main.cpp
#include "../../commonfiles/ymmval.h" extern "C" void AvxPfpArithmeticFloat(const YmmVal* a, const YmmVal* b, YmmVal c[6]); extern "C" void AvxPfpArithmeticDouble(const YmmVal* a, const YmmVal* b, YmmVal c[5]); void AvxPfpArithmeticFloatCpp(void) { __attribute__ ((aligned(32))) YmmVal a; __attribute__ ((aligned(32))) YmmVal b; __attribute__ ((aligned(32))) YmmVal c[6]; a.r32[0] = 2.0f; b.r32[0] = 12.5f; a.r32[1] = 3.5f; b.r32[1] = 52.125f; a.r32[2] = -10.75f; b.r32[2] = 17.5f; a.r32[3] = 15.0f; b.r32[3] = 13.982f; a.r32[4] = -12.125f; b.r32[4] = -4.75f; a.r32[5] = 3.875f; b.r32[5] = 3.0625f; a.r32[6] = 2.0f; b.r32[6] = 7.875f; a.r32[7] = -6.35f; b.r32[7] = -48.1875f; AvxPfpArithmeticFloat(&a, &b, c); printf("Results for AvxPfpArithmeticFloat()\n\n"); printf(" i a b Add Sub Mul Div Abs Neg\n"); printf("--------------------------------------------------------------------------\n"); for (int i = 0; i < 8; i++) { const char* fs = "%8.3f "; printf("%2d ", i); printf(fs, a.r32[i]); printf(fs, b.r32[i]); printf(fs, c[0].r32[i]); printf(fs, c[1].r32[i]); printf(fs, c[2].r32[i]); printf(fs, c[3].r32[i]); printf(fs, c[4].r32[i]); printf(fs, c[5].r32[i]); printf("\n"); } } void AvxPfpArithmeticDoubleCpp(void) { __attribute__ ((aligned(32))) YmmVal a; __attribute__ ((aligned(32))) YmmVal b; __attribute__ ((aligned(32))) YmmVal c[5]; a.r64[0] = 12.0; b.r64[0] = 0.875; a.r64[1] = 13.5; b.r64[1] = -125.25; a.r64[2] = 18.75; b.r64[2] = 72.5; a.r64[3] = 5.0; b.r64[3] = -98.375; AvxPfpArithmeticDouble(&a, &b, c); printf("\n\nResults for AvxPfpArithmeticDouble()\n\n"); printf(" i a b Min Max Sqrt a HorAdd HorSub\n"); printf("--------------------------------------------------------------------------\n"); for (int i = 0; i < 4; i++) { const char* fs = "%9.3lf "; printf("%2d ", i); printf(fs, a.r64[i]); printf(fs, b.r64[i]); printf(fs, c[0].r64[i]); printf(fs, c[1].r64[i]); printf(fs, c[2].r64[i]); printf(fs, c[3].r64[i]); printf(fs, c[4].r64[i]); printf("\n"); } } int main(int argc, char* argv[]) { AvxPfpArithmeticFloatCpp(); AvxPfpArithmeticDoubleCpp(); return 0; }
avxpackedfloatingpointarithmetic.asm
; Name: avxpackedfloatingpointarithmetic.asm ; ; Build: g++ -c -m32 main.cpp -o main.o ; nasm -f elf32 -o avxpackedfloatingpointarithmetic.o avxpackedfloatingpointarithmetic.asm ; g++ -m32 -o avxpackedfloatingpointarithmetic avxpackedfloatingpointarithmetic.o main.o ../../commonfiles/ymmval.o ; ; Source: Modern x86 Assembly Language Programming p. 378 global AvxPfpArithmeticFloat global AvxPfpArithmeticDouble section .data align 16 ; Mask value for packed SPFP absolute value AbsMask dd 0x7fffffff,0x7fffffff,0x7fffffff,0x7fffffff dd 0x7fffffff,0x7fffffff,0x7fffffff,0x7fffffff ; Mask value for packed SPFP negation NegMask dd 0x80000000,0x80000000,0x80000000,0x80000000 dd 0x80000000,0x80000000,0x80000000,0x80000000 section .text ; extern "C" void AvxPfpArithmeticFloat_(const YmmVal* a, const YmmVal* b, YmmVal c[6]); ; ; Description: The following function illustrates how to use common ; packed SPFP arithmetic instructions using the YMM ; registers. ; ; Requires: AVX %define a [ebp+8] %define b [ebp+12] %define c [ebp+16] AvxPfpArithmeticFloat: push ebp mov ebp,esp ; Load argument values. Note that the vmovaps instruction ; requires proper aligment of operands in memory. mov eax,a ;eax = ptr to a mov ecx,b ;ecx = ptr to b mov edx,c ;edx = ptr to c vmovaps ymm0,[eax] ;ymm0 = a vmovaps ymm1,[ecx] ;ymm1 = b ; Perform packed SPFP addition, subtraction, multiplication, ; and division vaddps ymm2,ymm0,ymm1 ;a + b vmovaps [edx],ymm2 vsubps ymm3,ymm0,ymm1 ;a - b vmovaps [edx+32],ymm3 vmulps ymm4,ymm0,ymm1 ;a * b vmovaps [edx+64],ymm4 vdivps ymm5,ymm0,ymm1 ;a / b vmovaps [edx+96],ymm5 ; Compute packed SPFP absolute value vmovups ymm6,[AbsMask] ;ymm6 = AbsMask vandps ymm7,ymm0,ymm6 ;ymm7 = packed fabs vmovaps [edx+128],ymm7 ; Compute packed SPFP negation vxorps ymm7,ymm0,[NegMask] ;ymm7 = packed neg. vmovaps [edx+160],ymm7 ; Zero upper 128-bit of all YMM registers to avoid potential x86-AVX ; to x86-SSE transition penalties. vzeroupper pop ebp ret ; extern "C" void AvxPfpArithmeticDouble_(const YmmVal* a, const YmmVal* b, YmmVal c[5]); ; ; Description: The following function illustrates how to use common ; packed DPFP arithmetic instructions using the YMM ; registers. ; ; Requires: AVX %define a [ebp+8] %define b [ebp+12] %define c [ebp+16] AvxPfpArithmeticDouble: push ebp mov ebp,esp ; Load argument values. Note that the vmovapd instruction ; requires proper aligment of operands in memory. mov eax,a ;eax = ptr to a mov ecx,b ;ecx = ptr to b mov edx,c ;edx = ptr to c vmovapd ymm0,[eax] ;ymm0 = a vmovapd ymm1,[ecx] ;ymm1 = b ; Compute packed min, max and square root vminpd ymm2,ymm0,ymm1 vmaxpd ymm3,ymm0,ymm1 vsqrtpd ymm4,ymm0 ; Perform horizontal addition and subtraction vhaddpd ymm5,ymm0,ymm1 vhsubpd ymm6,ymm0,ymm1 ; Save the results vmovapd [edx],ymm2 vmovapd [edx+32],ymm3 vmovapd [edx+64],ymm4 vmovapd [edx+96],ymm5 vmovapd [edx+128],ymm6 ; Zero upper 128-bit of all YMM registers to avoid potential x86-AVX ; to x86-SSE transition penalties. vzeroupper pop ebp ret
build
g++ -c -m32 main.cpp -o main.o nasm -f elf32 -o avxpackedfloatingpointarithmetic.o avxpackedfloatingpointarithmetic.asm g++ -m32 -o avxpackedfloatingpointarithmetic avxpackedfloatingpointarithmetic.o main.o ../../commonfiles/ymmval.o