avxfma.h
#pragma once #include "../../commonfiles/miscdefs.h" // These functions are defined in AvxFma.cpp extern void AvxFmaSmooth5Cpp(float* y, const float*x, Uint32 n, const float* sm5_mask); extern bool AvxFmaInitX(float* x, Uint32 n); // These functions are defined in AvxFma_.asm extern "C" void AvxFmaSmooth5a(float* y, const float*x, Uint32 n, const float* sm5_mask); extern "C" void AvxFmaSmooth5b(float* y, const float*x, Uint32 n, const float* sm5_mask); extern "C" void AvxFmaSmooth5c(float* y, const float*x, Uint32 n, const float* sm5_mask); // next function is removed because of windows-only functions. In time I need to figure out // how to reconstruct this function for g++ on Linux // These functions are defined in AvxFmaTimed.cpp //extern void AvxFmaTimed(void);
main.cpp
#include "avxfma.h" #include <stdio.h> #include <malloc.h> #define _USE_MATH_DEFINES #include <math.h> #include <stdlib.h> float* aligned_malloc(int n, int alignment){ void* ptrx; int rc = posix_memalign(&ptrx,alignment,n*sizeof(float)); return (float*)ptrx; } bool AvxFmaInitX(float* x, Uint32 n) { const float degtorad = (float)(M_PI / 180.0); const float t_start = 0; const float t_step = 0.002f; const Uint32 m = 3; const float amp[m] = {1.0f, 0.80f, 1.20f}; const float freq[m] = {5.0f, 10.0f, 15.0f}; const float phase[m] = {0.0f, 45.0f, 90.0f}; float t = t_start; srand(97); for (Uint32 i = 0; i < n; i++, t += t_step) { float x_total = 0; for (Uint32 j = 0; j < m; j++) { float omega = 2.0f * (float)M_PI * freq[j]; float x_temp1 = amp[j] * sin(omega * t + phase[j] * degtorad); float noise = (float)((rand() % 300) - 150) / 10.0f; float x_temp2 = x_temp1 + x_temp1 * noise / 100.0f; x_total += x_temp2; } x[i] = x_total; } return true; } void AvxFmaSmooth5Cpp(float* y, const float*x, Uint32 n, const float* sm5_mask) { for (Uint32 i = 2; i < n - 2; i++) { float sum = 0; sum += x[i - 2] * sm5_mask[0]; sum += x[i - 1] * sm5_mask[1]; sum += x[i + 0] * sm5_mask[2]; sum += x[i + 1] * sm5_mask[3]; sum += x[i + 2] * sm5_mask[4]; y[i] = sum; } } void AvxFma(void) { const Uint32 n = 512; float* x = aligned_malloc(n, 32); float* y_cpp = aligned_malloc(n, 32); float* y_a = aligned_malloc(n, 32); float* y_b = aligned_malloc(n, 32); float* y_c = aligned_malloc(n, 32); const float sm5_mask[] = { 0.0625f, 0.25f, 0.375f, 0.25f, 0.0625f }; printf("Results for AvxFma\n"); if (!AvxFmaInitX(x, n)) { printf("Data initialization failed\n"); return; } AvxFmaSmooth5Cpp(y_cpp, x, n, sm5_mask); AvxFmaSmooth5a(y_a, x, n, sm5_mask); AvxFmaSmooth5b(y_b, x, n, sm5_mask); AvxFmaSmooth5b(y_c, x, n, sm5_mask); FILE* fp; const char* fn = "AvxFmaRawData.csv"; fp=fopen(fn, "wr"); if (fp == NULL) { printf("File open failed (%s)\n", fn); return; } fprintf(fp, "i, x, y_cpp, y_a, y_b, y_c, "); fprintf(fp, "diff_ab, diff_ac, diff_bc\n"); Uint32 num_diff_ab = 0, num_diff_ac = 0, num_diff_bc = 0; for (Uint32 i = 2; i < n - 2; i++) { bool diff_ab = false, diff_ac = false, diff_bc = false; if (y_a[i] != y_b[i]) { diff_ab = true; num_diff_ab++; } if (y_a[i] != y_c[i]) { diff_ac = true; num_diff_ac++; } if (y_b[i] != y_c[i]) { diff_bc = true; num_diff_bc++; } const char* fs1 = "%15.8f, "; fprintf(fp, "%4d, ", i); fprintf(fp, fs1, x[i]); fprintf(fp, fs1, y_cpp[i]); fprintf(fp, fs1, y_a[i]); fprintf(fp, fs1, y_b[i]); fprintf(fp, fs1, y_c[i]); fprintf(fp, "%d, %d, %d, ", diff_ab, diff_ac, diff_bc); fprintf(fp, "\n"); } printf("\nRaw data saved to file %s\n", fn); printf("\nNumber of data point differences between\n"); printf(" y_a and y_b: %u\n", num_diff_ab); printf(" y_a and y_c: %u\n", num_diff_ac); printf(" y_b and y_c: %u\n", num_diff_bc); fclose(fp); free(x); free(y_cpp); free(y_a); free(y_b); free(y_c); } int main(int argc, char* argv[]) { try { AvxFma(); //AvxFmaTimed(); } catch (...) { printf("Unexpected exception has occurred!\n"); printf("File: %s (main)\n", __FILE__); } return 0; }
avxfma.asm
; Name: avxfma.asm ; ; Build: g++ -c -m32 main.cpp -o main.o ; nasm -f elf32 -o avxfma.o avxfma.asm ; g++ -m32 -o avxfma avxfma.o main.o ; ; Source: Modern x86 Assembly Language Programming p. 470 global AvxFmaSmooth5a global AvxFmaSmooth5b global AvxFmaSmooth5c section .text ; void AvxFmaSmooth5a(float* y, const float*x, Uint32 n, const float* sm5_mask); ; ; Description: The following function applies a weighted-average ; smoothing transformation to the input array x using ; scalar SPFP multiplication and addition. ; ; Requires: AVX %define y [ebp+8] %define x [ebp+12] %define n [ebp+16] %define sm5_mask [ebp+20] AvxFmaSmooth5a: push ebp mov ebp,esp push esi push edi ; Load argument values mov edi,y ;edi = ptr to y mov esi,x ;esi = ptr to x mov ecx,n ;ecx = n mov eax,sm5_mask ;eax = ptr to sm5_mask add esi,8 ;adjust pointers and add edi,8 ;counter to skip first 2 sub ecx,4 ;and last 2 elements align 16 ; Apply smoothing operator to each element of x .@1: vxorps xmm6,xmm6,xmm6 ;x_total=0 ; Compute x_total += x[i-2]*sm5_mask[0] vmovss xmm0,dword [esi-8] vmulss xmm1,xmm0,dword [eax] vaddss xmm6,xmm6,xmm1 ; Compute x_total += x[i-1]*sm5_mask[1] vmovss xmm2,dword [esi-4] vmulss xmm3,xmm2,dword [eax+4] vaddss xmm6,xmm6,xmm3 ; Compute x_total += x[i]*sm5_mask[2] vmovss xmm0,dword [esi] vmulss xmm1,xmm0,dword[eax+8] vaddss xmm6,xmm6,xmm1 ; Compute x_total += x[i+1]*sm5_mask[3] vmovss xmm2,dword[esi+4] vmulss xmm3,xmm2,dword[eax+12] vaddss xmm6,xmm6,xmm3 ; Compute x_total += x[i+2]*sm5_mask[4] vmovss xmm0,dword[esi+8] vmulss xmm1,xmm0,dword[eax+16] vaddss xmm6,xmm6,xmm1 ; Save x_total vmovss dword[edi],xmm6 add esi,4 add edi,4 sub ecx,1 jnz .