main.cpp
#include <stdio.h> #include "../../commonfiles/xmmval.h" #include "../../commonfiles/ymmval.h" #include <stdlib.h> extern "C" void AvxGatherFloat(YmmVal* des, YmmVal* indices, YmmVal* mask, const float* x); extern "C" void AvxGatherI64(YmmVal* des, XmmVal* indices, YmmVal* mask, const Int64* x); void AvxGatherFloatPrint(const char* msg, YmmVal& des, YmmVal& indices, YmmVal& mask) { printf("\n%s\n", msg); for (int i = 0; i < 8; i++) { printf("ElementID: %d ", i); printf("des: %8.1f ", des.r32[i]); printf("indices: %4d ", indices.i32[i]); printf("mask: 0x%08X\n", mask.i32[i]); } } void AvxGatherI64Print(const char* msg, YmmVal& des, XmmVal& indices, YmmVal& mask) { printf("\n%s\n", msg); for (int i = 0; i < 4; i++) { printf("ElementID: %d ", i); printf("des: %8lld ", des.i64[i]); printf("indices: %4d ", indices.i32[i]); printf("mask: 0x%016llX\n", mask.i64[i]); } } void AvxGatherFloatCpp(void) { const int merge_no = 0; const int merge_yes = 0x80000000; const int n = 15; float x[n]; __attribute__((aligned(32))) YmmVal des; __attribute__((aligned(32))) YmmVal indices; __attribute__((aligned(32))) YmmVal mask; // Initialize the test array srand(22); for (int i = 0; i < n; i++) x[i] = (float)(rand() % 1000); // Load des with initial values for (int i = 0; i < 8; i++) des.r32[i] = -1.0f; // Initialize the indices indices.i32[0] = 2; indices.i32[1] = 1; indices.i32[2] = 6; indices.i32[3] = 5; indices.i32[4] = 4; indices.i32[5] = 13; indices.i32[6] = 11; indices.i32[7] = 9; // Initialize the mask value mask.i32[0] = merge_yes; mask.i32[1] = merge_yes; mask.i32[2] = merge_no; mask.i32[3] = merge_yes; mask.i32[4] = merge_yes; mask.i32[5] = merge_no; mask.i32[6] = merge_yes; mask.i32[7] = merge_yes; printf("\nResults for AvxGatherFloat()\n"); printf("Test array\n"); for (int i = 0; i < n; i++) printf("x[%02d]: %6.1f\n", i, x[i]); printf("\n"); const char* s1 = "Values BEFORE call to AvxGatherFloat()"; const char* s2 = "Values AFTER call to AvxGatherFloat()"; AvxGatherFloatPrint(s1, des, indices, mask); AvxGatherFloat(&des, &indices, &mask, x); AvxGatherFloatPrint(s2, des, indices, mask); } void AvxGatherI64Cpp(void) { const Int64 merge_no = 0; const Int64 merge_yes = 0x8000000000000000LL; const int n = 15; Int64 x[n]; __attribute__((aligned(32))) YmmVal des; __attribute__((aligned(16))) XmmVal indices; __attribute__((aligned(32))) YmmVal mask; // Initialize the test array srand(36); for (int i = 0; i < n; i++) x[i] = (Int64)(rand() % 1000); // Load des with initial values for (int i = 0; i < 4; i++) des.i64[i] = -1; // Initialize the indices and mask elements indices.i32[0] = 2; indices.i32[1] = 7; indices.i32[2] = 9; indices.i32[3] = 12; mask.i64[0] = merge_yes; mask.i64[1] = merge_yes; mask.i64[2] = merge_no; mask.i64[3] = merge_yes; printf("\nResults for AvxGatherI64()\n"); printf("Test array\n"); for (int i = 0; i < n; i++) printf("x[%02d]: %8lld\n", i, x[i]); printf("\n"); const char* s1 = "Values BEFORE call to AvxGatherI64()"; const char* s2 = "Values AFTER call to AvxGatherI64()"; AvxGatherI64Print(s1, des, indices, mask); AvxGatherI64(&des, &indices, &mask, x); AvxGatherI64Print(s2, des, indices, mask); } int main(int argc, char* argv[]) { AvxGatherFloatCpp(); AvxGatherI64Cpp(); return 0; }
avxgather.asm
; Name: avxgather.asm ; ; Build: g++ -c -m32 main.cpp -o main.o ; nasm -f elf32 -o avxgather.o avxgather.asm ; g++ -m32 -o avxgather avxgather.o main.o ; ; Source: Modern x86 Assembly Language Programming p. 463 global AvxGatherFloat global AvxGatherI64 section .text ; extern "C" void AvxGatherFloat(YmmVal* des, YmmVal* indices, YmmVal* mask, const float* x); ; ; Description: The following function demonstrates use of the ; vgatherdps instruction. ; ; Requires: AVX2 %define des [ebp+8] %define indices [ebp+12] %define mask [ebp+16] %define x [ebp+20] AvxGatherFloat: push ebp mov ebp,esp push ebx ; Load argument values. The contents of des are loaded into ymm0 ; prior to execution of the vgatherdps instruction in order to ; demonstrate the conditional effects of the control mask. mov eax,des ;eax = ptr to des mov ebx,indices ;ebx = ptr to indices mov ecx,mask ;ecx = ptr to mask mov edx,x ;edx = ptr to x vmovaps ymm0,[eax] ;ymm0 = des (initial values) vmovdqa ymm1,[ebx] ;ymm1 = indices vmovdqa ymm2,[ecx] ;ymm2 = mask ; Perform the gather operation and save the results. vgatherdps ymm0,[edx+ymm1*4],ymm2 ;ymm0 = gathered elements vmovaps [eax],ymm0 ;save des vmovdqa [ebx],ymm1 ;save indices (unchanged) vmovdqa [ecx],ymm2 ;save mask (all zeros) vzeroupper pop ebx pop ebp ret ; extern "C" void AvxGatherI64(YmmVal* des, XmmVal* indices, YmmVal* mask, const Int64* x); ; ; Description: The following function demonstrates use of the vpgatherdq ; instruction. ; ; Requires: AVX2 %define des [ebp+8] %define indices [ebp+12] %define mask [ebp+16] %define x [ebp+20] AvxGatherI64: push ebp mov ebp,esp push ebx ; Load argument values. Note that the indices are loaded ; into register XMM1. mov eax,des ;eax = ptr to des mov ebx,indices ;ebx = ptr to indices mov ecx,mask ;ecx = ptr to mask mov edx,x ;edx = ptr to x vmovdqa ymm0,[eax] ;ymm0 = des (initial values) vmovdqa xmm1,[ebx] ;xmm1 = indices vmovdqa ymm2,[ecx] ;ymm2 = mask ; Perform the gather and save the results. Note that the ; scale factor matches the size of the gathered elements. vpgatherdq ymm0,[edx+xmm1*8],ymm2 ;ymm0 = gathered elements vmovdqa [eax],ymm0 ;save des vmovdqa [ebx],xmm1 ;save indices (unchanged) vmovdqa [ecx],ymm2 ;save mask (all zeros) vzeroupper pop ebx pop ebp ret
build
g++ -c -m32 main.cpp -o main.o nasm -f elf32 -o avxgather.o avxgather.asm g++ -m32 -o avxgather avxgather.o main.o ../../commonfiles/xmmval.o ../../commonfiles/ymmval.o