我想将我的答案与@asm一起发布,以便这个SO问题可以成为一个思路库。与@asm类似,我也检测并存储进位条件以及“carry-through”条件,即当中间字结果全部为1(0xF...FFF)时,如果进位传播到该字,则会“carry-through”到下一个字。
我没有在代码中使用任何PTX或asm,所以我选择使用64位无符号整数而不是32位,以实现2048x32位的能力,使用1024个线程。
与@asm的代码相比,我并行进位传播方案有更大的差异。我构建了一个位打包数组(“carry”),其中每个位表示来自1024个线程中每个独立中间64位加法生成的进位条件。我还构建了一个位打包数组(“carry_through”),其中每个位表示单个64位中间结果的carry_through条件。对于1024个线程,这相当于每个位打包数组16x64位共享内存,因此总共使用的共享内存量为64 + 3个32位数量。使用这些位打包数组,我执行以下操作以生成组合传播进位指示器:
carry = carry | (carry_through ^ ((carry & carry_through) + carry_through);
(请注意,进位向左移动了一位:carry[i]表示a[i-1]+b[i-1]的结果产生了一个进位)
以下是详细说明:
- 将carry和carry_through进行按位与运算,生成可能出现进位的地方,这些地方会与一个或多个进位条件序列相互作用
- 将步骤1的结果加上carry_through,生成一个有变化的结果,它代表着所有受到进位影响的单词
- 将carry_through和步骤2的结果进行异或操作,显示受影响的结果,其中包含1位
- 将步骤3的结果和普通的进位指示器进行按位或运算,得到组合进位条件,然后用它来更新所有中间结果。
请注意,步骤2中的加法需要进行另一种多词加法(对于由超过64个单词组成的大整数)。 我相信这种算法是有效的,并且已经通过了我投入的测试用例。
以下是我实现这个算法的示例代码:
#include <stdio.h>
#include <stdlib.h>
#define MAXSIZE 1024
#define LLBITS 64
#define BSIZE ((MAXSIZE + LLBITS -1)/LLBITS)
#define nTPB MAXSIZE
#define GPU
#define LOOPCNT 1000
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
__global__ void paradd(const unsigned size, const unsigned psize, unsigned long long *c, const unsigned long long *a, const unsigned long long *b){
__shared__ unsigned long long carry_through[BSIZE];
__shared__ unsigned long long carry[BSIZE+1];
__shared__ volatile unsigned mcarry;
__shared__ volatile unsigned mcarry_through;
unsigned idx = threadIdx.x + (psize * blockIdx.x);
if ((threadIdx.x < psize) && (idx < size)){
unsigned long long cr1 = a[idx];
unsigned long long lc = cr1 + b[idx];
if (threadIdx.x < BSIZE){
carry[threadIdx.x] = 0;
carry_through[threadIdx.x] = 0;
}
if (threadIdx.x == 0){
mcarry = 0;
mcarry_through = 0;
}
__syncthreads();
if (lc < cr1){
if ((threadIdx.x%LLBITS) != (LLBITS-1))
atomicAdd(&(carry[threadIdx.x/LLBITS]), (2ull<<(threadIdx.x%LLBITS)));
else atomicAdd(&(carry[(threadIdx.x/LLBITS)+1]), 1);
}
if (lc == 0xFFFFFFFFFFFFFFFFull)
atomicAdd(&(carry_through[threadIdx.x/LLBITS]), (1ull<<(threadIdx.x%LLBITS)));
__syncthreads();
if (threadIdx.x < ((psize + LLBITS-1)/LLBITS)){
unsigned long long cr3 = carry_through[threadIdx.x];
cr1 = carry[threadIdx.x] & cr3;
unsigned long long cr2 = cr3 + cr1;
if (cr2 < cr1) atomicAdd((unsigned *)&mcarry, (2u<<(threadIdx.x)));
if (cr2 == 0xFFFFFFFFFFFFFFFFull) atomicAdd((unsigned *)&mcarry_through, (1u<<threadIdx.x));
if (threadIdx.x == 0) {
unsigned cr4 = mcarry & mcarry_through;
cr4 += mcarry_through;
mcarry |= (mcarry_through ^ cr4);
}
if (mcarry & (1u<<threadIdx.x)) cr2++;
carry[threadIdx.x] |= (cr2 ^ cr3);
}
__syncthreads();
if (carry[threadIdx.x/LLBITS] & (1ull<<(threadIdx.x%LLBITS))) lc++;
c[idx] = lc;
}
}
int main() {
unsigned long long *h_a, *h_b, *h_c, *d_a, *d_b, *d_c, *c;
unsigned at_once = 256;
unsigned prob_size = MAXSIZE ;
unsigned dsize = at_once * prob_size;
cudaEvent_t t_start_gpu, t_start_cpu, t_end_gpu, t_end_cpu;
float et_gpu, et_cpu, tot_gpu, tot_cpu;
tot_gpu = 0;
tot_cpu = 0;
if (sizeof(unsigned long long) != (LLBITS/8)) {printf("Word Size Error\n"); return 1;}
