X-Git-Url: https://git.sesse.net/?p=qscale;a=blobdiff_plain;f=libqscale.c;fp=libqscale.c;h=fa7f2c3f8018c7329632a884eadb9699cc03866c;hp=b744a9b5f3b98e1a1da8182e15f9eaefc991f694;hb=78acccc81264b96a60e355cdab42e15dc84b9b07;hpb=cce321fe623a347fd2bda4e442bed42dfc9299c7 diff --git a/libqscale.c b/libqscale.c index b744a9b..fa7f2c3 100644 --- a/libqscale.c +++ b/libqscale.c @@ -1,9 +1,28 @@ +#include #include #include #include #include "libqscale.h" +/* The number of pixels to process at a time when scaling vertically. */ +#define CACHE_LINE_FACTOR 16 + +/* Whether to use SSE for horizontal scaling or not (requires SSE3). */ +#define USE_HORIZONTAL_SSE 1 + +/* Whether to use SSE for vertical scaling or not (requires only SSE1). */ +#define USE_VERTICAL_SSE 1 + +#if USE_VERTICAL_SSE +#undef CACHE_LINE_FACTOR +#define CACHE_LINE_FACTOR 16 +#endif + +#ifndef M_PI +#define M_PI 3.14159265358979323846264 +#endif + qscale_img *qscale_load_jpeg(const char *filename) { FILE *file = fopen(filename, "rb"); @@ -98,3 +117,403 @@ void qscale_destroy(qscale_img *img) free(img->data_cr); free(img); } + + +static 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; + } +} + +static 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; +}; + +static 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; + } + +#if USE_HORIZONTAL_SSE + /* 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); + } + } +#endif + + 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) { +#if USE_HORIZONTAL_SSE + int result; + float acc; + long tmp; + static const float low = 0.0, high = 255.0; + __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,%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) + : "memory", "xmm1", "xmm2" + ); + + *dptr++ = (unsigned char)result; +#else + 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++; + } + + if (acc < 0.0) + ch = 0; + else if (acc > 255.0) + ch = 255; + else + ch = (unsigned char)acc; + *dptr++ = ch; +#endif + } + ch = dptr[-1]; + for ( ; x < dstride; ++x) { + *dptr++ = ch; + } + } +} + +static 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 USE_VERTICAL_SSE + /* + * 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 */ + "0: \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 0b \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" ((long)w), /* 3: stride */ + "r" (dptr) /* 4: dstptr base */ + : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +#else + 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]; + } +#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; + } + } +} + +qscale_img *qscale_scale(qscale_img *src, unsigned width, unsigned height, unsigned samp_h0, unsigned samp_v0, unsigned samp_h1, unsigned samp_v1, unsigned samp_h2, unsigned samp_v2, enum qscale_scaling_filter scaling_filter) +{ + qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img)); + if (dst == NULL) { + return NULL; + } + + dst->width = width; + dst->height = height; + + 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; + + dst->w0 = width * samp_h0 / max_samp_h; + dst->h0 = height * samp_v0 / max_samp_v; + + dst->w1 = width * samp_h1 / max_samp_h; + dst->h1 = height * samp_v1 / max_samp_v; + + dst->w2 = width * samp_h2 / max_samp_h; + dst->h2 = height * samp_v2 / max_samp_v; + + unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1); + + unsigned sstride0 = (src->w0 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned sstride1 = (src->w1 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned sstride2 = (src->w2 + DCTSIZE-1) & ~(DCTSIZE-1); + + /* FIXME: handle out-of-memory gracefully */ + { + float *npix = (float*)memalign(16, sstride0 * dst->h0 * sizeof(float)); + vscale(src->data_y, npix, sstride0, src->h0, dst->h0, sstride0); + dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0); + hscale(npix, dst->data_y, src->w0, dst->h0, dst->w0, sstride0, dstride0); + free(npix); + } + { + float *npix = (float*)memalign(16, sstride1 * dst->h1 * sizeof(float)); + vscale(src->data_cr, npix, sstride1, src->h1, dst->h1, sstride1); + dst->data_cr = (unsigned char *)malloc(dst->h1 * dstride1); + hscale(npix, dst->data_cr, src->w1, dst->h1, dst->w1, sstride1, dstride1); + free(npix); + } + { + float *npix = (float*)memalign(16, sstride2 * dst->h2 * sizeof(float)); + vscale(src->data_cb, npix, sstride2, src->h2, dst->h2, sstride2); + dst->data_cb = (unsigned char *)malloc(dst->h2 * dstride2); + hscale(npix, dst->data_cb, src->w2, dst->h2, dst->w2, sstride2, dstride2); + free(npix); + } + + return dst; +}