[PD-cvs] pd/src m_simd_sse_gcc.c,1.1.2.1,1.1.2.2 m_simd_sse_gcc.h,1.1.2.3,1.1.2.4
Tim Blechmann
timblech at users.sourceforge.net
Sat Jul 10 21:57:23 CEST 2004
Update of /cvsroot/pure-data/pd/src
In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv20643
Modified Files:
Tag: devel_0_37
m_simd_sse_gcc.c m_simd_sse_gcc.h
Log Message:
port of thomas grill's vectorized simd functions for vc to gcc
Index: m_simd_sse_gcc.c
===================================================================
RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.c,v
retrieving revision 1.1.2.1
retrieving revision 1.1.2.2
diff -C2 -d -r1.1.2.1 -r1.1.2.2
*** m_simd_sse_gcc.c 23 Dec 2003 01:15:39 -0000 1.1.2.1
--- m_simd_sse_gcc.c 10 Jul 2004 19:57:20 -0000 1.1.2.2
***************
*** 10,12 ****
--- 10,997 ----
+ /* 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"
+ );
+
+
+ /* sig_tilde_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 sig_tilde_perf_sse_gcc \n"
+ "sig_tilde_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 2*T_INT(%esi), %edx \n" /* out */
+ "movl 1*T_INT(%esi), %eax \n" /* f */
+ "movl 3*T_INT(%esi), %ecx \n" /* n */
+
+ /* set registers to f */
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "movl 12(%esi), %ecx \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = f */
+ "sigtpsg_loop: \n"
+ "movaps %xmm0, (%edx) \n"
+ "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"
+
+ /* return w+4; */
+ "movl -4(%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 sig_tilde_perf_sse_gcc, @function \n"
+
+ );
+
+
+
+ /* plus_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 plus_perf_sse_gcc \n"
+ "plus_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 vector */
+
+ "movl 1*T_INT(%esi), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset index */
+
+ /* loop: *out = *in1 + *in2 */
+ "ppsg_loop: \n"
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "addps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "addps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "addps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "addps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+
+ "addl $64, %esi \n" /* out+=16; */
+ "loop ppsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+
+ ".align 4 \n" /* alignment */
+ ".type plus_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalarplus_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 scalarplus_perf_sse_gcc \n"
+ "scalarplus_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = *in + value */
+ "sppsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "addps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "addps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "addps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "addps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop sppsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalarplus_perf_sse_gcc, @function \n"
+
+ );
+
+
+ /* minus_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 minus_perf_sse_gcc \n"
+ "minus_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), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset register */
+
+
+ /* loop: *out = *in1 - *in2 */
+ "mpsg_loop: \n"
+
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "subps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "subps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "subps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "subps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+ "addl $64, %esi \n" /* out +=16 */
+
+ "loop mpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type minus_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalarminus_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 scalarminus_perf_sse_gcc \n"
+ "scalarminus_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = *in - value */
+ "smpsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "subps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "subps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "subps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "subps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop smpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalarminus_perf_sse_gcc, @function\n"
+
+ );
+
+ /* times_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 times_perf_sse_gcc \n"
+ "times_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 vector */
+
+ "movl 1*T_INT(%esi), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset index */
