fix(nrLDPC_decoder_LYC): apply clang-format

This commit is contained in:
Romain Beurdouche
2025-08-08 14:48:41 +00:00
parent 8d2df31d7b
commit 37bf1deb7d
@@ -32,10 +32,13 @@
#include "bgs/BG2_I6" #include "bgs/BG2_I6"
#include "bgs/BG2_I7" #include "bgs/BG2_I7"
#define MAX_ITERATION 2 #define MAX_ITERATION 2
#define MC 1 #define MC 1
typedef void decode_abort_t; typedef void decode_abort_t;
#define cudaCheck(ans) { cudaAssert((ans), __FILE__, __LINE__); } #define cudaCheck(ans) \
{ \
cudaAssert((ans), __FILE__, __LINE__); \
}
inline void cudaAssert(cudaError_t code, const char *file, int line) inline void cudaAssert(cudaError_t code, const char *file, int line)
{ {
if (code != cudaSuccess) { if (code != cudaSuccess) {
@@ -44,23 +47,23 @@ inline void cudaAssert(cudaError_t code, const char *file, int line)
} }
} }
typedef struct{ typedef struct {
char x; char x;
char y; char y;
short value; short value;
} h_element; } h_element;
#include "bgs/BG1_compact_in_C.h" #include "bgs/BG1_compact_in_C.h"
__device__ char dev_const_llr[68*384]; __device__ char dev_const_llr[68 * 384];
__device__ char dev_dt [46*68*384]; __device__ char dev_dt[46 * 68 * 384];
__device__ char dev_llr[68*384]; __device__ char dev_llr[68 * 384];
__device__ unsigned char dev_tmp[68*384]; __device__ unsigned char dev_tmp[68 * 384];
h_element h_compact1 [46*23] = {}; h_element h_compact1[46 * 23] = {};
h_element h_compact2 [68*30] = {}; h_element h_compact2[68 * 30] = {};
__device__ h_element dev_h_compact1[46*23]; // used in kernel 1 __device__ h_element dev_h_compact1[46 * 23]; // used in kernel 1
__device__ h_element dev_h_compact2[68*30]; // used in kernel 2 __device__ h_element dev_h_compact2[68 * 30]; // used in kernel 2
// __device__ __constant__ h_element dev_h_compact1[46*23]; // used in kernel 1 // __device__ __constant__ h_element dev_h_compact1[46*23]; // used in kernel 1
// __device__ __constant__ h_element dev_h_compact2[68*30]; // used in kernel 2 // __device__ __constant__ h_element dev_h_compact2[68*30]; // used in kernel 2
@@ -82,13 +85,13 @@ __global__ void warmup()
// warm up gpu for time measurement // warm up gpu for time measurement
} }
extern "C" extern "C" void warmup_for_GPU()
void warmup_for_GPU(){ {
warmup<<<20, 1024>>>(); warmup<<<20, 1024>>>();
} }
extern "C" extern "C" void set_compact_BG(int Zc, short BG)
void set_compact_BG(int Zc,short BG){ {
int row, col; int row, col;
if (BG == 1) { if (BG == 1) {
row = 46; row = 46;
@@ -162,7 +165,6 @@ void set_compact_BG(int Zc,short BG){
// return 0; // return 0;
} }
// Kernel 1 // Kernel 1
__global__ void ldpc_cnp_kernel_1st_iter(/*char * dev_llr,*/ int BG, int row, int col, int Zc) __global__ void ldpc_cnp_kernel_1st_iter(/*char * dev_llr,*/ int BG, int row, int col, int Zc)
{ {
@@ -303,8 +305,7 @@ __global__ void ldpc_cnp_kernel(/*char * dev_llr, char * dev_dt,*/ int BG, int r
} }
// Kernel 2: VNP processing // Kernel 2: VNP processing
__global__ void __global__ void 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)
{ {
int iMCW = blockIdx.y; int iMCW = blockIdx.y;
int iBlkCol = blockIdx.x; int iBlkCol = blockIdx.x;
