main.h
//modifications #pragma once #include "../../commonfiles/mat4x4.h" // SsePackedFloatingPointMatrix4x4.cpp function declarations //extern void SsePfpMatrix4x4MultiplyCpp(Mat4x4 m_des, Mat4x4 m_src1, Mat4x4 m_src2); //extern void SsePfpMatrix4x4TransformVectorsCpp(Vec4x1* v_des, Mat4x4 m_src, Vec4x1* v_src, int num_vec); // SsePackedFloatingPointMatrix4x4_.asm function declarations extern "C" void SsePfpMatrix4x4Multiply(Mat4x4 m_des, Mat4x4 m_src1, Mat4x4 m_src2); extern "C" void SsePfpMatrix4x4TransformVectors(Vec4x1* v_des, Mat4x4 m_src, Vec4x1* v_src, int num_vec);
main.cpp
#include "main.h" // The functions Mat4x4Mul and Mat4x4MulVec are defined in // the file CommonFiles\Mat4x4.cpp /* void SsePfpMatrix4x4MultiplyCpp(Mat4x4 m_des, Mat4x4 m_src1, Mat4x4 m_src2) { Mat4x4Mul(m_des, m_src1, m_src2); } void SsePfpMatrix4x4TransformVectorsCpp(Vec4x1* v_des, Mat4x4 m_src, Vec4x1* v_src, int num_vec) { for (int i= 0; i < num_vec; i++) Mat4x4MulVec(v_des[i], m_src, v_src[i]); } */ void SsePfpMatrix4x4MultiplyCpp(void) { __attribute__ ((aligned(16))) Mat4x4 m_src1; __attribute__ ((aligned(16))) Mat4x4 m_src2; __attribute__ ((aligned(16))) Mat4x4 m_des1; __attribute__ ((aligned(16))) Mat4x4 m_des2; Mat4x4SetRow(m_src1, 0, 10.5, 11, 12, -13.625); Mat4x4SetRow(m_src1, 1, 14, 15, 16, 17.375); Mat4x4SetRow(m_src1, 2, 18.25, 19, 20.125, 21); Mat4x4SetRow(m_src1, 3, 22, 23.875, 24, 25); Mat4x4SetRow(m_src2, 0, 7, 1, 4, 8); Mat4x4SetRow(m_src2, 1, 14, -5, 2, 9); Mat4x4SetRow(m_src2, 2, 10, 9, 3, 6); Mat4x4SetRow(m_src2, 3, 2, 11, -14, 13); //SsePfpMatrix4x4MultiplyCpp(m_des1, m_src1, m_src2); Mat4x4Mul(m_des1, m_src1, m_src2); SsePfpMatrix4x4Multiply(m_des2, m_src1, m_src2); printf("\nResults for SsePfpMatrix4x4Multiply()\n"); Mat4x4Printf(m_src1, "\nMatrix m_src1\n"); Mat4x4Printf(m_src2, "\nMatrix m_src2\n"); Mat4x4Printf(m_des1, "\nMatrix m_des1\n"); Mat4x4Printf(m_des2, "\nMatrix m_des2\n"); } void SsePfpMatrix4x4TransformVectorsCpp(void) { const int n = 8; __attribute__ ((aligned(16))) Mat4x4 m_src; __attribute__ ((aligned(16))) Vec4x1 v_src[n]; __attribute__ ((aligned(16))) Vec4x1 v_des1[n]; __attribute__ ((aligned(16))) Vec4x1 v_des2[n]; Vec4x1Set(v_src[0], 10, 10, 10, 1); Vec4x1Set(v_src[1], 10, 11, 10, 1); Vec4x1Set(v_src[2], 11, 10, 10, 1); Vec4x1Set(v_src[3], 11, 11, 10, 1); Vec4x1Set(v_src[4], 10, 10, 12, 1); Vec4x1Set(v_src[5], 10, 11, 12, 1); Vec4x1Set(v_src[6], 11, 10, 12, 1); Vec4x1Set(v_src[7], 11, 11, 12, 1); // m_src = scale(2, 3, 4) Mat4x4SetRow(m_src, 0, 2, 0, 0, 0); Mat4x4SetRow(m_src, 1, 0, 3, 0, 0); Mat4x4SetRow(m_src, 2, 0, 0, 7, 0); Mat4x4SetRow(m_src, 3, 0, 0, 0, 1); //SsePfpMatrix4x4TransformVectorsCpp(v_des1, m_src, v_src, n); for (int i= 0; i < n; i++) Mat4x4MulVec(v_des1[i], m_src, v_src[i]); SsePfpMatrix4x4TransformVectors(v_des2, m_src, v_src, n); printf("\nResults for