Skip to content

O
OpenXG-RAN
Commit 96023eeb authored by NCTU OpinConnect Terng-Yin Hsu/WEI-YING,LIN's avatar NCTU OpinConnect Terng-Yin Hsu/WEI-YING,LIN
Browse files
Options

Fix CUDA latency

parents 61bdea75 f7dc17ec
Hide whitespace changes
Inline Side-by-side
Showing with 221 additions and 0 deletions
openair1/PHY/CODING/TESTBENCH/ldpctest.c
View file @ 96023eeb
......@@ -402,9 +402,17 @@ int test_ldpc(short No_iteration,
start_meas(time_decoder);
#ifdef CUDA_FLAG
if(run_cuda){
<<<<<<< HEAD
n_iter = nrLDPC_decoder_LYC(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], block_length, time_decoder);
}
else{
=======
printf("***********run ldpc by cuda\n");
n_iter = nrLDPC_decoder_LYC(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], block_length, time_decoder);
}
else{
printf("**************run ldpc by cpu\n");
>>>>>>> origin/develop
// decode the sequence
// decoder supports BG2, Z=128 & 256
//esimated_output=ldpc_decoder(channel_output_fixed, block_length, No_iteration, (double)((float)nom_rate/(float)denom_rate));
......
openair1/PHY/CODING/nrLDPC_decoder_LYC/nrLDPC_decoder_LYC.cu
View file @ 96023eeb
......@@ -32,7 +32,11 @@
#include "bgs/BG2_I6"
#include "bgs/BG2_I7"
<<<<<<< HEAD
#define MAX_ITERATION 2
=======
#define MAX_ITERATION 5
>>>>>>> origin/develop
#define MC 1
#define cudaCheck(ans) { cudaAssert((ans), __FILE__, __LINE__); }
......@@ -49,21 +53,30 @@ typedef struct{
char y;
short value;
} h_element;
<<<<<<< HEAD
#include "bgs/BG1_compact_in_C.h"
__device__ char dev_const_llr[68*384];
__device__ char dev_dt [46*68*384];
__device__ char dev_llr[68*384];
__device__ unsigned char dev_tmp[68*384];
=======
>>>>>>> origin/develop
h_element h_compact1 [46*19] = {};
h_element h_compact2 [68*30] = {};
<<<<<<< HEAD
__device__ h_element dev_h_compact1[46*19]; // used in kernel 1
__device__ h_element dev_h_compact2[68*30]; // used in kernel 2
// __device__ __constant__ h_element dev_h_compact1[46*19]; // used in kernel 1
// __device__ __constant__ h_element dev_h_compact2[68*30]; // used in kernel 2
=======
__device__ __constant__ h_element dev_h_compact1[46*19]; // used in kernel 1
__device__ __constant__ h_element dev_h_compact2[68*30]; // used in kernel 2
>>>>>>> origin/develop
// row and col element count
__device__ __constant__ char h_ele_row_bg1_count[46] = {
......@@ -100,6 +113,7 @@ __global__ void warmup()
// warm up gpu for time measurement
}
<<<<<<< HEAD
extern "C"
void warmup_for_GPU(){
......@@ -187,6 +201,11 @@ void set_compact_BG(int Zc,short BG){
// Kernel 1
__global__ void ldpc_cnp_kernel_1st_iter(/*char * dev_llr,*/ int BG, int row, int col, int Zc)
=======
// Kernel 1
__global__ void ldpc_cnp_kernel_1st_iter(char * dev_llr, char * dev_dt, int BG, int row, int col, int Zc)
>>>>>>> origin/develop
{
// if(blockIdx.x == 0 && threadIdx.x == 1) printf("cnp %d\n", threadIdx.x);
int iMCW = blockIdx.y; // codeword id
......@@ -245,7 +264,11 @@ __global__ void ldpc_cnp_kernel_1st_iter(/*char * dev_llr,*/ int BG, int row, in
for(int i = 0; i < s; i++){
// v0: Best performance so far. 0.75f is the value of alpha.
