LDPC implementation on GPU
Signed-off-by: NCTU CS ISIP <tyhsu@cs.nctu.edu.tw>
This commit is contained in:
NCTU CS ISIP
@@ -0,0 +1,228 @@
|
||||
#include <stdio.h>
|
||||
#include <stdint.h>
|
||||
#include <string.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <cuda.h>
|
||||
#include "test.h"
|
||||
|
||||
void print_arr_cpu(const char *file, int *arr, int size)
|
||||
{
|
||||
FILE *fp = fopen(file, "w");
|
||||
for(int i = 0; i < size; i++){
|
||||
fprintf(fp, "%s[%d]: %d\n", file, i, arr[i]);
|
||||
}
|
||||
fclose(fp);
|
||||
}
|
||||
|
||||
void print_arr(const char *file, int *arr, int size)
|
||||
{
|
||||
int *tmp = (int*)malloc(sizeof(int)*size);
|
||||
FILE *fp = fopen(file, "w");
|
||||
cudaMemcpy((void*)tmp, (const void*)arr, size*sizeof(int), cudaMemcpyDeviceToHost);
|
||||
cudaDeviceSynchronize();
|
||||
for(int i = 0; i < size; i++){
|
||||
fprintf(fp, "%s[%d]: %d\n", file, i, tmp[i]);
|
||||
}
|
||||
free(tmp);
|
||||
fclose(fp);
|
||||
}
|
||||
|
||||
__global__ void llr2CN(int *llr, int *cnbuf, int *l2c_idx)
|
||||
{
|
||||
int tid = blockIdx.x*blockDim.x + threadIdx.x;
|
||||
|
||||
cnbuf[tid] = llr[l2c_idx[tid]];
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
__global__ void llr2BN(int *llr, int *const_llr, int *l2b_idx)
|
||||
{
|
||||
int tid = blockIdx.x*blockDim.x + threadIdx.x;
|
||||
|
||||
const_llr[tid] = llr[l2b_idx[tid]];
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
__global__ void CNProcess(int *cnbuf, int *bnbuf, int *b2c_idx, int *cnproc_idx)
|
||||
{
|
||||
int tid = blockIdx.x*blockDim.x + threadIdx.x;
|
||||
|
||||
int start = cnproc_idx[tid*2];
|
||||
int end = cnproc_idx[tid*2+1];
|
||||
/*
|
||||
if (tid == 6){
|
||||
printf("start %d, end %d\n", start, end);
|
||||
}
|
||||
*/
|
||||
int sgn = 1, val = INT32_MAX;
|
||||
for(int i = start; i < end; i++){
|
||||
if(i == tid) continue;
|
||||
|
||||
int tmp = cnbuf[i];
|
||||
// if(tid == 6) printf("tmp %d\n", tmp);
|
||||
if(tmp < 0){
|
||||
tmp = -tmp;
|
||||
sgn = -sgn;
|
||||
}
|
||||
if(val > tmp){
|
||||
val = tmp;
|
||||
// if(tid == 6)printf("val = tmp\n");
|
||||
}
|
||||
// if(tid == 6) printf("%d\n", val);
|
||||
}
|
||||
/*
|
||||
if(tid == 6){
|
||||
printf("b2c_idx[6] %d\n", b2c_idx[tid]);
|
||||
printf("sgn*val = %d\n", sgn*val);
|
||||
printf("\n\n");
|
||||
}
|
||||
*/
|
||||
bnbuf[b2c_idx[tid]] = sgn*val;// + const_llr[tid];
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
__global__ void BNProcess(int *const_llr, int *bnbuf, int *cnbuf, int *c2b_idx, int *bnproc_idx)
|
||||
{
|
||||
int tid = blockIdx.x*blockDim.x + threadIdx.x;
|
||||
|
||||
int start = bnproc_idx[tid*2];
|
||||
int end = bnproc_idx[tid*2+1];
|
||||
|
||||
int val = 0;
|
||||
for(int i = start; i < end; i++){
|
||||
if(i == tid) continue;
|
||||
val += bnbuf[i];
|
||||
}
|
||||
cnbuf[c2b_idx[tid]] = val + const_llr[tid];
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
__global__ void BN2llr(int *bnbuf, int *llrbuf, int *llr_idx)
|
||||
{
|
||||
int tid = blockIdx.x*blockDim.x + threadIdx.x;
|
||||
|
||||
int start = llr_idx[tid];
|
||||
int end = llr_idx[tid+1];
|
||||
|
||||
int res = 0;
|
||||
for(int i = start; i < end; i++){
|
||||
res += bnbuf[i];
|
||||
}
|
||||
llrbuf[tid] = res;
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
__global__ void BitDetermination(int *BN, unsigned int *decode_d)
|
||||
{
|
||||
__shared__ int tmp[256];
|
||||
int tid = blockIdx.x*256 + threadIdx.x;
|
||||
int bid = threadIdx.x;
|
||||
tmp[bid] = 0;
|
||||
|
||||
|
||||
if(BN[tid] < 0)
|
||||
{
|
||||
tmp[bid] = 1 << (bid&7);
|
||||
}
|
||||
|
||||
__syncthreads();
|
||||
|
||||
if(threadIdx.x < 32)
|
||||
{
|
||||
decode_d[blockIdx.x*32 + threadIdx.x] = 0;
|
||||
for(int i = 0; i < 8; i++)
|
||||
{
|
||||
