a, const MatGFP &b)
dim3 block(THREAD_X, THREAD_Y);
dim3 grid((width - 1) / block.x + 1, (nrows - 1) / block.y + 1);
gfp_gpu_mul_kernel<<<grid, block>>>(a.data, a.rowstride, b.data, b.rowstride, data, rowstride, nrows, width, a.width);
gpu_mul_kernel<<<grid, block>>>(a.data, a.rowstride, b.data, b.rowstride, data, rowstride, nrows, width, a.width);
cudaDeviceSynchronize();
}
}
#endif

17
include/header.cuh
View File

@ -6,14 +6,6 @@
#include <cpp_progress.hpp>
// matrix
// #include <map>
// #include <vector>
// #include <algorithm>
// #include <numeric>
// #include <omp.h>
using namespace std;
using base_t = uint64_t;
@ -25,13 +17,8 @@ static const base_t base_one = (base_t)0x00'00'00'00'00'00'00'01;
static const base_t base_fullmask = (base_t)0xFF'FF'FF'FF'FF'FF'FF'FF;
static const size_t THREAD_X = 32; // 列
static const size_t THREAD_Y = 8; // 行
// __host__ __device__ base_t *at_base(base_t *base, size_t rowstride, size_t r, size_t w)
// {
// return base + r * rowstride + w;
// }
static const size_t THREAD_X = 16; // 列
static const size_t THREAD_Y = 16; // 行
template <typename T>
__host__ __device__ T *at_base(T *base, size_t rowstride, size_t r, size_t w)

2
src/main.cu
View File

@ -4,6 +4,8 @@
#undef SHOW_PROGRESS_BAR
using namespace gfp;
int main()
{
int m = 1000, k = 1000, n = 1000;

2
test/test_gf256_elim.cu
View File

@ -1,6 +1,8 @@
#include <gtest/gtest.h>
#include "test_header.cuh"
using namespace gf256;
bool test_gf256_elim(size_t rank, size_t rank_col, size_t nrows, size_t ncols, const GF256 &gf256, uint_fast32_t seed)
{
assert(rank <= nrows && rank <= rank_col && rank_col <= ncols);

2
test/test_gf256_header.cu
View File

@ -1,6 +1,8 @@
#include <gtest/gtest.h>
#include "test_header.cuh"
using namespace gf256;
vector<gf256_t> expect_inv_table{
0x00, 0x01, 0x8E, 0xF4, 0x47, 0xA7, 0x7A, 0xBA, 0xAD, 0x9D, 0xDD, 0x98, 0x3D, 0xAA, 0x5D, 0x96,
0xD8, 0x72, 0xC0, 0x58, 0xE0, 0x3E, 0x4C, 0x66, 0x90, 0xDE, 0x55, 0x80, 0xA0, 0x83, 0x4B, 0x2A,

2
test/test_gf256_matrix.cu
View File

@ -1,6 +1,8 @@
#include <gtest/gtest.h>
#include "test_header.cuh"
using namespace gf256;
TEST(TestGF256Matrix, Equal)
{
MatGF256 a(50, 50);

2
test/test_gfp_mul.cu
View File

@ -1,6 +1,8 @@
#include <gtest/gtest.h>
#include "test_header.cuh"
using namespace gfp;
bool test_gfp_mul(size_t m, size_t k, size_t n, uint_fast32_t seed)
{
MatGFP a(m, k);