sq = 1 - 2 * ((Q_sign >> i) & 0x01);
<<<<<<< HEAD
R_temp = 0.75f * sign * sq * (i != idx_min ? rmin1 : rmin2);
=======
R_temp = 0.8 * sign * sq * (i != idx_min ? rmin1 : rmin2);
>>>>>>> origin/develop
// write results to global memory
h_element_t = dev_h_compact1[i*row+iBlkRow];
int addr_temp = offsetR + h_element_t.y * row * Zc;
......@@ -255,7 +278,11 @@ __global__ void ldpc_cnp_kernel_1st_iter(/*char * dev_llr,*/ int BG, int row, in
}
// Kernel_1
<<<<<<< HEAD
__global__ void ldpc_cnp_kernel(/*char * dev_llr, char * dev_dt,*/ int BG, int row, int col, int Zc)
=======
__global__ void ldpc_cnp_kernel(char * dev_llr, char * dev_dt, int BG, int row, int col, int Zc)
>>>>>>> origin/develop
{
// if(blockIdx.x == 0 && threadIdx.x == 1) printf("cnp\n");
int iMCW = blockIdx.y;
......@@ -315,7 +342,11 @@ __global__ void ldpc_cnp_kernel(/*char * dev_llr, char * dev_dt,*/ int BG, int r
// The 2nd recursion
for(int i = 0; i < s; i ++){
sq = 1 - 2 * ((Q_sign >> i) & 0x01);
<<<<<<< HEAD
R_temp = 0.75f * sign * sq * (i != idx_min ? rmin1 : rmin2);
=======
R_temp = 0.8 * sign * sq * (i != idx_min ? rmin1 : rmin2);
>>>>>>> origin/develop
// write results to global memory
......@@ -328,7 +359,11 @@ __global__ void ldpc_cnp_kernel(/*char * dev_llr, char * dev_dt,*/ int BG, int r
// Kernel 2: VNP processing
__global__ void
<<<<<<< HEAD
ldpc_vnp_kernel_normal(/*char * dev_llr, char * dev_dt, char * dev_const_llr,*/ int BG, int row, int col, int Zc)
=======
ldpc_vnp_kernel_normal(char * dev_llr, char * dev_dt, char * dev_const_llr, int BG, int row, int col, int Zc)
>>>>>>> origin/develop
{
int iMCW = blockIdx.y;
int iBlkCol = blockIdx.x;
......@@ -368,7 +403,11 @@ ldpc_vnp_kernel_normal(/*char * dev_llr, char * dev_dt, char * dev_const_llr,*/
}
<<<<<<< HEAD
__global__ void pack_decoded_bit(/*char *dev, unsigned char *host,*/ int col, int Zc)
=======
__global__ void pack_decoded_bit(char *dev, unsigned char *host, int col, int Zc)
>>>>>>> origin/develop
{
__shared__ unsigned char tmp[128];
int iMCW = blockIdx.y;
......@@ -376,15 +415,25 @@ __global__ void pack_decoded_bit(/*char *dev, unsigned char *host,*/ int col, in
int btid = threadIdx.x;
tmp[btid] = 0;
<<<<<<< HEAD
if(dev_llr[tid] < 0){
=======
if(dev[tid] < 0){
>>>>>>> origin/develop
tmp[btid] = 1 << (7-(btid&7));
}
__syncthreads();
if(threadIdx.x < 16){
<<<<<<< HEAD
dev_tmp[iMCW * col*Zc + blockIdx.x*16+threadIdx.x] = 0;
for(int i = 0; i < 8; i++){
dev_tmp[iMCW * col*Zc + blockIdx.x*16+threadIdx.x] += tmp[threadIdx.x*8+i];
=======
host[iMCW * col*Zc + blockIdx.x*16+threadIdx.x] = 0;
for(int i = 0; i < 8; i++){
host[iMCW * col*Zc + blockIdx.x*16+threadIdx.x] += tmp[threadIdx.x*8+i];
>>>>>>> origin/develop
}
}
}
......@@ -461,6 +510,7 @@ void read_BG(int BG, int *h, int row, int col)
*/
}
<<<<<<< HEAD
extern "C"
