[PD-cvs] pd/src m_simd_sse_gcc.h,1.1.2.9,1.1.2.10 m_simd_sse_gcc.c,1.1.2.5,1.1.2.6

Update of /cvsroot/pure-data/pd/src In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv31156

Modified Files: Tag: devel_0_37 m_simd_sse_gcc.h m_simd_sse_gcc.c Log Message: port of thomas' sse optimized code to gcc inline assembly

Index: m_simd_sse_gcc.c =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.c,v retrieving revision 1.1.2.5 retrieving revision 1.1.2.6 diff -C2 -d -r1.1.2.5 -r1.1.2.6 *** m_simd_sse_gcc.c 2 Sep 2004 14:23:01 -0000 1.1.2.5 --- m_simd_sse_gcc.c 9 Oct 2004 07:44:29 -0000 1.1.2.6 *************** *** 10,123 ****

! /* TB: adapted from the xxx_sse_vc routines ! since gcc can't access the c code as vc can, the functions itself have ! been adapted to assembler */

/* zero_perf_sse_gcc (t_int * w)*/ ! asm( ! ".set T_FLOAT,4 \n" /* sizeof(t_float) */ ! ".set T_INT,4 \n" /* sizeof(t_int) */

! /* header */ ! ".text \n" ! ".align 4 \n" /* alignment */

! ".globl zero_perf_sse_gcc \n" ! "zero_perf_sse_gcc: \n"

! /* head of function */ ! "pushl %ebp \n" ! "movl %esp, %ebp \n" ! "pushl %esi \n" ! "movl %esi, -4(%ebp) \n"

! /* get arguments */ ! "movl 8(%ebp), %esi \n" /* argument pointer */ ! "movl T_INT(%esi), %edx \n" /* out */ ! "movl 2*T_INT(%esi), %ecx \n" /* n */ ! "xorps %xmm0, %xmm0 \n" /* load zero */ ! "shrl $4, %ecx \n" /* divide by 16 */

! /* loop: *out=0 */ ! "zpsg_loop: \n" ! "movaps %xmm0, (%edx) \n" ! "movaps %xmm0, 4*T_FLOAT(%edx) \n" ! "movaps %xmm0, 8*T_FLOAT(%edx) \n" ! "movaps %xmm0, 12*T_FLOAT(%edx) \n"

! "addl $64, %edx \n" /* out+=16 */ ! "loop zpsg_loop \n"

! /* return w+3; */ ! "movl -4(%ebp), %esi \n" ! "movl 8(%ebp), %eax \n" ! "addl $12, %eax \n" ! "movl %ebp, %esp \n" ! "popl %ebp \n" ! "ret \n"

! ".align 4 \n" /* alignment */ ! ".type zero_perf_sse_gcc, @function \n"

! );

- /* copy_perf_sse_gcc (t_int * w)*/ - asm( - ".set T_FLOAT,4 \n" - ".set T_INT,4 \n"

! /* header */ ! ".text \n" ! ".align 4 \n" /* alignment */

! ".globl copy_perf_sse_gcc \n" ! "copy_perf_sse_gcc: \n"

! /* head of function */ ! "pushl %ebp \n" ! "movl %esp, %ebp \n" ! "subl $8, %esp \n" ! "movl %ebx, -4(%ebp) \n" ! "movl %esi, -8(%ebp) \n"

! /* get arguments */ ! "movl 8(%ebp), %esi \n" /* argument pointer */ ! "movl 1*T_INT(%esi), %ebx \n" /* in1 */ ! "movl 2*T_INT(%esi), %edx \n" /* out */ ! "movl 3*T_INT(%esi), %ecx \n" /* n */ ! "shrl $4, %ecx \n" /* divide by 16 */

! /* loop: *out = *in */ ! "cpsg_loop: \n" ! "movaps (%ebx), %xmm0 \n" ! "movaps %xmm0, (%edx) \n" ! "movaps 4*T_FLOAT(%ebx), %xmm1 \n" ! "movaps %xmm1, 4*T_FLOAT(%edx) \n"