+
+ /* loop: *out = *in1 * *in2 */
+ "tpsg_loop: \n"
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "mulps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "mulps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "mulps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "mulps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+
+ "addl $64, %esi \n" /* out+=16; */
+ "loop tpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+
+ ".align 4 \n" /* alignment */
+ ".type times_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalartimes_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 scalartimes_perf_sse_gcc \n"
+ "scalartimes_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = *in * n */
+ "stpsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "mulps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "mulps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "mulps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "mulps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop stpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalartimes_perf_sse_gcc, @function\n"
+
+ );
+
+ /* over_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 over_perf_sse_gcc \n"
+ "over_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 vector */
+
+ "movl 1*T_INT(%esi), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset index */
+
+ /* loop: *out = *in1 / *in2 */
+ "opsg_loop: \n"
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "divps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "divps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "divps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "divps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+
+ "addl $64, %esi \n" /* out+=16; */
+ "loop opsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+
+ ".align 4 \n" /* alignment */
+ ".type over_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalarover_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 scalarover_perf_sse_gcc \n"
+ "scalarover_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = *in / n */
+ "sopsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "divps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "divps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "divps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "divps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop sopsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalarover_perf_sse_gcc, @function \n"
+
+ );
+
+ /* min_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 min_perf_sse_gcc \n"
+ "min_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 vector */
+
+ "movl 1*T_INT(%esi), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset index */
+
+ /* loop: *out = min(*in1, *in2) */
+ "minpsg_loop: \n"
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "minps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "minps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "minps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "minps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+
+ "addl $64, %esi \n" /* out+=16; */
+ "loop minpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+
+ ".align 4 \n" /* alignment */
+ ".type min_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalarmin_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 scalarmin_perf_sse_gcc \n"
+ "scalarmin_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = min(*in,value) */
+ "sminopsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "minps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "minps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "minps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "minps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop sminopsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalarmin_perf_sse_gcc, @function \n"
+
+ );
+
+ /* max_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 max_perf_sse_gcc \n"
+ "max_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 vector */
+
+ "movl 1*T_INT(%esi), %eax \n" /* in1 */
+ "movl 2*T_INT(%esi), %ebx \n" /* in2 */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+ "shrl $4, %ecx \n" /* divide by 16 */
+ "xorl %esi, %esi \n" /* reset index */
+
+ /* loop: *out = max(*in1, *in2) */
+ "maxpsg_loop: \n"
+ "movaps (%eax,%esi), %xmm0 \n"
+ "movaps (%ebx,%esi), %xmm1 \n"
+ "maxps %xmm1, %xmm0 \n"
+ "movaps %xmm0, (%edx,%esi) \n"
+
+ "movaps 4*T_FLOAT(%eax,%esi), %xmm2 \n"
