在NVIDIA精良的reduction optimization documentation中，他们最终得到一个如下所示的 warpReduce:
Template <unsigned int blockSize>
__device__ void warpReduce(volatile int* sdata, int tid) {
if (blockSize >= 64) sdata[tid] += sdata[tid + 32];
if (blockSize >= 32) sdata[tid] += sdata[tid + 16];
if (blockSize >= 16) sdata[tid] += sdata[tid + 8];
if (blockSize >= 8) sdata[tid] += sdata[tid + 4];
if (blockSize >= 4) sdata[tid] += sdata[tid + 2];
if (blockSize >= 2) sdata[tid] += sdata[tid + 1];
}
这伤害了我作为 Don't Repeat Yourself 支持者的情感，为什么他们不节省程序员的时间并引入更像这样的结构
Template <unsigned int blockSize>
__device__ void warpReduce(volatile int* sdata, int tid) {
#pragma unroll(6)
for (int i = 64; i>1; i>>=1){
if (blockSize >= i) sdata[tid] += sdata[tid + (i/2)];
}
}
这会编译成相同的代码吗？或者这是否会让编译器过于困惑而无法对其进行优化？我知道 case 语句遇到类似的 DRY 问题，但我不知道有办法解决这个问题，因为它将运行时输入映射到编译器时间常量定义的模板内核，但我不知道如何让程序员需要多次写出相同的内容是有帮助的。
更不用说，如果扭曲大小发生变化，这可能不会(但谁知道)我引入的代码将更容易更改，甚至可以使用常量来定义扭曲大小，从而允许一次更改更改此处以及依赖扭曲大小的其他任何地方的优化。
非常相关的子问题，我读到使用内置的 warpSize 出于编译器优化的原因也是有问题的。否则我会把它包含在上面的代码中。转到device_launch_parameters.h中的定义，它调用内置设备查询__cudaGet_warpSize()。如果这不是问题，NVIDIA 可以在该文档中提供完整的、通用的 warpReduce 优化以供引用，但没有，我想知道为什么？
Tl;dr:我可以把它写成
Template <unsigned int blockSize>
__device__ void warpReduce(volatile int* sdata, int tid) {
#pragma unroll(6)
for (int i = warpSize * 2; i>1; i>>=1){
if (blockSize >= i) sdata[tid] += sdata[tid + (i/2)];
}
}
并享受与其文档中相同的优化？
最佳答案
这是一个富有洞察力的观察(我认为)。基于我的测试代码:
#include <stdio.h>
template <unsigned int blockSize>
__device__ void warpReduce1(volatile int* sdata, int tid) {
if (blockSize >= 64) sdata[tid] += sdata[tid + 32];
if (blockSize >= 32) sdata[tid] += sdata[tid + 16];
if (blockSize >= 16) sdata[tid] += sdata[tid + 8];
if (blockSize >= 8) sdata[tid] += sdata[tid + 4];
if (blockSize >= 4) sdata[tid] += sdata[tid + 2];
if (blockSize >= 2) sdata[tid] += sdata[tid + 1];
}
template <unsigned int blockSize>
__device__ void warpReduce2(volatile int* sdata, int tid) {
#pragma unroll 6
for (int i = 64; i>1; i>>=1){
if (blockSize >= i) sdata[tid] += sdata[tid + (i/2)];
}
}
template <unsigned int blockSize>
__global__ void reduce6(int *g_idata, int *g_odata, unsigned int n) {
extern __shared__ int sdata[];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*(blockSize*2) + tid;
unsigned int gridSize = blockSize*2*gridDim.x;
sdata[tid] = 0;
while (i < n){sdata[tid] += g_idata[i] + g_idata[i+blockSize]; i += gridSize; }
__syncthreads();
if (blockSize >= 512) { if (tid < 256) { sdata[tid] += sdata[tid + 256]; } __syncthreads(); }
if (blockSize >= 256) { if (tid < 128) { sdata[tid] += sdata[tid + 128]; } __syncthreads(); }
if (blockSize >= 128) { if (tid < 64) { sdata[tid] += sdata[tid + 64]; } __syncthreads(); }
if (tid < 32) warpReduce1<blockSize>(sdata, tid);
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}
#define DSIZE 1048576
#define BSIZE 256
#define NBLKS 64
int main(){
int *h_data, *d_idata, *d_odata;