! "movaps 8*T_FLOAT(%ebx), %xmm2 \n" ! "movaps %xmm2, 8*T_FLOAT(%edx) \n" ! "movaps 12*T_FLOAT(%ebx), %xmm3 \n" ! "movaps %xmm3, 12*T_FLOAT(%edx) \n"

! "addl $64, %ebx \n" /* in1 +=16 */ ! "addl $64, %edx \n" /* out +=16 */ ! "loop cpsg_loop \n"

! /* return w+4; */ ! "movl -4(%ebp), %ebx \n" ! "movl -8(%ebp), %esi \n" ! "movl 8(%ebp), %eax \n" ! "addl $16, %eax \n" ! "movl %ebp, %esp \n" ! "popl %ebp \n" ! "ret \n"

! ".align 4 \n" /* alignment */ ! ".type copy_perf_sse_gcc, @function \n" ! );

--- 10,211 ----

! /* TB: adapted from the xxx_sse_vc routines */ ! ! /* dst is assumed to be aligned */ ! static void zerovec_sse_gcc(t_float *dst,int n) ! { ! asm( ! ".set T_FLOAT,4 \n" /* sizeof(t_float) */ ! "xorps %%xmm0, %%xmm0 \n" /* zero value */ ! "shr $4, %0 \n" ! ! /* should we do more loop unrolling? */ ! "1: \n" ! "movaps %%xmm0, (%1) \n" ! "movaps %%xmm0, 4*T_FLOAT(%1) \n" ! "movaps %%xmm0, 8*T_FLOAT(%1) \n" ! "movaps %%xmm0, 12*T_FLOAT(%1) \n" ! ! "addl $64,%1 \n" ! "loop 1b \n" ! : ! :"c"(n),"r"(dst) ! :"%xmm0"); ! } ! ! /* dst is assumed to be aligned */ ! void setvec_sse_gcc(t_float *dst,t_float v,int n) ! { ! asm( ! ".set T_FLOAT,4 \n" /* sizeof(t_float) */ ! "shufps $0,%2,%2 \n" /* load value */ ! "shr $4,%0 \n" ! ! /* should we do more loop unrolling? */ ! "1: \n" ! "movaps %2, (%1) \n" ! "movaps %2, 4*T_FLOAT(%1) \n" ! "movaps %2, 8*T_FLOAT(%1) \n" ! "movaps %2, 12*T_FLOAT(%1) \n" ! ! "addl $64,%1 \n" ! "loop 1b \n" ! : ! :"c"(n),"r"(dst),"x"(v) ! ); ! } ! ! ! /* dst and src are assumed to be aligned */ ! void copyvec_sse_gcc(t_float *dst,const t_float *src,int n) ! { ! asm( ! ".set T_FLOAT,4 \n" /* sizeof(t_float) */ ! "shr $4, %0 \n" ! ! /* loop: *out = *in */ ! "1: \n" ! "movaps (%1), %%xmm0 \n" ! "movaps 4*T_FLOAT(%1), %%xmm1 \n" ! "movaps 8*T_FLOAT(%1), %%xmm2 \n" ! "movaps 16*T_FLOAT(%1), %%xmm3 \n" ! "movaps %%xmm0, (%2) \n" ! "movaps %%xmm1, 4*T_FLOAT(%2) \n" ! "movaps %%xmm2, 8*T_FLOAT(%2) \n" ! "movaps %%xmm3, 16*T_FLOAT(%2) \n" ! ! "addl $64,%1 \n" ! "addl $64,%2 \n" ! "loop 1b \n" ! : ! :"r"(n),"r"(src),"r"(dst) ! :"%xmm0","%xmm1","%xmm2","%xmm3","%eax"); ! } ! ! t_int *zero_perf_sse_gcc(t_int *w) ! { ! zerovec_sse_gcc((t_float *)w[1],w[2]); ! return w+3; ! } ! ! t_int *copy_perf_sse_gcc(t_int *w) ! { ! copyvec_sse_gcc((t_float *)w[2],(const t_float *)w[1],w[3]); ! return w+4; ! } ! ! /* t_int *sig_tilde_perf_sse_gcc(t_int *w) */ ! /* { */ ! /* setvec_sse_gcc((t_float *)w[2],*(const t_float *)w[1],w[3]); */ ! /* return w+4; */ ! /* } */