@1 pop edi pop esi pop ebp ret ; void AvxFmaSmooth5b(float* y, const float*x, Uint32 n, const float* sm5_mask); ; ; Description: The following function applies a weighted-average ; smoothing transformation to the input array x using ; scalar SPFP fused-multiply-add operations. ; ; Requires: AVX2, FMA %define y [ebp+8] %define x [ebp+12] %define n [ebp+16] %define sm5_mask [ebp+20] AvxFmaSmooth5b: push ebp mov ebp,esp push esi push edi ; Load argument values mov edi,y ;edi = ptr to y mov esi,x ;esi = ptr to x mov ecx,n ;ecx = n mov eax,sm5_mask ;eax = ptr to sm5_mask add esi,8 ;adjust pointers and add edi,8 ;counter to skip first 2 sub ecx,4 ;and last 2 elements align 16 ; Apply smoothing operator to each element of x .@1: vxorps xmm6,xmm6,xmm6 ;set x_total1 = 0 vxorps xmm7,xmm7,xmm7 ;set x_total2 = 0 ; Compute x_total1 = x[i-2] * sm5_mask[0] + x_total1 vmovss xmm0,dword[esi-8] vfmadd231ss xmm6,xmm0,dword[eax] ; Compute x_total2 = x[i-1] * sm5_mask[1] + x_total2 vmovss xmm2,dword[esi-4] vfmadd231ss xmm7,xmm2,dword[eax+4] ; Compute x_total1 = x[i] * sm5_mask[2] + x_total1 vmovss xmm0,dword[esi] vfmadd231ss xmm6,xmm0,dword[eax+8] ; Compute x_total2 = x[i+1] * sm5_mask[3] + x_total2 vmovss xmm2,dword[esi+4] vfmadd231ss xmm7,xmm2,dword[eax+12] ; Compute x_total1 = x[i+2] * sm5_mask[4] + x_total1 vmovss xmm0,dword[esi+8] vfmadd231ss xmm6,xmm0,dword[eax+16] ; Compute final x_total and save result vaddss xmm5,xmm6,xmm7 vmovss dword[edi],xmm5 add esi,4 add edi,4 sub ecx,1 jnz .@1 pop edi pop esi pop ebp ret ; void AvxFmaSmooth5c(float* y, const float*x, Uint32 n, const float* sm5_mask); ; ; Description: The following function applies a weighted-average ; smoothing transformation to the input array x using ; scalar SPFP fused-multiply-add operations. ; ; Requires: AVX2, FMA %define y [ebp+8] %define x [ebp+12] %define n [ebp+16] %define sm5_mask [ebp+20] AvxFmaSmooth5c: push ebp mov ebp,esp push esi push edi ; Load argument values mov edi,y ;edi = ptr to y mov esi,x ;esi = ptr to x mov ecx,n ;ecx = n mov eax,sm5_mask ;eax = ptr to sm5_mask add esi,8 ;adjust pointers and add edi,8 ;counter to skip first 2 sub ecx,4 ;and last 2 elements align 16 ; Apply smoothing operator to each element of x, save result to y .@1: vxorps xmm6,xmm6,xmm6 ;set x_total = 0 ; Compute x_total = x[i-2] * sm5_mask[0] + x_total vmovss xmm0,dword[esi-8] vfmadd231ss xmm6,xmm0,dword[eax] ; Compute x_total = x[i-1] * sm5_mask[1] + x_total vmovss xmm0,dword[esi-4] vfmadd231ss xmm6,xmm0,dword[eax+4] ; Compute x_total = x[i] * sm5_mask[2] + x_total vmovss xmm0,dword[esi] vfmadd231ss xmm6,xmm0,dword[eax+8] ; Compute x_total = x[i+1] * sm5_mask[3] + x_total vmovss xmm0,dword[esi+4] vfmadd231ss xmm6,xmm0,dword[eax+12] ; Compute x_total = x[i+2] * sm5_mask[4] + x_total vmovss xmm0,dword[esi+8] vfmadd231ss xmm6,xmm0,dword[eax+16] ; Save result vmovss dword[edi],xmm6 add esi,4 add edi,4 sub ecx,1 jnz .@1 pop edi pop esi pop ebp ret
build
g++ -c -m32 main.cpp -o main.o -std=c++11 nasm -f elf32 -o avxfma.o avxfma.asm g++ -m32 -o avxfma avxfma.o main.o