SsePfpMatrix4x4TransformVectors()\n"); Mat4x4Printf(m_src, "Matrix m_src\n"); printf("\n"); for (int i = 0; i < n; i++) { const char* fmt = "%4s %4d: %12.6f %12.6f %12.6f %12.6f\n"; printf(fmt, "v_src ", i, v_src[i][0], v_src[i][1], v_src[i][2], v_src[i][3]); printf(fmt, "v_des1 ", i, v_des1[i][0], v_des1[i][1], v_des1[i][2], v_des1[i][3]); printf(fmt, "v_des2 ", i, v_des2[i][0], v_des2[i][1], v_des2[i][2], v_des2[i][3]); printf("\n"); } } int main(int argc, char* argv[]) { SsePfpMatrix4x4MultiplyCpp(); SsePfpMatrix4x4TransformVectorsCpp(); //SsePfpMatrix4x4MultiplyTimed(); //SsePfpMatrix4x4TransformVectorsTimed(); return 0; }
ssepackedfloatingpointmatrix4x4.asm
; Name: ssepackedfloatingpointmatrix4x4.asm ; ; Build: g++ -c -m32 main.cpp -o main.o ; nasm -f elf32 -o ssepackedfloatingpointmatrix4x4.o ssepackedfloatingpointmatrix4x4.asm ; g++ -m32 -o ssepackedfloatingpointmatrix4x4 ssepackedfloatingpointmatrix4x4.o main.o ../../commonfiles/mat4x4.o ; ; Source: Modern x86 Assembly Language Programming p. 260 global SsePfpMatrix4x4Multiply global SsePfpMatrix4x4TransformVectors ; Mat4x4Transpose macro ; ; Description: This macro computes the transpose of a 4x4 ; single-precision floating-point matrix. ; ; Input Matrix Output Matrtix ; xmm0 a3 a2 a1 a0 xmm4 d0 c0 b0 a0 ; xmm1 b3 b2 b1 b0 xmm5 d1 c1 b1 a1 ; xmm2 c3 c2 c1 c0 xmm6 d2 c2 b2 a2 ; xmm3 d3 d2 d1 d0 xmm7 d3 c3 b3 a3 ; ; Note: The row of a 4x4 matrix is reversed when loaded into an ; XMM register due to x86 little-endian ordering. ; ; Requires: SSE %MACRO Mat4x4Transpose 0 movaps xmm4,xmm0 unpcklps xmm4,xmm1 ;xmm4 = b1 a1 b0 a0 unpckhps xmm0,xmm1 ;xmm0 = b3 a3 b2 a2 movaps xmm5,xmm2 unpcklps xmm5,xmm3 ;xmm5 = d1 c1 d0 c0 unpckhps xmm2,xmm3 ;xmm2 = d3 c3 d2 c2 movaps xmm1,xmm4 movlhps xmm4,xmm5 ;xmm4 = d0 c0 b0 a0 movhlps xmm5,xmm1 ;xmm5 = d1 c1 b1 a1 movaps xmm6,xmm0 movlhps xmm6,xmm2 ;xmm6 = d2 c2 b2 a2 movaps xmm7,xmm2 movhlps xmm7,xmm0 ;xmm7 = d3 c3 b2 a3 %ENDMACRO section .text ; extern "C" void SsePfpMatrix4x4Multiply(Mat4x4 m_des, Mat4x4 m_src1, Mat4x4 m_src2); ; ; Description: The following function computes the product of two ; 4x4 single-precision floating-point matrices. ; ; Requires: SSE4.1 %define m_des [ebp+8] %define m_src1 [ebp+12] %define m_src2 [ebp+16] SsePfpMatrix4x4Multiply: push ebp mov ebp,esp push ebx ; Compute transpose of m_src2 (m_src2_T) mov ebx,m_src2 ;ebx = m_src2 movaps xmm0,[ebx] movaps xmm1,[ebx+16] movaps xmm2,[ebx+32] movaps xmm3,[ebx+48] ;xmm3:xmm0 = m_src2 Mat4x4Transpose ;xmm7:xmm4 = m_src2_T ; Perform initializations for matrix product mov edx,m_des ;edx = m_des mov ebx,m_src1 ;ebx = m_src1 mov ecx,4 ;ecx = number of rows xor eax,eax ;eax = offset into arrays ; Repeat loop until matrix product is calculated. align 16 .