/* zero_perf_sse_gcc (t_int * w)*/ ! /* asm( */ ! /* ".set T_FLOAT,4 \n" /* sizeof(t_float) */ */ ! /* ".set T_INT,4 \n" /* sizeof(t_int) */ */

! /* /* header */ */ ! /* ".text \n" */ ! /* ".align 4 \n" /* alignment */ */

! /* ".globl zero_perf_sse_gcc \n" */ ! /* "zero_perf_sse_gcc: \n" */

! /* /* head of function */ */ ! /* "pushl %ebp \n" */ ! /* "movl %esp, %ebp \n" */ ! /* "pushl %esi \n" */ ! /* "movl %esi, -4(%ebp) \n" */

! /* /* get arguments */ */ ! /* "movl 8(%ebp), %esi \n" /* argument pointer */ */ ! /* "movl T_INT(%esi), %edx \n" /* out */ */ ! /* "movl 2*T_INT(%esi), %ecx \n" /* n */ */ ! /* "xorps %xmm0, %xmm0 \n" /* load zero */ */ ! /* "shrl $4, %ecx \n" /* divide by 16 */ */

! /* /* loop: *out=0 */ */ ! /* "zpsg_loop: \n" */ ! /* "movaps %xmm0, (%edx) \n" */ ! /* "movaps %xmm0, 4*T_FLOAT(%edx) \n" */ ! /* "movaps %xmm0, 8*T_FLOAT(%edx) \n" */ ! /* "movaps %xmm0, 12*T_FLOAT(%edx) \n" */

! /* "addl $64, %edx \n" /* out+=16 */ */ ! /* "loop zpsg_loop \n" */

! /* /* return w+3; */ */ ! /* "movl -4(%ebp), %esi \n" */ ! /* "movl 8(%ebp), %eax \n" */ ! /* "addl $12, %eax \n" */ ! /* "movl %ebp, %esp \n" */ ! /* "popl %ebp \n" */ ! /* "ret \n" */

! /* ".align 4 \n" /* alignment */ */ ! /* ".type zero_perf_sse_gcc, @function \n" */

! /* ); */

! /* /* copy_perf_sse_gcc (t_int * w)*/ */ ! /* asm( */ ! /* ".set T_FLOAT,4 \n" */ ! /* ".set T_INT,4 \n" */ ! ! /* /* header */ */ ! /* ".text \n" */ ! /* ".align 4 \n" /* alignment */ */

! /* ".globl copy_perf_sse_gcc \n" */ ! /* "copy_perf_sse_gcc: \n" */

! /* /* head of function */ */ ! /* "pushl %ebp \n" */ ! /* "movl %esp, %ebp \n" */ ! /* "subl $8, %esp \n" */ ! /* "movl %ebx, -4(%ebp) \n" */ ! /* "movl %esi, -8(%ebp) \n" */

! /* /* get arguments */ */ ! /* "movl 8(%ebp), %esi \n" /* argument pointer */ */ ! /* "movl 1*T_INT(%esi), %ebx \n" /* in1 */ */ ! /* "movl 2*T_INT(%esi), %edx \n" /* out */ */ ! /* "movl 3*T_INT(%esi), %ecx \n" /* n */ */ ! /* "shrl $4, %ecx \n" /* divide by 16 */ */

