X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=qscale.c;h=992bf7d24e3aae137f887bd3215876db66921ed5;hb=139d3c0c4a2e228d7f8dbca7e6ac0bef6222d6e5;hp=c847a12122cc38e1deb80fe3e4c75a60c6d27428;hpb=82b2da83a6924b542e0cc399437ad848315ed8c5;p=qscale diff --git a/qscale.c b/qscale.c index c847a12..992bf7d 100644 --- a/qscale.c +++ b/qscale.c @@ -1,18 +1,25 @@ #include +#include #include #include #include #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) @@ -25,29 +32,27 @@ double lanczos_tap(double x) return sinc(x*M_PI) * sinc(x*M_PI / 3.0); } -struct filter { - struct pix_desc *pd; - float *coeffs; -}; + struct pix_desc { unsigned start, end; unsigned startcoeff; }; -void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned nw) +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, sx; + 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) { @@ -57,6 +62,19 @@ void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned 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; @@ -68,7 +86,7 @@ void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned size_coeffs <<= 1; coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float)); } - + coeffs[num_coeffs++] = f; sum += f; } @@ -78,38 +96,61 @@ void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned } } - filter->pd = pd; - filter->coeffs = coeffs; -} - -void hscale(struct filter *filter, unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride) -{ - struct pix_desc *pd = filter->pd; - float *coeffs = filter->coeffs; - int y; - for (y = 0; y < h; ++y) { - unsigned char *sptr = pix + y*sstride; - float *dptr = npix + y*dstride; - float acc = 0.0; - int x; + float *sptr = pix + y*sstride; + unsigned char *dptr = npix + y*dstride; + unsigned char ch; for (x = 0; x < nw; ++x) { - float *cf = &coeffs[pd[x].startcoeff]; - unsigned sx; - acc = 0.0; - - for (sx = pd[x].start; sx <= pd[x].end; ++sx) { - acc += sptr[sx] * *cf++; - } - *dptr++ = acc; + int result; + float acc; + long tmp; + static const float low = 0.0, high = 255.0; + asm ( + "pxor %1, %1 \n" + "xor %2, %2 \n" + ".lbl2: \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 .lbl2 \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) + : "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)result; } + ch = dptr[-1]; for ( ; x < dstride; ++x) { - *dptr++ = acc; + *dptr++ = ch; } } } -void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride) +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; @@ -154,17 +195,18 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh } #if CACHE_LINE_FACTOR > 1 - for (x = 0; x < w; x += CACHE_LINE_FACTOR) { - float *sptr = pix + x; - unsigned char *dptr = npix + x; + 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; @@ -173,15 +215,97 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh } for (i = 0; i < CACHE_LINE_FACTOR; ++i) { - unsigned char ch; - if (acc[i] < 0.0) - ch = 0; - else if (acc[i] > 255.0) - ch = 255; - else - ch = (unsigned char)acc[i]; - dptr[i] = ch; + 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" ((long)w), /* 3: stride */ + "r" (dptr) /* 4: dstptr base */ + : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +#endif dptr += dstride; } } @@ -189,8 +313,8 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh #else for (x = 0; x < w; ++x) { #endif - float *sptr = pix + x; - unsigned char *dptr = npix + x; + 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]; @@ -200,14 +324,7 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh acc += sptr[sy * w] * *cf++; } - unsigned char ch; - if (acc < 0.0) - ch = 0; - else if (acc > 255.0) - ch = 255; - else - ch = (unsigned char)acc; - *dptr = ch; + *dptr = acc; dptr += dstride; } } @@ -240,42 +357,28 @@ int main(int argc, char **argv) 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); 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); - - float *npix_y = (float *)malloc((dinfo.comp_info[0].height_in_blocks + dinfo.comp_info[0].v_samp_factor - 1) * DCTSIZE * stride0 * sizeof(float)); - float *npix_cb = (float *)malloc((dinfo.comp_info[1].height_in_blocks + dinfo.comp_info[1].v_samp_factor - 1) * DCTSIZE * stride1 * sizeof(float)); - float *npix_cr = (float *)malloc((dinfo.comp_info[2].height_in_blocks + dinfo.comp_info[2].v_samp_factor - 1) * DCTSIZE * stride2 * sizeof(float)); - JSAMPLE *data_y = (unsigned char *)malloc(nh0 * stride0); - JSAMPLE *data_cb = (unsigned char *)malloc(nh1 * stride1); - JSAMPLE *data_cr = (unsigned char *)malloc(nh2 * stride2); - - struct filter filt0, filt1, filt2; - - hscale_calc_filter(&filt0, - /* w= */ dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[0].v_samp_factor * DCTSIZE, - /* nw= */ nw0 - ); - hscale_calc_filter(&filt1, - /* w= */ dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[1].v_samp_factor * DCTSIZE, - /* nw= */ nw1 - ); - hscale_calc_filter(&filt2, - /* w= */ dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[2].v_samp_factor * DCTSIZE, - /* nw= */ nw2 - ); + + 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; while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) { unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE; - JSAMPLE tmp_y [(dinfo.comp_info[0].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[0].v_samp_factor)]; - JSAMPLE tmp_cb[(dinfo.comp_info[1].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[1].v_samp_factor)]; - JSAMPLE tmp_cr[(dinfo.comp_info[2].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[2].v_samp_factor)]; JSAMPROW y_row_ptrs[max_lines]; JSAMPROW cb_row_ptrs[max_lines]; @@ -284,49 +387,39 @@ int main(int argc, char **argv) int i; for (i = 0; i < max_lines; ++i) { - y_row_ptrs[i] = tmp_y + i * dinfo.comp_info[0].width_in_blocks * DCTSIZE; - cb_row_ptrs[i] = tmp_cb + i * dinfo.comp_info[1].width_in_blocks * DCTSIZE; - cr_row_ptrs[i] = tmp_cr + i * dinfo.comp_info[2].width_in_blocks * DCTSIZE; + y_row_ptrs[i] = data_y + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE; + cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE; + cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE; } - int lines = jpeg_read_raw_data(&dinfo, ptrs, max_lines); - if (lines == 0) - break; - - hscale(/* filter= */ &filt0, - /* from= */ tmp_y, - /* to= */ npix_y + blocks * DCTSIZE * dinfo.comp_info[0].v_samp_factor * stride0, - /* w= */ dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[0].v_samp_factor * DCTSIZE, - /* nw= */ nw0, - /* sstride= */ dinfo.comp_info[0].width_in_blocks * DCTSIZE, - /* dstride= */ stride0); - - hscale(/* filter= */ &filt1, - /* from= */ tmp_cb, - /* to= */ npix_cb + blocks * DCTSIZE * dinfo.comp_info[1].v_samp_factor * stride1, - /* w= */ dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[1].v_samp_factor * DCTSIZE, - /* nw= */ nw1, - /* sstride= */ dinfo.comp_info[1].width_in_blocks * DCTSIZE, - /* dstride= */ stride1); - - hscale(/* filter= */ &filt2, - /* from= */ tmp_cr, - /* to= */ npix_cr + blocks * DCTSIZE * dinfo.comp_info[2].v_samp_factor * stride2, - /* w= */ dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor, - /* h= */ dinfo.comp_info[2].v_samp_factor * DCTSIZE, - /* nw= */ nw2, - /* sstride= */ dinfo.comp_info[2].width_in_blocks * DCTSIZE, - /* dstride= */ stride2); - total_lines += max_lines; ++blocks; + + if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0) + break; } - vscale(npix_y, data_y, stride0, dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor, nh0, stride0); - vscale(npix_cb, data_cb, stride1, dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor, nh1, stride1); - vscale(npix_cr, data_cr, stride2, dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor, nh2, stride2); + { + 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, 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, h1, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE); + data_ncr = (unsigned char *)malloc(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, h2, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE); + data_ncb = (unsigned char *)malloc(nw2 * stride2); + hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2); + free(npix); + } jpeg_destroy_decompress(&dinfo); struct jpeg_compress_struct cinfo; @@ -373,9 +466,9 @@ int main(int argc, char **argv) if (crline > nh2 - 1) crline = nh2 - 1; - y_row_ptrs[i] = data_y + yline * stride0; - cb_row_ptrs[i] = data_cb + cbline * stride1; - cr_row_ptrs[i] = data_cr + crline * stride2; + y_row_ptrs[i] = data_ny + yline * stride0; + cb_row_ptrs[i] = data_ncb + cbline * stride1; + cr_row_ptrs[i] = data_ncr + crline * stride2; } total_lines += max_lines;