@1: movaps xmm0,[ebx+eax] ;xmm0 = row i of m_src1 ; Compute dot product of m_src1 row i and m_src2_T row 0 movaps xmm1,xmm0 dpps xmm1,xmm4,11110001b ;xmm1[31:0] = dot product insertps xmm3,xmm1,00000000b ;xmm3[31:0] = xmm1[31:0] ; Compute dot product of m_src1 row i and m_src2_T row 1 movaps xmm2,xmm0 dpps xmm2,xmm5,11110001b ;xmm2[31:0] = dot product insertps xmm3,xmm2,00010000b ;xmm3[63:32] = xmm2[31:0] ; Compute dot product of m_src1 row i and m_src2_T row 2 movaps xmm1,xmm0 dpps xmm1,xmm6,11110001b ;xmm1[31:0] = dot product insertps xmm3,xmm1,00100000b ;xmm3[95:64] = xmm1[31:0] ; Compute dot product of m_src1 row i and m_src2_T row 3 movaps xmm2,xmm0 dpps xmm2,xmm7,11110001b ;xmm2[31:0] = dot product insertps xmm3,xmm2,00110000b ;xmm3[127:96] = xmm2[31:0] ; Save m_des.row i and update loop variables movaps [edx+eax],xmm3 ;save current row result add eax,16 ;set array offset to next row dec ecx jnz .@1 pop ebx pop ebp ret ; extern void SsePfpMatrix4x4TransformVectors(Vec4x1* v_des, Mat4x4 m_src, Vec4x1* v_src, int num_vec); ; ; Description: The following function applies a transformation matrix ; to an array 4x1 single-precision floating-point vectors. ; ; Requires: SSE4.1 %define v_des [ebp+8] %define m_src [ebp+12] %define v_src [ebp+16] %define num_vec [ebp+20] SsePfpMatrix4x4TransformVectors: push ebp mov ebp,esp push esi push edi ; Make sure num_vec is valid mov ecx,num_vec ;ecx = num_vec test ecx,ecx jle .done ;jump if num_vec <= 0 ; Load m_src into xmm3:xmm0 mov eax,m_src ;eax = pointer to m_src movaps xmm0,[eax] ;xmm0 = row 0 movaps xmm1,[eax+16] ;xmm1 = row 1 movaps xmm2,[eax+32] ;xmm2 = row 2 movaps xmm3,[eax+48] ;xmm3 = row 3 ; Initialize pointers to v_src and v_des mov esi,v_src ;esi = pointer to v_src mov edi,v_des ;edi = pointer to v_des xor eax,eax ;eax = array offset ; Compute v_des[i] = m_src * v_src[i] align 16 .@1: movaps xmm4,[esi+eax] ;xmm4 = vector v_src[i] ; Compute dot product of m_src row 0 and v_src[i] movaps xmm5,xmm4 dpps xmm5,xmm0,11110001b ;xmm5[31:0] = dot product insertps xmm7,xmm5,00000000b ;xmm7[31:0] = xmm5[31:0] ; Compute dot product of m_src row 1 and v_src[i] movaps xmm6,xmm4 dpps xmm6,xmm1,11110001b ;xmm6[31:0] = dot product insertps xmm7,xmm6,00010000b ;xmm7[63:32] = xmm6[31:0] ; Compute dot product of m_src row 2 and v_src[i] movaps xmm5,xmm4 dpps xmm5,xmm2,11110001b ;xmm5[31:0] = dot product insertps xmm7,xmm5,00100000b ;xmm7[95:64] = xmm5[31:0] ; Compute dot product of m_src row 3 and v_src[i] movaps xmm6,xmm4 dpps xmm6,xmm3,11110001b ;xmm6[31:0] = dot product insertps xmm7,xmm6,00110000b ;xmm7[127:96] = xmm6[31:0] ; Save v_des[i] and update loop variables movaps [edi+eax],xmm7 ;save transformed vector add eax,16 dec ecx jnz .@1 .done: pop edi pop esi pop ebp ret
build
g++ -c -m32 main.cpp -o main.o nasm -f elf32 -o ssepackedfloatingpointmatrix4x4.o ssepackedfloatingpointmatrix4x4.asm g++ -m32 -o ssepackedfloatingpointmatrix4x4 ssepackedfloatingpointmatrix4x4.o main.o ../../commonfiles/mat4x4.o