#define CACHE_LINE_FACTOR 16
+#ifndef M_PI
+#define M_PI 3.14159265358979323846264
+#endif
+
double sinc(double x)
{
- static const double cutoff = 1.220703668e-4; // sqrt(sqrt(eps))
+ static const double cutoff = 1.220703668e-4; /* sqrt(sqrt(eps)) */
if (abs(x) < cutoff) {
- // For small |x|, use Taylor series instead
+ /* For small |x|, use Taylor series instead */
const double x2 = x * x;
const double x4 = x2 * x2;
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;
end = w - 1;
}
- // round up so we get a multiple of four for the SSE code
+ /* 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
+ /* prefer aligning it if possible */
if (start % 4 != 0 && start % 4 <= num % 4) {
num += start % 4;
start -= start % 4;
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;
+ long tmp;
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"
+ __asm__ (
+ "pxor %1, %1 \n"
+ "xor %2, %2 \n"
+ "0: \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"
- "jnz .lbl2 \n"
- "haddps %0,%0 \n"
- "haddps %0,%0 \n"
- "maxss %4,%0 \n"
- "minss %5,%0 \n"
- : "=x" (acc)
+ "addps %%xmm1,%1 \n"
+ "add $16,%2 \n"
+ "dec %5 \n"
+ "jnz 0b \n"
+ "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"
+ : "memory", "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) {
* xmm4: current filter coefficient
* xmm5, xmm6, xmm7: scratchpad
*/
- asm (
+ __asm__ (
/* clear */
- "pxor %%xmm0, %%xmm0 \n"
- "pxor %%xmm1, %%xmm1 \n"
- "pxor %%xmm2, %%xmm2 \n"
- "pxor %%xmm3, %%xmm3 \n"
+ "pxor %%xmm0, %%xmm0 \n"
+ "pxor %%xmm1, %%xmm1 \n"
+ "pxor %%xmm2, %%xmm2 \n"
+ "pxor %%xmm3, %%xmm3 \n"
/* main loop */
- ".lbl: \n"
+ "0: \n"
/* a zero is useful during unpacking */
- "pxor %%xmm4, %%xmm4 \n"
+ "pxor %%xmm4, %%xmm4 \n"
/* fetch all 16 source bytes */
- "movups (%0), %%xmm5 \n"
- "prefetcht0 (%0,%3,4) \n"
+ "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"
+ "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"
+ "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"
+ "cvtdq2ps %%xmm5, %%xmm5 \n"
+ "cvtdq2ps %%xmm6, %%xmm6 \n"
/* fetch the coefficient */
- "movss (%2), %%xmm4 \n"
- "shufps $0x0, %%xmm4, %%xmm4 \n"
+ "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"
+ "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"
+ "pxor %%xmm4, %%xmm4 \n"
/* unpack xmm7 into dwords (xmm7, xmm6) */
- "movaps %%xmm7, %%xmm6 \n"
- "punpcklwd %%xmm4, %%xmm7 \n"
- "punpckhwd %%xmm4, %%xmm6 \n"
+ "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"
+ "cvtdq2ps %%xmm7, %%xmm7 \n"
+ "cvtdq2ps %%xmm6, %%xmm6 \n"
/* fetch the coefficient */
- "movss (%2), %%xmm4 \n"
- "shufps $0x0, %%xmm4, %%xmm4 \n"
+ "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"
+ "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"
+ "add $4, %2 \n"
+ "add %3, %0 \n"
+ "dec %1 \n"
+ "jnz 0b \n"
/* store the values */
- "movaps %%xmm0, (%4) \n"
- "movaps %%xmm1, 16(%4) \n"
- "movaps %%xmm2, 32(%4) \n"
- "movaps %%xmm3, 48(%4) \n"
+ "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" ((long)w), /* 3: stride */
"r" (dptr) /* 4: dstptr base */
: "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
);
int main(int argc, char **argv)
{
+ /* user-settable parameters */
unsigned nominal_w = atoi(argv[1]);
unsigned nominal_h = atoi(argv[2]);
-
unsigned samp_h0 = 2, samp_v0 = 2;
unsigned samp_h1 = 1, samp_v1 = 1;
unsigned samp_h2 = 1, samp_v2 = 1;
- unsigned max_samp_h = 2, max_samp_v = 2;
+ unsigned jpeg_quality = 85;
+ /* end */
+
+ unsigned max_samp_h, max_samp_v;
+ max_samp_h = samp_h0;
+ if (samp_h1 > max_samp_h)
+ max_samp_h = samp_h1;
+ if (samp_h2 > max_samp_h)
+ max_samp_h = samp_h2;
+
+ max_samp_v = samp_v0;
+ if (samp_v1 > max_samp_v)
+ max_samp_v = samp_v1;
+ if (samp_v2 > max_samp_v)
+ max_samp_v = samp_v2;
unsigned nw0 = nominal_w * samp_h0 / max_samp_h, nh0 = nominal_h * samp_v0 / max_samp_v;
unsigned nw1 = nominal_w * samp_h1 / max_samp_h, nh1 = nominal_h * samp_v1 / max_samp_v;
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);
- data_ny = (unsigned char *)malloc(nw0 * stride0);
- hscale(npix, data_ny, dinfo.comp_info[0].width_in_blocks * DCTSIZE, nh0, nw0, stride0);
+ 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(nh0 * 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);
- data_ncr = (unsigned char *)malloc(nw1 * stride1);
- hscale(npix, data_ncr, dinfo.comp_info[1].width_in_blocks * DCTSIZE, nh1, nw1, stride1);
+ 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(nh1 * 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);
- data_ncb = (unsigned char *)malloc(nw2 * stride2);
- hscale(npix, data_ncb, dinfo.comp_info[2].width_in_blocks * DCTSIZE, nh2, nw2, stride2);
+ 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(nh2 * stride2);
+ hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2);
free(npix);
}
jpeg_destroy_decompress(&dinfo);
jpeg_stdio_dest(&cinfo, stdout);
cinfo.input_components = 3;
jpeg_set_defaults(&cinfo);
- jpeg_set_quality(&cinfo, 85, FALSE);
+ jpeg_set_quality(&cinfo, jpeg_quality, FALSE);
cinfo.image_width = nominal_w;
cinfo.image_height = nominal_h;
cinfo.raw_data_in = TRUE;
int i;
for (i = 0; i < max_lines; ++i) {
- // simple edge extension
+ /* simple edge extension */
int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
if (yline > nh0 - 1)
yline = nh0 - 1;