void init_LLR_DMA_for_CUDA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length){
......@@ -482,17 +532,28 @@ void init_LLR_DMA_for_CUDA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8
cudaDeviceSynchronize();
}
=======
>>>>>>> origin/develop
extern "C"
int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length, time_stats_t *time_decoder)
{
<<<<<<< HEAD
=======
// alloc mem
//unsigned char *decision = (unsigned char*)p_out;
>>>>>>> origin/develop
uint16_t Zc = p_decParams->Z;
uint8_t BG = p_decParams->BG;
uint8_t numMaxIter = p_decParams->numMaxIter;
e_nrLDPC_outMode outMode = p_decParams->outMode;
<<<<<<< HEAD
cudaError_t cudaStatus;
=======
>>>>>>> origin/develop
uint8_t row,col;
if(BG == 1){
row = 46;
......@@ -502,6 +563,7 @@ int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8
row = 42;
col = 52;
}
<<<<<<< HEAD
// alloc memory
unsigned char *hard_decision = (unsigned char*)p_out;
......@@ -510,6 +572,98 @@ int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8
cudaCheck( cudaMemcpyToSymbol(dev_const_llr, p_llr, memorySize_llr_cuda) );
cudaCheck( cudaMemcpyToSymbol(dev_llr, p_llr, memorySize_llr_cuda) );
=======
int compact_row = 30, compact_col = 19, lift_index=0;;
if(BG==2){compact_row = 10, compact_col = 23;}
short lift_set[][9] = {
{2,4,8,16,32,64,128,256},
{3,6,12,24,48,96,192,384},
{5,10,20,40,80,160,320},
{7,14,28,56,112,224},
{9,18,36,72,144,288},
{11,22,44,88,176,352},
{13,26,52,104,208},
{15,30,60,120,240},
{0}
};
for(int i = 0; lift_set[i][0] != 0; i++){
for(int j = 0; lift_set[i][j] != 0; j++){
if(Zc == lift_set[i][j]){
lift_index = i;
break;
}
}
}
int *h = NULL;
switch(lift_index){
case 0:
h = (BG == 1)? h_base_0:h_base_8;
break;
case 1:
h = (BG == 1)? h_base_1:h_base_9;
break;
case 2:
h = (BG == 1)? h_base_2:h_base_10;
break;
case 3:
h = (BG == 1)? h_base_3:h_base_11;
break;
case 4:
h = (BG == 1)? h_base_4:h_base_12;
break;
case 5:
h = (BG == 1)? h_base_5:h_base_13;
break;
case 6:
h = (BG == 1)? h_base_6:h_base_14;
break;
case 7:
h = (BG == 1)? h_base_7:h_base_15;
break;
}
/* pack BG in compact graph */
read_BG(BG, h, row, col);
int memorySize_h_compact1 = row * compact_col * sizeof(h_element);
int memorySize_h_compact2 = compact_row * col * sizeof(h_element);
// cpu
int memorySize_hard_decision = col * Zc * sizeof(unsigned char) * MC;
// alloc memory
unsigned char *hard_decision = (unsigned char*)p_out;
// gpu
int memorySize_llr_cuda = col * Zc * sizeof(char) * MC;
int memorySize_dt_cuda = row * Zc * col * sizeof(char) * MC;
// alloc memory
char *dev_llr;
char *dev_dt;
char *dev_const_llr;
unsigned char *dev_tmp;
cudaCheck( cudaMalloc((void **)&dev_tmp, memorySize_hard_decision) );
cudaCheck( cudaMalloc((void **)&dev_llr, memorySize_llr_cuda) );
cudaCheck( cudaMalloc((void **)&dev_const_llr, memorySize_llr_cuda) );
cudaCheck( cudaMalloc((void **)&dev_dt, memorySize_dt_cuda) );