! /* /* loop: *out = *in */ */ ! /* "cpsg_loop: \n" */ ! /* "movaps (%ebx), %xmm0 \n" */ ! /* "movaps %xmm0, (%edx) \n" */ ! /* "movaps 4*T_FLOAT(%ebx), %xmm1 \n" */ ! /* "movaps %xmm1, 4*T_FLOAT(%edx) \n" */

! /* "movaps 8*T_FLOAT(%ebx), %xmm2 \n" */ ! /* "movaps %xmm2, 8*T_FLOAT(%edx) \n" */ ! /* "movaps 12*T_FLOAT(%ebx), %xmm3 \n" */ ! /* "movaps %xmm3, 12*T_FLOAT(%edx) \n" */

! /* "addl $64, %ebx \n" /* in1 +=16 */ */ ! /* "addl $64, %edx \n" /* out +=16 */ */ ! /* "loop cpsg_loop \n" */

! /* /* return w+4; */ */ ! /* "movl -4(%ebp), %ebx \n" */ ! /* "movl -8(%ebp), %esi \n" */ ! /* "movl 8(%ebp), %eax \n" */ ! /* "addl $16, %eax \n" */ ! /* "movl %ebp, %esp \n" */ ! /* "popl %ebp \n" */ ! /* "ret \n" */

! /* ".align 4 \n" /* alignment */ */ ! /* ".type copy_perf_sse_gcc, @function \n" */ ! /* ); */

*************** *** 157,161 **** "movaps %xmm0, 16(%edx) \n" "movaps %xmm0, 32(%edx) \n" ! "movaps %xmm0, 48(%edx) \n" "addl $64, %edx \n" /* out+=16 */ "loop sigtpsg_loop \n" --- 245,249 ---- "movaps %xmm0, 16(%edx) \n" "movaps %xmm0, 32(%edx) \n" ! "movaps %xmm0, 48(%edx) \n" "addl $64, %edx \n" /* out+=16 */ "loop sigtpsg_loop \n"

Index: m_simd_sse_gcc.h =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.h,v retrieving revision 1.1.2.9 retrieving revision 1.1.2.10 diff -C2 -d -r1.1.2.9 -r1.1.2.10 *** m_simd_sse_gcc.h 7 Oct 2004 17:56:38 -0000 1.1.2.9 --- m_simd_sse_gcc.h 9 Oct 2004 07:44:29 -0000 1.1.2.10 *************** *** 11,16 **** /* SIMD functions for SSE with GCC */ //void zerovec_sse_gcc(t_float *dst,int n); ! //void setvec_sse_gcc(t_float *dst,t_float v,int n); ! //void copyvec_sse_gcc(t_float *dst,const t_float *src,int n); //void addvec_sse_gcc(t_float *dst,const t_float *src,int n); //void testcopyvec_sse_gcc(t_float *dst,const t_float *src,int n); --- 11,16 ---- /* SIMD functions for SSE with GCC */ //void zerovec_sse_gcc(t_float *dst,int n); ! void setvec_sse_gcc(t_float *dst,t_float v,int n); ! void copyvec_sse_gcc(t_float *dst,const t_float *src,int n); //void addvec_sse_gcc(t_float *dst,const t_float *src,int n); //void testcopyvec_sse_gcc(t_float *dst,const t_float *src,int n); *************** *** 39,45 ****

! #define zerovec zerovec_8 /* SIMD not implemented */ ! #define setvec setvec_8 /* SIMD not implemented */ ! #define copyvec copyvec_8 /* SIMD not implemented */ #define addvec addvec_8 /* SIMD not implemented */ #define testcopyvec testcopyvec_8 /* SIMD not implemented */ --- 39,45 ----

! #define zerovec zerovec_8 //sse_gcc ! #define setvec setvec_sse_gcc ! #define copyvec copyvec_sse_gcc #define addvec addvec_8 /* SIMD not implemented */ #define testcopyvec testcopyvec_8 /* SIMD not implemented */