h_data=(int *)malloc(DSIZE * sizeof(int));
cudaMalloc(&d_idata, DSIZE * sizeof(int));
cudaMalloc(&d_odata, NBLKS * sizeof(int));
for (int i = 0; i < DSIZE; i++) h_data[i] = rand()%2;
cudaMemcpy(d_idata, h_data, DSIZE*sizeof(int), cudaMemcpyHostToDevice);
reduce6<BSIZE><<<64, BSIZE>>>(d_idata, d_odata, DSIZE);
cudaMemcpy(h_data, d_odata, NBLKS*sizeof(int), cudaMemcpyDeviceToHost);
return 0;
}
看起来生成的机器代码是相同的:
Fatbin elf code:
================
arch = sm_20
code version = [1,7]
producer = <unknown>
host = linux
compile_size = 64bit
identifier = t576.cu
code for sm_20
Fatbin elf code:
================
arch = sm_20
code version = [1,7]
producer = cuda
host = linux
compile_size = 64bit
identifier = t576.cu
code for sm_20
Function : _Z7reduce6ILj256EEvPiS0_j
.headerflags @"EF_CUDA_SM20 EF_CUDA_PTX_SM(EF_CUDA_SM20)"
/*0000*/ MOV R1, c[0x1][0x100]; /* 0x2800440400005de4 */
/*0008*/ S2R R0, SR_CTAID.X; /* 0x2c00000094001c04 */
/*0010*/ S2R R2, SR_TID.X; /* 0x2c00000084009c04 */
/*0018*/ MOV R4, c[0x0][0x14]; /* 0x2800400050011de4 */
/*0020*/ MOV R10, RZ; /* 0x28000000fc029de4 */
/*0028*/ ISCADD R8, R0, R2, 0x9; /* 0x4000000008021d23 */
/*0030*/ SHL.W R3, R2, 0x2; /* 0x6000c0000820de03 */
/*0038*/ SHL R9, R4, 0x9; /* 0x6000c00024425c03 */
/*0040*/ ISETP.GE.U32.AND P0, PT, R8, c[0x0][0x30], PT; /* 0x1b0e4000c081dc03 */
/*0048*/ STS [R3], RZ; /* 0xc9000000003fdc85 */
/*0050*/ SSY 0xf0; /* 0x6000000260000007 */
/*0058*/ @P0 NOP.S; /* 0x40000000000001f4 */
/*0060*/ MOV32I R11, 0x4; /* 0x180000001002dde2 */
/*0068*/ IMAD.U32.U32 RZ, R1, RZ, RZ; /* 0x207e0000fc1fdc03 */
/*0070*/ SSY 0xe8; /* 0x60000001c0000007 */
/*0078*/ NOP; /* 0x4000000000001de4 */
/*0080*/ IMAD.U32.U32 R6.CC, R8, R11, c[0x0][0x20]; /* 0x2017800080819c03 */
/*0088*/ IADD R5, R8, 0x100; /* 0x4800c00400815c03 */
/*0090*/ IMAD.U32.U32.HI.X R7, R8, R11, c[0x0][0x24]; /* 0x209680009081dc43 */
/*0098*/ IMAD.U32.U32 R4.CC, R5, R11, c[0x0][0x20]; /* 0x2017800080511c03 */
/*00a0*/ IADD R8, R8, R9; /* 0x4800000024821c03 */
/*00a8*/ LD.E R7, [R6]; /* 0x840000000061dc85 */
/*00b0*/ IMAD.U32.U32.HI.X R5, R5, R11, c[0x0][0x24]; /* 0x2096800090515c43 */
/*00b8*/ ISETP.LT.U32.AND P0, PT, R8, c[0x0][0x30], PT; /* 0x188e4000c081dc03 */
/*00c0*/ LD.E R4, [R4]; /* 0x8400000000411c85 */
/*00c8*/ IADD R5, R7, R10; /* 0x4800000028715c03 */
/*00d0*/ IADD R10, R5, R4; /* 0x4800000010529c03 */
/*00d8*/ @P0 BRA 0x80; /* 0x4003fffe800001e7 */
/*00e0*/ NOP.S; /* 0x4000000000001df4 */
/*00e8*/ STS.S [R3], R10; /* 0xc900000000329c95 */
/*00f0*/ IMAD.U32.U32 RZ, R1, RZ, RZ; /* 0x207e0000fc1fdc03 */
/*00f8*/ BAR.RED.POPC RZ, RZ, RZ, PT; /* 0x50ee0000ffffdc04 */
/*0100*/ ISETP.GT.U32.AND P0, PT, R2, 0x7f, PT; /* 0x1a0ec001fc21dc03 */
/*0108*/ @!P0 LDS R5, [R3]; /* 0xc100000000316085 */
/*0110*/ @!P0 LDS R4, [R3+0x200]; /* 0xc100000800312085 */
/*0118*/ @!P0 IADD R4, R5, R4; /* 0x4800000010512003 */
/*0120*/ @!P0 STS [R3], R4; /* 0xc900000000312085 */
