Implement first round of SSE.

diff --git a/qscale.c b/qscale.c
index 8fecd666a972b1fad7a35b302dc5f8445677aca0..d5536bc36a359aff34690a0e67485c15561af83c 100644 (file)
--- a/qscale.c
+++ b/qscale.c
@@ -1,10 +1,11 @@
 #include <stdio.h>
+#include <malloc.h>
 #include <math.h>
 #include <string.h>
 #include <stdlib.h>
 #include "jpeglib.h"
 
-#define CACHE_LINE_FACTOR 8
+#define CACHE_LINE_FACTOR 16
 
 double sinc(double x)
 {
@@ -147,19 +148,19 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh
        }
 
 #if CACHE_LINE_FACTOR > 1
-       for (x = 0; x < w; x += CACHE_LINE_FACTOR) {
+       for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
                unsigned char *sptr = pix + x;
                float *dptr = npix + x;
                for (y = 0; y < nh; ++y) {
+#if 0
                        int i;
                        float acc[CACHE_LINE_FACTOR];
                        for (i = 0; i < CACHE_LINE_FACTOR; ++i)
                                acc[i] = 0.0;
                        float *cf = &coeffs[pd[y].startcoeff];
                        unsigned sy;
-                       
+               
                        for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
-                               //asm volatile ("prefetcht0 %0" :: "m" (sptr[(sy+1) * w]));
                                for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
                                        acc[i] += sptr[sy * w + i] * *cf;
                                }
@@ -169,6 +170,95 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh
                        for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
                                dptr[i] = acc[i];
                        }
+#else
+                       /*
+                        * xmm0 - xmm3: acc[0..15]
+                        * xmm4: current filter coefficient
+                        * xmm5, xmm6, xmm7: scratchpad
+                        */
+                       asm (
+                               /* clear */
+                               "pxor %%xmm0, %%xmm0    \n"
+                               "pxor %%xmm1, %%xmm1    \n"
+                               "pxor %%xmm2, %%xmm2    \n"
+                               "pxor %%xmm3, %%xmm3    \n"
+
+                               /* main loop */
+                               ".lbl:                   \n"
+                               
+                               /* a zero is useful during unpacking */
+                               "pxor %%xmm4, %%xmm4     \n"
+                               
+                               /* fetch all 16 source bytes */
+                               "movups (%0), %%xmm5     \n"
+                               "prefetcht0 (%0,%3,4)    \n"
+
+                               /* unpack into words (xmm5, xmm7) */
+                               "movaps %%xmm5, %%xmm7    \n"
+                               "punpcklbw %%xmm4, %%xmm5 \n"
+                               "punpckhbw %%xmm4, %%xmm7 \n"
+
+                               /* unpack xmm5 into dwords (xmm5, xmm6) */
+                               "movaps %%xmm5, %%xmm6    \n"
+                               "punpcklwd %%xmm4, %%xmm5 \n"
+                               "punpckhwd %%xmm4, %%xmm6 \n"
+
+                               /* convert xmm5, xmm6 to floats */
+                               "cvtdq2ps %%xmm5, %%xmm5 \n"
+                               "cvtdq2ps %%xmm6, %%xmm6 \n"
+
+                               /* fetch the coefficient */
+                               "movss (%2), %%xmm4      \n"
+                               "shufps $0x0, %%xmm4, %%xmm4 \n"
+
+                               /* do the muls for xmm5 and xmm6 */
+                               "mulps %%xmm4, %%xmm5    \n"
+                               "mulps %%xmm4, %%xmm6    \n"
+                               "addps %%xmm5, %%xmm0    \n"
+                               "addps %%xmm6, %%xmm1    \n"
+
+                               /* get the zero back again */
+                               "pxor %%xmm4, %%xmm4     \n"
+
+                               /* unpack xmm7 into dwords (xmm7, xmm6) */
+                               "movaps %%xmm7, %%xmm6    \n"
+                               "punpcklwd %%xmm4, %%xmm7 \n"
+                               "punpckhwd %%xmm4, %%xmm6 \n"
+
+                               /* convert xmm7, xmm6 to floats */
+                               "cvtdq2ps %%xmm7, %%xmm7 \n"
+                               "cvtdq2ps %%xmm6, %%xmm6 \n"
+
+                               /* fetch the coefficient */
+                               "movss (%2), %%xmm4      \n"
+                               "shufps $0x0, %%xmm4, %%xmm4 \n"
+
+                               /* do the second set of muls */
+                               "mulps %%xmm4, %%xmm7    \n"
+                               "mulps %%xmm4, %%xmm6    \n"
+                               "addps %%xmm7, %%xmm2    \n"
+                               "addps %%xmm6, %%xmm3    \n"
+
+                               /* move along, and loop */
+                               "addl $4, %2             \n"
+                               "addl %3, %0             \n"
+                               "decl %1                 \n"
+                               "jnz .lbl                \n"
+
+                               /* store the values */
+                               "movaps %%xmm0, (%4)     \n"
+                               "movaps %%xmm1, 16(%4)   \n"
+                               "movaps %%xmm2, 32(%4)   \n"
+                               "movaps %%xmm3, 48(%4)   \n"
+                               : :
+                               "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
+                               "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
+                               "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
+                               "r" (w),                             /* 3: stride */
+                               "r" (dptr)                           /* 4: dstptr base */
+                               : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+                       );
+#endif
                        dptr += dstride;
                }
        }
@@ -225,9 +315,9 @@ int main(int argc, char **argv)
        fprintf(stderr, "  Cb component: %ux%u -> %ux%u\n", dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nw1, nh1);
        fprintf(stderr, "  Cr component: %ux%u -> %ux%u\n", dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nw2, nh2);
 
-       JSAMPLE *data_y  = (JSAMPLE*)malloc(dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
-       JSAMPLE *data_cb = (JSAMPLE*)malloc(dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
-       JSAMPLE *data_cr = (JSAMPLE*)malloc(dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
        JSAMPLE *data_ny, *data_ncb, *data_ncr;
 
        int total_lines = 0, blocks = 0;
@@ -254,21 +344,21 @@ int main(int argc, char **argv)
        }
 
        {
-               float *npix = (float*)malloc(dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));       
+               float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float)); 
                vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
                data_ny = (unsigned char *)malloc(nw0 * stride0);
                hscale(npix, data_ny, dinfo.comp_info[0].width_in_blocks * DCTSIZE, nh0, nw0, stride0);
                free(npix);
        }
        {
-               float *npix = (float*)malloc(dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));       
+               float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float)); 
                vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
                data_ncr = (unsigned char *)malloc(nw1 * stride1);
                hscale(npix, data_ncr, dinfo.comp_info[1].width_in_blocks * DCTSIZE, nh1, nw1, stride1);
                free(npix);
        }
        {
-               float *npix = (float*)malloc(dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));       
+               float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float)); 
                vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
                data_ncb = (unsigned char *)malloc(nw2 * stride2);
                hscale(npix, data_ncb, dinfo.comp_info[2].width_in_blocks * DCTSIZE, nh2, nw2, stride2);