// memcpy host to device
cudaCheck( cudaMemcpyToSymbol(dev_h_compact1, h_compact1, memorySize_h_compact1) );
cudaCheck( cudaMemcpyToSymbol(dev_h_compact2, h_compact2, memorySize_h_compact2) );
cudaCheck( cudaMemcpy((void*)dev_const_llr, p_llr, memorySize_llr_cuda, cudaMemcpyHostToDevice) );
start_meas(time_decoder);
cudaCheck( cudaMemcpy((void*)dev_llr, p_llr, memorySize_llr_cuda, cudaMemcpyHostToDevice) );
>>>>>>> origin/develop
// Define CUDA kernel dimension
int blockSizeX = Zc;
dim3 dimGridKernel1(row, MC, 1); // dim of the thread blocks
......@@ -518,14 +672,33 @@ int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8
dim3 dimGridKernel2(col, MC, 1);
dim3 dimBlockKernel2(blockSizeX, 1, 1);
cudaDeviceSynchronize();
<<<<<<< HEAD
// lauch kernel
=======
cudaEvent_t start, end;
float time;
warmup<<<dimGridKernel1, dimBlockKernel1>>>();
warmup<<<dimGridKernel2, dimBlockKernel2>>>();
cudaEventCreate(&start);
cudaEventCreate(&end);
cudaEventRecord(start, 0);
// cudaProfilerStart();
// lauch kernel
>>>>>>> origin/develop
for(int ii = 0; ii < MAX_ITERATION; ii++){
// first kernel
if(ii == 0){
ldpc_cnp_kernel_1st_iter
<<<dimGridKernel1, dimBlockKernel1>>>
<<<<<<< HEAD
(/*dev_llr,*/ BG, row, col, Zc);
}else{
ldpc_cnp_kernel
......@@ -547,6 +720,43 @@ int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8
cudaDeviceSynchronize();
=======
(dev_llr, dev_dt, BG, row, col, Zc);
}else{
ldpc_cnp_kernel
<<<dimGridKernel1, dimBlockKernel1>>>
(dev_llr, dev_dt, BG, row, col, Zc);
}
// second kernel
ldpc_vnp_kernel_normal
<<<dimGridKernel2, dimBlockKernel2>>>
(dev_llr, dev_dt, dev_const_llr, BG, row, col, Zc);
}
int pack = (block_length/128)+1;
dim3 pack_block(pack, MC, 1);
pack_decoded_bit<<<pack_block,128>>>(dev_llr, dev_tmp, col, Zc);
cudaEventRecord(end, 0);
cudaEventSynchronize(end);
cudaEventElapsedTime(&time, start, end);
//cudaCheck( cudaMemcpy((*)hard_decision, (const void*)dev_tmp, memorySize_hard_decision, cudaMemcpyDeviceToHost) );
cudaCheck( cudaMemcpy((void*)hard_decision, (const void*)dev_tmp, (block_length/8)*sizeof(unsigned char), cudaMemcpyDeviceToHost) );
cudaDeviceSynchronize();
stop_meas(time_decoder);
cudaCheck( cudaFree(dev_llr) );
cudaCheck( cudaFree(dev_dt) );
cudaCheck( cudaFree(dev_const_llr) );
cudaCheck( cudaFree(dev_tmp) );
//free(hard_decision);
>>>>>>> origin/develop
return MAX_ITERATION;
}
openair1/PHY/CODING/nrLDPC_decoder_LYC/nrLDPC_decoder_LYC.h
View file @ 96023eeb
......@@ -23,10 +23,13 @@
int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length, time_stats_t *time_decoder);
<<<<<<< HEAD
void init_LLR_DMA_for_CUDA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length);
void warmup_for_GPU(void);
void set_compact_BG(int Zc, short BG);
=======
>>>>>>> origin/develop
#endif
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment