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;
size_coeffs <<= 1;
coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
}
-
+
coeffs[num_coeffs++] = f;
sum += f;
}
coeffs[pd[x].startcoeff + sx - start] /= sum;
}
}
-
+
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;
for (x = 0; x < nw; ++x) {
+ int result;
float acc;
+ int tmp;
static const float low = 0.0, high = 255.0;
asm (
- "pxor %0, %0 \n"
- "xor %%eax, %%eax \n"
+ "pxor %1, %1 \n"
+ "xor %2, %2 \n"
".lbl2: \n"
- "movups (%2,%%eax),%%xmm1 \n"
- "movups (%1,%%eax),%%xmm2 \n"
+ "movups (%4,%2),%%xmm1 \n"
+ "movups (%3,%2),%%xmm2 \n"
"mulps %%xmm2,%%xmm1 \n"
- "addps %%xmm1,%0 \n"
- "addl $16,%%eax \n"
- "dec %3 \n"
+ "addps %%xmm1,%1 \n"
+ "addl $16,%2 \n"
+ "dec %5 \n"
"jnz .lbl2 \n"
- "haddps %0,%0 \n"
- "haddps %0,%0 \n"
- "maxss %4,%0 \n"
- "minss %5,%0 \n"
- : "=x" (acc)
+ "haddps %1,%1 \n"
+ "haddps %1,%1 \n"
+ "maxss %6,%1 \n"
+ "minss %7,%1 \n"
+ "cvtss2si %1,%0 \n"
+ : "=r" (result),
+ "=&x" (acc),
+ "=&r" (tmp)
: "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"
+ : "xmm1", "xmm2"
);
#if 0
ch = (unsigned char)acc;
*dptr++ = ch;
#endif
- *dptr++ = (unsigned char)acc;
+ *dptr++ = (unsigned char)result;
}
ch = dptr[-1];
for ( ; x < dstride; ++x) {
dinfo.raw_data_out = TRUE;
jpeg_start_decompress(&dinfo);
+ unsigned w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
+ unsigned h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
+
+ unsigned w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
+ unsigned h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
+
+ unsigned w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
+ unsigned h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
+
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);
{
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);
+ 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*)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);
+ 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*)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);
+ 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);