if ((c = (unsigned long long *)malloc(dsize * sizeof(unsigned long long))) == 0) {printf("Malloc Fail\n"); return 1;}
cudaHostAlloc((void **)&h_a, dsize * sizeof(unsigned long long), cudaHostAllocDefault);
cudaCheckErrors("cudaHostAlloc1 fail");
cudaHostAlloc((void **)&h_b, dsize * sizeof(unsigned long long), cudaHostAllocDefault);
cudaCheckErrors("cudaHostAlloc2 fail");
cudaHostAlloc((void **)&h_c, dsize * sizeof(unsigned long long), cudaHostAllocDefault);
cudaCheckErrors("cudaHostAlloc3 fail");
cudaMalloc((void **)&d_a, dsize * sizeof(unsigned long long));
cudaCheckErrors("cudaMalloc1 fail");
cudaMalloc((void **)&d_b, dsize * sizeof(unsigned long long));
cudaCheckErrors("cudaMalloc2 fail");
cudaMalloc((void **)&d_c, dsize * sizeof(unsigned long long));
cudaCheckErrors("cudaMalloc3 fail");
cudaMemset(d_c, 0, dsize*sizeof(unsigned long long));
cudaEventCreate(&t_start_gpu);
cudaEventCreate(&t_end_gpu);
cudaEventCreate(&t_start_cpu);
cudaEventCreate(&t_end_cpu);
for (unsigned loops = 0; loops <LOOPCNT; loops++){
if (loops == 0){
for (int j=0; j<at_once; j++)
for (int k=0; k<prob_size; k++){
int i= (j*prob_size) + k;
h_a[i] = 0xFFFFFFFFFFFFFFFFull;
h_b[i] = 0;
}
h_a[prob_size-1] = 0;
h_b[prob_size-1] = 1;
h_b[0] = 1;
}
else if (loops == 1){
for (int i=0; i<dsize; i++){
h_a[i] = 0xFFFFFFFFFFFFFFFFull;
h_b[i] = 0;
}
h_b[0] = 1;
}
else if (loops == 2){
for (int i=0; i<dsize; i++){
h_a[i] = 0xFFFFFFFFFFFFFFFEull;
h_b[i] = 2;
}
h_b[0] = 1;
}
else {
for (int i = 0; i<dsize; i++){
h_a[i] = (((unsigned long long)lrand48())<<33) + (unsigned long long)lrand48();
h_b[i] = (((unsigned long long)lrand48())<<33) + (unsigned long long)lrand48();
}
}
#ifdef GPUCOPY
cudaEventRecord(t_start_gpu, 0);
#endif
cudaMemcpy(d_a, h_a, dsize*sizeof(unsigned long long), cudaMemcpyHostToDevice);
cudaCheckErrors("cudaMemcpy1 fail");
cudaMemcpy(d_b, h_b, dsize*sizeof(unsigned long long), cudaMemcpyHostToDevice);
cudaCheckErrors("cudaMemcpy2 fail");
#ifdef GPU
cudaEventRecord(t_start_gpu, 0);
#endif
paradd<<<at_once, nTPB>>>(dsize, prob_size, d_c, d_a, d_b);
cudaCheckErrors("Kernel Fail");
#ifdef GPU
cudaEventRecord(t_end_gpu, 0);
#endif
cudaMemcpy(h_c, d_c, dsize*sizeof(unsigned long long), cudaMemcpyDeviceToHost);
cudaCheckErrors("cudaMemcpy3 fail");
#ifdef GPUCOPY
cudaEventRecord(t_end_gpu, 0);
#endif
cudaEventSynchronize(t_end_gpu);
cudaEventElapsedTime(&et_gpu, t_start_gpu, t_end_gpu);
tot_gpu += et_gpu;
cudaEventRecord(t_start_cpu, 0);
for (int j=0; j<at_once; j++) {
unsigned rc=0;
for (int n=0; n<prob_size; n++){
unsigned i = (j*prob_size) + n;
c[i] = h_a[i] + h_b[i];
if (c[i] < h_a[i]) {
c[i] += rc;
rc=1;}
else {
if ((c[i] += rc) != 0) rc=0;
}
if (c[i] != h_c[i]) {printf("Results mismatch at offset %d, GPU = 0x%lX, CPU = 0x%lX\n", i, h_c[i], c[i]); return 1;}
}
}
cudaEventRecord(t_end_cpu, 0);
cudaEventSynchronize(t_end_cpu);
cudaEventElapsedTime(&et_cpu, t_start_cpu, t_end_cpu);
tot_cpu += et_cpu;
if ((loops%(LOOPCNT/10)) == 0) printf("*\n");
}
printf("\nResults Match!\n");
printf("Average GPU time = %fms\n", (tot_gpu/LOOPCNT));
printf("Average CPU time = %fms\n", (tot_cpu/LOOPCNT));
return 0;
}