+ "movaps 4*T_FLOAT(%ebx,%esi), %xmm3 \n"
+ "maxps %xmm3, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 8*T_FLOAT(%eax,%esi), %xmm4 \n"
+ "movaps 8*T_FLOAT(%ebx,%esi), %xmm5 \n"
+ "maxps %xmm5, %xmm4 \n"
+ "movaps %xmm4, 8*T_FLOAT(%edx,%esi) \n"
+
+ "movaps 12*T_FLOAT(%eax,%esi), %xmm6 \n"
+ "movaps 12*T_FLOAT(%ebx,%esi), %xmm7 \n"
+ "maxps %xmm7, %xmm6 \n"
+ "movaps %xmm6, 12*T_FLOAT(%edx,%esi) \n"
+
+ "addl $64, %esi \n" /* out+=16; */
+ "loop maxpsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+
+ ".align 4 \n" /* alignment */
+ ".type max_perf_sse_gcc, @function \n"
+
+ );
+
+ /* scalarmax_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 scalarmax_perf_sse_gcc \n"
+ "scalarmax_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 T_INT(%esi), %ebx \n" /* in */
+ "movl 3*T_INT(%esi), %edx \n" /* out */
+ "movl 2*T_INT(%esi), %eax \n" /* value */
+ "movl 4*T_INT(%esi), %ecx \n" /* n */
+
+ "movss (%eax), %xmm0 \n"
+ "shufps $0, %xmm0, %xmm0 \n"
+ "shrl $4, %ecx \n" /* divide by 16 */
+
+ /* loop: *out = max(*in,value) */
+ "smaxopsg_loop: \n"
+
+ "movaps (%ebx), %xmm1 \n"
+ "maxps %xmm0, %xmm1 \n"
+ "movaps %xmm1, (%edx) \n"
+
+ "movaps 4*T_FLOAT(%ebx), %xmm2 \n"
+ "maxps %xmm0, %xmm2 \n"
+ "movaps %xmm2, 4*T_FLOAT(%edx) \n"
+
+ "movaps 8*T_FLOAT(%ebx), %xmm3 \n"
+ "maxps %xmm0, %xmm3 \n"
+ "movaps %xmm3, 8*T_FLOAT(%edx) \n"
+
+ "movaps 12*T_FLOAT(%ebx), %xmm4 \n"
+ "maxps %xmm0, %xmm4 \n"
+ "movaps %xmm4, 12*T_FLOAT(%edx) \n"
+
+ "addl $64, %ebx \n" /* in += 16 */
+ "addl $64, %edx \n" /* out += 16 */
+ "loop smaxopsg_loop \n"
+
+ /* return w+5; */
+ "movl -4(%ebp), %ebx \n"
+ "movl -8(%ebp), %esi \n"
+ "movl 8(%ebp), %eax \n"
+ "addl $20, %eax \n"
+ "movl %ebp, %esp \n"
+ "popl %ebp \n"
+ "ret \n"
+
+ ".align 4 \n" /* alignment */
+ ".type scalarmax_perf_sse_gcc, @function \n"
+
+ );
+
+
#endif
+
Index: m_simd_sse_gcc.h
===================================================================
RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.h,v
retrieving revision 1.1.2.3
retrieving revision 1.1.2.4
diff -C2 -d -r1.1.2.3 -r1.1.2.4
*** m_simd_sse_gcc.h 29 Dec 2003 02:01:57 -0000 1.1.2.3
--- m_simd_sse_gcc.h 10 Jul 2004 19:57:21 -0000 1.1.2.4
***************
*** 33,58 ****
/* functions in d_ugen.c */
! #define zero_perf_simd zero_perf8 /* SIMD not implemented */
/* functions in d_dac.c */
! #define copy_perf_simd copy_perf8 /* SIMD not implemented */
/* functions in d_ctl.c */
! #define sig_tilde_perf_simd sig_tilde_perf8 /* SIMD not implemented */
/* functions in d_arithmetic.c */
! #define plus_perf_simd plus_perf8 /* SIMD not implemented */
! #define scalarplus_perf_simd scalarplus_perf8 /* SIMD not implemented */
! #define minus_perf_simd minus_perf8 /* SIMD not implemented */
! #define scalarminus_perf_simd scalarminus_perf8 /* SIMD not implemented */
! #define times_perf_simd times_perf8 /* SIMD not implemented */
! #define scalartimes_perf_simd scalartimes_perf8 /* SIMD not implemented */
#define sqr_perf_simd sqr_perf8 /* SIMD not implemented */
! #define over_perf_simd over_perf8 /* SIMD not implemented */
! #define scalarover_perf_simd scalarover_perf8 /* SIMD not implemented */
! #define min_perf_simd min_perf8 /* SIMD not implemented */
! #define scalarmin_perf_simd scalarmin_perf8 /* SIMD not implemented */
! #define max_perf_simd max_perf8 /* SIMD not implemented */
! #define scalarmax_perf_simd scalarmax_perf8 /* SIMD not implemented */
/* functions in d_math.c */
--- 33,58 ----
/* functions in d_ugen.c */
! #define zero_perf_simd zero_perf_sse_gcc
/* functions in d_dac.c */
! #define copy_perf_simd copy_perf_sse_gcc
/* functions in d_ctl.c */
! #define sig_tilde_perf_simd sig_tilde_perf_sse_gcc
/* functions in d_arithmetic.c */
! #define plus_perf_simd plus_perf_sse_gcc
! #define scalarplus_perf_simd scalarplus_perf_sse_gcc
! #define minus_perf_simd minus_perf_sse_gcc
! #define scalarminus_perf_simd scalarminus_perf_sse_gcc
! #define times_perf_simd times_perf_sse_gcc
! #define scalartimes_perf_simd scalartimes_perf_sse_gcc
#define sqr_perf_simd sqr_perf8 /* SIMD not implemented */
! #define over_perf_simd over_perf_sse_gcc
! #define scalarover_perf_simd scalarover_perf_sse_gcc
! #define min_perf_simd min_perf_sse_gcc
! #define scalarmin_perf_simd scalarmin_perf_sse_gcc
! #define max_perf_simd max_perf_sse_gcc
! #define scalarmax_perf_simd scalarmax_perf_sse_gcc
/* functions in d_math.c */
More information about the Pd-cvs
mailing list