@@ -344,7 +345,6 @@ ldpc_vnp_kernel_normal(/*char * dev_llr, char * dev_dt, char * dev_const_llr,*/
dev_llr[iMCW * col * Zc + iCol] = APP; dev_llr[iMCW * col * Zc + iCol] = APP;
} }
__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)
{ {
__shared__ unsigned char tmp[128]; __shared__ unsigned char tmp[128];
@@ -440,8 +440,8 @@ void read_BG(int BG, int *h, int row, int col)
*/ */
} }
extern "C" extern "C" void init_LLR_DMA(t_nrLDPC_dec_params *p_decParams, int8_t *p_llr, int8_t *p_out)
void init_LLR_DMA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out){ {
uint16_t Zc = p_decParams->Z; uint16_t Zc = p_decParams->Z;
uint8_t BG = p_decParams->BG; uint8_t BG = p_decParams->BG;
uint8_t col; uint8_t col;
@@ -456,11 +456,11 @@ void init_LLR_DMA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out
cudaDeviceSynchronize(); cudaDeviceSynchronize();
} }
using namespace std ; using namespace std;
/* from here: entry points in decoder shared lib */ /* from here: entry points in decoder shared lib */
extern "C" extern "C" int ldpc_autoinit(void)
int ldpc_autoinit(void) { // called by the library loader { // called by the library loader
/*int devices = 0; /*int devices = 0;
cudaError_t err = cudaGetDeviceCount(&devices); cudaError_t err = cudaGetDeviceCount(&devices);
@@ -493,7 +493,7 @@ int ldpc_autoinit(void) { // called by the library loader
} }
*/ */
warmup_for_GPU(); warmup_for_GPU();
return 0; return 0;
} }
extern "C" int32_t LDPCinit() extern "C" int32_t LDPCinit()
@@ -505,11 +505,11 @@ extern "C" void LDPCshutdown()
{ {
} }
extern "C" int32_t LDPCdecoder(t_nrLDPC_dec_params *p_decParams, extern "C" int32_t LDPCdecoder(t_nrLDPC_dec_params *p_decParams,
int8_t *p_llr, int8_t *p_llr,
int8_t *p_out, int8_t *p_out,
t_nrLDPC_time_stats *, t_nrLDPC_time_stats *,
decode_abort_t *ab) decode_abort_t *ab)
{ {
set_compact_BG(p_decParams->Z, p_decParams->BG); set_compact_BG(p_decParams->Z, p_decParams->BG);
init_LLR_DMA(p_decParams, p_llr, p_out); init_LLR_DMA(p_decParams, p_llr, p_out);
@@ -526,7 +526,7 @@ extern "C" int32_t LDPCdecoder(t_nrLDPC_dec_params *p_decParams,
} }
// alloc memory // alloc memory
unsigned char* hard_decision = (unsigned char*)p_out; unsigned char *hard_decision = (unsigned char *)p_out;
// gpu // gpu
int memorySize_llr_cuda = col * Zc * sizeof(char) * MC; int memorySize_llr_cuda = col * Zc * sizeof(char) * MC;
cudaCheck(cudaMemcpyToSymbol(dev_const_llr, p_llr, memorySize_llr_cuda)); cudaCheck(cudaMemcpyToSymbol(dev_const_llr, p_llr, memorySize_llr_cuda));
@@ -560,7 +560,7 @@ extern "C" int32_t LDPCdecoder(t_nrLDPC_dec_params *p_decParams,
dim3 pack_block(pack, MC, 1); dim3 pack_block(pack, MC, 1);
pack_decoded_bit<<<pack_block, 128>>>(/*dev_llr,*/ /*dev_tmp,*/ col, Zc); pack_decoded_bit<<<pack_block, 128>>>(/*dev_llr,*/ /*dev_tmp,*/ col, Zc);
cudaCheck(cudaMemcpyFromSymbol((void*)hard_decision, (const void*)dev_tmp, (block_length / 8) * sizeof(unsigned char))); cudaCheck(cudaMemcpyFromSymbol((void *)hard_decision, (const void *)dev_tmp, (block_length / 8) * sizeof(unsigned char)));
cudaDeviceSynchronize(); cudaDeviceSynchronize();
return MAX_ITERATION; return MAX_ITERATION;