/*0128*/ BAR.RED.POPC RZ, RZ, RZ, PT; /* 0x50ee0000ffffdc04 */
/*0130*/ ISETP.GT.U32.AND P0, PT, R2, 0x3f, PT; /* 0x1a0ec000fc21dc03 */
/*0138*/ @!P0 LDS R5, [R3]; /* 0xc100000000316085 */
/*0140*/ @!P0 LDS R4, [R3+0x100]; /* 0xc100000400312085 */
/*0148*/ @!P0 IADD R4, R5, R4; /* 0x4800000010512003 */
/*0150*/ @!P0 STS [R3], R4; /* 0xc900000000312085 */
/*0158*/ SSY 0x240; /* 0x6000000380000007 */
/*0160*/ BAR.RED.POPC RZ, RZ, RZ, PT; /* 0x50ee0000ffffdc04 */
/*0168*/ ISETP.GT.U32.AND P0, PT, R2, 0x1f, PT; /* 0x1a0ec0007c21dc03 */
/*0170*/ @P0 NOP.S; /* 0x40000000000001f4 */
/*0178*/ SHL.W R3, R2, 0x2; /* 0x6000c0000820de03 */
/*0180*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*0188*/ LDS R4, [R3+0x80]; /* 0xc100000200311c85 */
/*0190*/ IADD R6, R5, R4; /* 0x4800000010519c03 */
/*0198*/ STS [R3], R6; /* 0xc900000000319c85 */
/*01a0*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*01a8*/ LDS R4, [R3+0x40]; /* 0xc100000100311c85 */
/*01b0*/ IADD R7, R5, R4; /* 0x480000001051dc03 */
/*01b8*/ STS [R3], R7; /* 0xc90000000031dc85 */
/*01c0*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*01c8*/ LDS R4, [R3+0x20]; /* 0xc100000080311c85 */
/*01d0*/ IADD R6, R5, R4; /* 0x4800000010519c03 */
/*01d8*/ STS [R3], R6; /* 0xc900000000319c85 */
/*01e0*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*01e8*/ LDS R4, [R3+0x10]; /* 0xc100000040311c85 */
/*01f0*/ IADD R7, R5, R4; /* 0x480000001051dc03 */
/*01f8*/ STS [R3], R7; /* 0xc90000000031dc85 */
/*0200*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*0208*/ LDS R4, [R3+0x8]; /* 0xc100000020311c85 */
/*0210*/ IADD R6, R5, R4; /* 0x4800000010519c03 */
/*0218*/ STS [R3], R6; /* 0xc900000000319c85 */
/*0220*/ LDS R5, [R3]; /* 0xc100000000315c85 */
/*0228*/ LDS R4, [R3+0x4]; /* 0xc100000010311c85 */
/*0230*/ IADD R4, R5, R4; /* 0x4800000010511c03 */
/*0238*/ STS.S [R3], R4; /* 0xc900000000311c95 */
/*0240*/ ISETP.NE.AND P0, PT, R2, RZ, PT; /* 0x1a8e0000fc21dc23 */
/*0248*/ @P0 BRA.U 0x278; /* 0x40000000a00081e7 */
/*0250*/ @!P0 MOV32I R3, 0x4; /* 0x180000001000e1e2 */
/*0258*/ @!P0 LDS R2, [RZ]; /* 0xc100000003f0a085 */
/*0260*/ @!P0 IMAD.U32.U32 R4.CC, R0, R3, c[0x0][0x28]; /* 0x20078000a0012003 */
/*0268*/ @!P0 IMAD.U32.U32.HI.X R5, R0, R3, c[0x0][0x2c]; /* 0x20868000b0016043 */
/*0270*/ @!P0 ST.E [R4], R2; /* 0x940000000040a085 */
/*0278*/ EXIT; /* 0x8000000000001de7 */
..........................................
Fatbin ptx code:
================
arch = sm_20
code version = [4,1]
producer = cuda
host = linux
compile_size = 64bit
compressed
identifier = t576.cu
我是否使用:
if (tid < 32) warpReduce1<blockSize>(sdata, tid);
或者:
if (tid < 32) warpReduce2<blockSize>(sdata, tid);
回想起来，编译器可以轻松确定 for 循环的行程计数:
for (int i = 64; i>1; i>>=1){
展开的循环将生成与内联代码完全相同的序列。因此编译器生成相同的代码。
关于c++ - #pragma unroll 语句可以用来清理编译器预评估的代码吗？，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/26062378/