-#include "jpeglib.h"
-
-#define CACHE_LINE_FACTOR 16
-
-double sinc(double x)
-{
- 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)
-{
- if (x < -3.0 || x > 3.0)
- return 0.0;
- if (x < 0.0)
- return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
- else
- return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
-}
-
-
-struct pix_desc {
- unsigned start, end;
- unsigned startcoeff;
-};
-
-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;
- double sf = (double)w / (double)nw;
- double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
-
- /* calculate the filter */
- 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) {
- start = 0;
- }
- if (end > w - 1) {
- 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 (sx = start; sx <= end; ++sx) {
- double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
- double f = lanczos_tap(nd);
- if (num_coeffs == size_coeffs) {
- size_coeffs <<= 1;
- coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
- }
-
- coeffs[num_coeffs++] = f;
- sum += f;
- }
-
- for (sx = start; sx <= end; ++sx) {
- coeffs[pd[x].startcoeff + sx - start] /= sum;
- }
- }
-
- for (y = 0; y < h; ++y) {
- float *sptr = pix + y*sstride;
- unsigned char *dptr = npix + y*dstride;
- unsigned char ch;
- for (x = 0; x < nw; ++x) {
- float acc;
- int tmp;
- static const float low = 0.0, high = 255.0;
- asm (
- "pxor %0, %0 \n"
- "xor %1, %1 \n"
- ".lbl2: \n"
- "movups (%3,%1),%%xmm1 \n"
- "movups (%2,%1),%%xmm2 \n"
- "mulps %%xmm2,%%xmm1 \n"
- "addps %%xmm1,%0 \n"
- "addl $16,%1 \n"
- "dec %4 \n"
- "jnz .lbl2 \n"
- "haddps %0,%0 \n"
- "haddps %0,%0 \n"
- "maxss %5,%0 \n"
- "minss %6,%0 \n"
- : "=&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)
- : "xmm1", "xmm2"
- );
-
-#if 0
- if (acc < 0.0)
- ch = 0;
- else if (acc > 255.0)
- ch = 255;
- else
- ch = (unsigned char)acc;
- *dptr++ = ch;
-#endif
- *dptr++ = (unsigned char)acc;
- }
- ch = dptr[-1];
- for ( ; x < dstride; ++x) {
- *dptr++ = ch;
- }
- }
-}
-
-void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
-{
- struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
- int size_coeffs = 8;
- float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
- int num_coeffs = 0;
- int x, y, sy;
- double sf = (double)h / (double)nh;
- double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
-
- /* calculate the filter */
- for (y = 0; y < nh; ++y) {
- int start = ceil(y * sf - support);
- int end = floor(y * sf + support);
- double sum = 0.0;
-
- if (start < 0) {
- start = 0;
- }
- if (end > h - 1) {
- end = h - 1;
- }
-
- pd[y].start = start;
- pd[y].end = end;
- pd[y].startcoeff = num_coeffs;
-
- for (sy = start; sy <= end; ++sy) {
- double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
- double f = lanczos_tap(nd);
- if (num_coeffs == size_coeffs) {
- size_coeffs <<= 1;
- coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
- }
-
- coeffs[num_coeffs++] = f;
- sum += f;
- }
-
- for (sy = start; sy <= end; ++sy) {
- coeffs[pd[y].startcoeff + sy - start] /= sum;
- }
- }
-
-#if CACHE_LINE_FACTOR > 1
- 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) {
- for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
- acc[i] += sptr[sy * w + i] * *cf;
- }
- ++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;
- }
- }
- for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
-#else
- for (x = 0; x < w; ++x) {
-#endif
- unsigned char *sptr = pix + x;
- float *dptr = npix + x;
- for (y = 0; y < nh; ++y) {
- float acc = 0.0;
- float *cf = &coeffs[pd[y].startcoeff];
- unsigned sy;
-
- for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
- acc += sptr[sy * w] * *cf++;
- }
-
- *dptr = acc;
- dptr += dstride;
- }
- }
-}