paizaでインラインアセンブラ その5

概要

paizaで、インラインアセンブラやってみた。
コード見つけたので、やってみた。

参考にしたページ

サンプルコード

#include <time.h>
#include <sys/time.h>
#include <stdint.h>
#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>

#define GHz 2.00

static inline uint64_t rdtsc() {
    uint64_t ret;
#if defined _LP64
    __asm__ volatile (
        "rdtsc  \n\t"
        "mov    $32, %%rdx\n\t"
        "orq    %%rdx, %%rax\n\t"
        "mov    %%rax, %0\n\t"
        :"=m"(ret)
        :
        :"%rax", "%rdx"
    );
#else
    __asm__ volatile (
        "rdtsc  \n\t"
        "mov    %%eax, %0\n\t"
        "mov    %%edx, %1\n\t"
        :"=m"(((uint32_t<span style="font-weight:bold;">)&ret)[0]), "=m"(((uint32_t</span>)&ret)[1])
        :
        :"%eax", "%edx"
    );
#endif
    return ret;
}

void print_GFLOPS(double flops, uint64_t cycles) {
    double GFLOPS = flops * GHz / cycles;
    double sec = (double) cycles * 1e-9 / GHz;
    printf("GFLOPS @ %.2fGHz:\n  %.3f [flops/clock] = %.3f [GFLOPS]  (%.0f flops in %"PRIu64" clock = %f sec)\n",
         GHz, flops / (double) cycles, GFLOPS, flops, cycles, sec);
}

void print_throughput(uint64_t instructions, uint64_t cycles) {
    printf("Throughput:\n  %.3f [instructions/clock]   (%"PRIu64" instrucions in %"PRIu64" clock)\n",
         (double) instructions / (double) cycles, instructions, cycles);
}

#define zero_all_xmm() \
    do { \
        __asm__ volatile (\
            "xorps %xmm0, %xmm0\n\t" \
            "xorps %xmm1, %xmm1\n\t" \
            "xorps %xmm2, %xmm2\n\t" \
            "xorps %xmm3, %xmm3\n\t" \
            "xorps %xmm4, %xmm4\n\t" \
            "xorps %xmm5, %xmm5\n\t" \
            "xorps %xmm6, %xmm6\n\t" \
            "xorps %xmm7, %xmm7\n\t" \
        ); \
    } while (0)

#define LOOP (1 << 21)

void sse_mulps_addps_movups_shufps_fromL1() {
    printf("-- sse_mulps_addps_movups_shufps_fromL1 --\n");
    unsigned long long clk0, 
        clk1, 
        cycles;
    zero_all_xmm();
    const int instructions_per_loop = 16;
    const int flops_per_instruction = 4;
    const double flops = flops_per_instruction * instructions_per_loop * LOOP;
    int i;
    float __attribute__ ((aligned(16))) a[4] = {0.0, 0.0, 0.0, 0.0};
    float __attribute__ ((aligned(16))) b[4] = {0.0, 0.0, 0.0, 0.0};
    float __attribute__ ((aligned(16))) c[4] = {0.0, 0.0, 0.0, 0.0};
    float __attribute__ ((aligned(16))) d[4] = {0.0, 0.0, 0.0, 0.0};
    clk0 = rdtsc();
    for (i = 0; i < LOOP; ++i) 
    {
        __asm__ volatile (
            "movups %0, %%xmm4\n\t"
            "shufps $0, %%xmm4, %%xmm4\n\t"
            "mulps  %%xmm4, %%xmm0\n\t"
            "addps  %%xmm0, %%xmm8\n\t"
            "movups %1, %%xmm5\n\t"
            "shufps $0, %%xmm5, %%xmm5\n\t"
            "mulps  %%xmm5, %%xmm1\n\t"
            "addps  %%xmm1, %%xmm9\n\t"
            "movups %2, %%xmm6\n\t"
            "shufps $0, %%xmm6, %%xmm6\n\t"
            "mulps  %%xmm6, %%xmm2\n\t"
            "addps  %%xmm2, %%xmm10\n\t"
            "movups %3, %%xmm7\n\t"
            "shufps $0, %%xmm7, %%xmm7\n\t"
            "mulps  %%xmm7, %%xmm3\n\t"
            "addps  %%xmm3, %%xmm11\n\t"
            "movups %0, %%xmm4\n\t"
            "shufps $0, %%xmm4, %%xmm4\n\t"
            "mulps  %%xmm4, %%xmm0\n\t"
            "addps  %%xmm0, %%xmm8\n\t"
            "movups %1, %%xmm5\n\t"
            "shufps $0, %%xmm5, %%xmm5\n\t"
            "mulps  %%xmm5, %%xmm1\n\t"
            "addps  %%xmm1, %%xmm9\n\t"
            "movups %2, %%xmm6\n\t"
            "shufps $0, %%xmm6, %%xmm6\n\t"
            "mulps  %%xmm6, %%xmm2\n\t"
            "addps  %%xmm2, %%xmm10\n\t"
            "movups %3, %%xmm7\n\t"
            "shufps $0, %%xmm7, %%xmm7\n\t"
            "mulps  %%xmm7, %%xmm3\n\t"
            "addps  %%xmm3, %%xmm11\n\t"
            :
            :"m"(a[0]), "m"(b[0]), "m"(c[0]), "m"(d[0])
        );
    }
    clk1 = rdtsc();
    cycles = clk1 - clk0;
    print_GFLOPS(flops, cycles);
    print_throughput(instructions_per_loop * LOOP, cycles);
}

int main() {
    sse_mulps_addps_movups_shufps_fromL1();
    return 0;
}

実行結果

-- sse_mulps_addps_movups_shufps_fromL1 --
GFLOPS @ 2.00GHz:
  7.522 [flops/clock] = 15.044 [GFLOPS]  (134217728 flops in 17843454 clock = 0.008922 sec)
Throughput:
  1.880 [instructions/clock]   (33554432 instrucions in 17843454 clock)

成果物

以上。