我正在学习 CUDA,在查看了 PTX 手册后,我发现有一条名为 red.shared 的指令可以在扭曲中执行缩减。
我很好奇硬件是否具有对减少的本机支持。如果是这样,如何在 CUDA 代码中使用它?也许有人已经尝试过了?
实际上恰好是这样,我出于好奇也尝试了“红色”指令。我不知道这在开普勒上是怎么回事,但在费米架构上,“红色”指令只是映射到另一个指令序列。也许他们把它留给了未来的 GPU。这是我玩过的代码:
#define WS 32
#define HF 16
__global__ void test_red_kernel(unsigned *g_R, const unsigned *g_U) {
extern __shared__ unsigned shared[];
unsigned thid = threadIdx.x, bidx_x = blockIdx.x;
unsigned *r = shared;
unsigned ofs = bidx_x << 7, thid_in_warp = thid & WS-1;
unsigned a = (g_U + ofs)[thid];
volatile unsigned *t = (volatile unsigned *)r + HF + UMUL(thid >> 5,
WS + HF + 1) + thid_in_warp;
t[-HF] = 0;
t[0] = a;
// warp reduction
a = a + t[-HF], t[0] = a;
a = a + t[-8], t[0] = a;
a = a + t[-4], t[0] = a;
a = a + t[-2], t[0] = a;
a = a + t[-1], t[0] = a;
CU_SYNC
volatile unsigned *t2 = r + HF + UMUL(WS*4 >> 5, WS + HF + 1);
if(thid < 4) {
unsigned loc_ofs = HF + WS-1 + UMUL(thid, WS + HF + 1);
unsigned a2;
volatile unsigned *ps = t2 + thid;
ps[-2] = 0;
a2 = r[loc_ofs]; ps[0] = a2;
a2 = a2 + ps[-2], ps[0] = a2;
a2 = a2 + ps[-1], ps[0] = a2;
}
CU_SYNC
a = a + t2[(thid >> 5) - 1];
unsigned b;
asm volatile("mov.u32 %r11, shared;" : );
asm volatile("red.shared.add.u32 [%r11], %0;" :
"+r"(b) : );
b = r[0]; // results of 'red.shared', compare it with a
(g_R + ofs)[thid] = a - b;
}
要查看“红色”指令是如何在硬件中实现的,您可以在生成的“cubin”文件上使用 cuobjdump 工具(使用选项 -keep 和 nvcc)