#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)
{
- // This is bad for very small x, should use power series instead.
- if (x == 0.0)
- return 1.0;
- else
+ static const double cutoff = 1.220703668e-4; // sqrt(sqrt(eps))
+
+ if (abs(x) < cutoff) {
+ // For small |x|, use Taylor series instead
+ const double x2 = x * x;
+ const double x4 = x2 * x2;
+
+ return 1.0 - x2 / 6.0 + x4 / 120.0;
+ } else {
return sin(x) / x;
+ }
}
double lanczos_tap(double x)
unsigned startcoeff;
};
-void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned dstride)
+void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
{
struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
int size_coeffs = 8;
float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
int num_coeffs = 0;
- int x, y, sx;
+ int x, y;
double sf = (double)w / (double)nw;
double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
for (x = 0; x < nw; ++x) {
int start = ceil(x * sf - support);
int end = floor(x * sf + support);
+ int sx;
double sum = 0.0;
if (start < 0) {
end = w - 1;
}
+ // round up so we get a multiple of four for the SSE code
+ int num = (end - start + 1);
+ if (num % 4 != 0) {
+ // prefer aligning it if possible
+ if (start % 4 != 0 && start % 4 <= num % 4) {
+ num += start % 4;
+ start -= start % 4;
+ }
+ if (num % 4 != 0) {
+ end += 4 - (num % 4);
+ }
+ }
+
pd[x].start = start;
pd[x].end = end;
pd[x].startcoeff = num_coeffs;
}
for (y = 0; y < h; ++y) {
- float *sptr = pix + y*w;
+ float *sptr = pix + y*sstride;
unsigned char *dptr = npix + y*dstride;
- unsigned char ch = 0;
+ unsigned char ch;
for (x = 0; x < nw; ++x) {
- float acc = 0.0;
- float *cf = &coeffs[pd[x].startcoeff];
- unsigned sx;
-
- for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
- acc += sptr[sx] * *cf++;
- }
-
+ float acc;
+ static const float low = 0.0, high = 255.0;
+ asm (
+ "pxor %0, %0 \n"
+ "xor %%eax, %%eax \n"
+ ".lbl2: \n"
+ "movups (%2,%%eax),%%xmm1 \n"
+ "movups (%1,%%eax),%%xmm2 \n"
+ "mulps %%xmm2,%%xmm1 \n"
+ "addps %%xmm1,%0 \n"
+ "addl $16,%%eax \n"
+ "dec %3 \n"
+ "jnz .lbl2 \n"
+ "haddps %0,%0 \n"
+ "haddps %0,%0 \n"
+ "maxss %4,%0 \n"
+ "minss %5,%0 \n"
+ : "=x" (acc)
+ : "r" (&coeffs[pd[x].startcoeff]),
+ "r" (&sptr[pd[x].start]),
+ "r" ((pd[x].end - pd[x].start + 1)/4),
+ "m" (low),
+ "m" (high)
+ : "eax", "xmm1", "xmm2"
+ );
+
+#if 0
if (acc < 0.0)
ch = 0;
else if (acc > 255.0)
else
ch = (unsigned char)acc;
*dptr++ = ch;
+#endif
+ *dptr++ = (unsigned char)acc;
}
+ ch = dptr[-1];
for ( ; x < dstride; ++x) {
*dptr++ = ch;
}
}
#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;
}
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 */
+ "add $4, %2 \n"
+ "add %3, %0 \n"
+ "dec %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;
}
}
dinfo.raw_data_out = TRUE;
jpeg_start_decompress(&dinfo);
+ unsigned w0 = dinfo.image_width * samp_h0 / max_samp_h, h0 = dinfo.image_height * samp_v0 / max_samp_v;
+ unsigned w1 = dinfo.image_width * samp_h1 / max_samp_h, h1 = dinfo.image_height * samp_v1 / max_samp_v;
+ unsigned w2 = dinfo.image_width * samp_h2 / max_samp_h, h2 = dinfo.image_height * samp_v2 / max_samp_v;
+
fprintf(stderr, "Scaling using Lanczos filter:\n");
fprintf(stderr, " Y component: %ux%u -> %ux%u\n", dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nw0, nh0);
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;
}
{
- float *npix = (float*)malloc(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);
+ 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, h0, 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);
+ hscale(npix, data_ny, w0, nh0, nw0, dinfo.comp_info[0].width_in_blocks * DCTSIZE, stride0);
free(npix);
}
{
- float *npix = (float*)malloc(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);
+ 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, h1, 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);
+ hscale(npix, data_ncr, w1, nh1, nw1, dinfo.comp_info[1].width_in_blocks * DCTSIZE, stride1);
free(npix);
}
{
- float *npix = (float*)malloc(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);
+ 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, h2, 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);
+ hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2);
free(npix);
}
jpeg_destroy_decompress(&dinfo);
projects / qscale / blobdiff