decode_d[blockIdx.x*32 + threadIdx.x] += tmp[threadIdx.x*8+i];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
int main(int argc, char **argv)
|
||||
{
|
||||
int llr[15] = {-48, 27, -47, 13, 34,
|
||||
-41, 51, 29, -6, -19,
|
||||
-45, -42, -40, -6, -33};
|
||||
|
||||
int *llr_d, *llrbuf_d, *const_llr_d, *cnbuf_d, *bnbuf_d;
|
||||
int *l2c_idx_d, *cnproc_idx_d, *c2b_idx_d, *bnproc_idx_d, *b2c_idx_d, *llr_idx_d, *l2b_idx_d;
|
||||
|
||||
int blockNum = 6, threadNum = 11;
|
||||
int rounds = 5, Zc = 3;
|
||||
|
||||
size_t p_llr;
|
||||
|
||||
cudaMallocPitch((void**)&llr_d, &p_llr, 15*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&llrbuf_d, &p_llr, 15*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&const_llr_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&cnbuf_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&bnbuf_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&l2c_idx_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&l2b_idx_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&cnproc_idx_d, &p_llr, 33*2*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&c2b_idx_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&bnproc_idx_d, &p_llr, 33*2*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&b2c_idx_d, &p_llr, 33*sizeof(int), 1);
|
||||
cudaMallocPitch((void**)&llr_idx_d, &p_llr, 16*sizeof(int), 1);
|
||||
|
||||
cudaMemcpyAsync((void*)llr_d, (const void*)llr, 15*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)l2c_idx_d, (const void*)l2c_idx, 33*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)l2b_idx_d, (const void*)l2b_idx, 33*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)cnproc_idx_d, (const void*)cnproc_idx, 33*2*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)c2b_idx_d, (const void*)c2b_idx, 33*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)bnproc_idx_d, (const void*)bnproc_idx, 33*2*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)b2c_idx_d, (const void*)b2c_idx, 33*sizeof(int), cudaMemcpyHostToDevice);
|
||||
cudaMemcpyAsync((void*)llr_idx_d, (const void*)llr_idx, 16*sizeof(int), cudaMemcpyHostToDevice);
|
||||
|
||||
cudaDeviceSynchronize();
|
||||
|
||||
|
||||
|
||||
llr2CN<<<blockNum, threadNum>>>(llr_d, cnbuf_d, l2c_idx_d);
|
||||
llr2CN<<<blockNum, threadNum>>>(llr_d, const_llr_d, l2b_idx_d);
|
||||
|
||||
/*
|
||||
print_arr("debug/const_llr_d", const_llr_d, 26112);
|
||||
print_arr("debug/cnbuf_d", cnbuf_d, 316*384);
|
||||
print_arr("debug/const_llrbuf_d", const_llrbuf_d, 316*384);
|
||||
*/
|
||||
|
||||
char debug[] = "debug/";
|
||||
char cn[] = "cnbuf";
|
||||
char bn[] = "bnbuf";
|
||||
char llrstr[] = "llrbuf_d";
|
||||
char str[100] = {};
|
||||
for(int i = 0; i < rounds; i++){
|
||||
CNProcess<<<blockNum, threadNum>>>(cnbuf_d, bnbuf_d, b2c_idx_d, cnproc_idx_d);
|
||||
snprintf(str, 20, "%s%s_%d", debug, bn, i+1);
|
||||
cudaDeviceSynchronize();
|
||||
print_arr(str, bnbuf_d, 33);
|
||||
|
||||
BNProcess<<<blockNum, threadNum>>>(const_llr_d, bnbuf_d, cnbuf_d, c2b_idx_d, bnproc_idx_d);
|
||||
snprintf(str, 20, "%s%s_%d", debug, cn, i+1);
|
||||
print_arr(str, cnbuf_d, 33);
|
||||
|
||||
BN2llr<<<3,5>>>(bnbuf_d, llrbuf_d, llr_idx_d);
|
||||
snprintf(str, 20, "%s%s_%d", debug, llrstr, i+1);
|
||||
print_arr(str, llrbuf_d, 15);
|
||||
}
|
||||
|
||||
// BitDetermination<<<33, 256>>>(llrbuf_d, decode_output_d);
|
||||
cudaDeviceSynchronize();
|
||||
|
||||
|
||||
cudaFree(llrbuf_d);
|
||||
cudaFree(bnbuf_d);
|
||||
cudaFree(cnbuf_d);
|
||||
cudaFree(l2c_idx_d);
|
||||
cudaFree(l2b_idx_d);
|
||||
cudaFree(cnproc_idx_d);
|
||||
cudaFree(c2b_idx_d);
|
||||
cudaFree(bnproc_idx_d);
|
||||
cudaFree(b2c_idx_d);
|
||||
cudaFree(llr_d);
|
||||
cudaFree(const_llr_d);
|
||||
cudaFree(llr_idx_d);
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user