X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=qscale.c;h=992bf7d24e3aae137f887bd3215876db66921ed5;hb=139d3c0c4a2e228d7f8dbca7e6ac0bef6222d6e5;hp=042d3411a110f5464e218a3ef8c4b55f0af66ca8;hpb=9c2b4088f2f72806dfd05136845e8e11a2dfb7f8;p=qscale diff --git a/qscale.c b/qscale.c index 042d341..992bf7d 100644 --- a/qscale.c +++ b/qscale.c @@ -1,39 +1,25 @@ #include +#include #include #include #include +#include "jpeglib.h" -#define CACHE_LINE_FACTOR 1 +#define CACHE_LINE_FACTOR 16 -void myreadline(char *buf) +double sinc(double x) { - char *ptr; - - if (fgets(buf, 256, stdin) == NULL) { - fprintf(stderr, "Premuture EOF\n"); - exit(1); - } + static const double cutoff = 1.220703668e-4; // sqrt(sqrt(eps)) - if (strlen(buf) == 0) { - return; - } + if (abs(x) < cutoff) { + // For small |x|, use Taylor series instead + const double x2 = x * x; + const double x4 = x2 * x2; - ptr = &buf[strlen(buf) - 1]; - if (*ptr == '\n') { - *ptr-- = 0; - if (*ptr == '\r') { - *ptr = 0; - } - } -} - -double sinc(double x) -{ - // This is bad for very small x, should use power series instead. - if (x == 0.0) - return 1.0; - else + return 1.0 - x2 / 6.0 + x4 / 120.0; + } else { return sin(x) / x; + } } double lanczos_tap(double x) @@ -52,13 +38,13 @@ struct pix_desc { unsigned startcoeff; }; -void hscale(float *pix, unsigned char *npix, 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, y, sx; + int x, y; double sf = (double)w / (double)nw; double support = (w > nw) ? (3.0 * sf) : (3.0 / sf); @@ -66,6 +52,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw 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) { @@ -75,6 +62,19 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw 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; @@ -86,7 +86,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw size_coeffs <<= 1; coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float)); } - + coeffs[num_coeffs++] = f; sum += f; } @@ -95,20 +95,44 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw coeffs[pd[x].startcoeff + sx - start] /= sum; } } - + for (y = 0; y < h; ++y) { - float *sptr = pix + y*w; - unsigned char *dptr = npix + y*nw; + float *sptr = pix + y*sstride; + unsigned char *dptr = npix + y*dstride; + unsigned char ch; for (x = 0; x < nw; ++x) { - float acc = 0.0; - float *cf = &coeffs[pd[x].startcoeff]; - unsigned sx; - unsigned char ch; - - for (sx = pd[x].start; sx <= pd[x].end; ++sx) { - acc += sptr[sx] * *cf++; - } - + 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) @@ -116,11 +140,17 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw else ch = (unsigned char)acc; *dptr++ = ch; +#endif + *dptr++ = (unsigned char)result; + } + ch = dptr[-1]; + for ( ; x < dstride; ++x) { + *dptr++ = ch; } } } -void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh) +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; @@ -165,17 +195,18 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh } #if CACHE_LINE_FACTOR > 1 - for (x = 0; x < w; x += CACHE_LINE_FACTOR) { + 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; @@ -186,7 +217,96 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh for (i = 0; i < CACHE_LINE_FACTOR; ++i) { dptr[i] = acc[i]; } - dptr += w; +#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; } } for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) { @@ -205,74 +325,160 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh } *dptr = acc; - dptr += w; + dptr += dstride; } } } -int main(void) +int main(int argc, char **argv) { - unsigned w, h; - unsigned nw = 600, nh = 400; - //unsigned nw = 3504, nh = 2336; - char buf[256]; - unsigned char *pix; - float *npix; - unsigned char *npix2; - - myreadline(buf); - if (strcmp(buf, "P5") != 0) { - fprintf(stderr, "Error: Not a PGM file\n"); - exit(1); - } - myreadline(buf); /* comment */ - myreadline(buf); + 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 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; + unsigned nw2 = nominal_w * samp_h2 / max_samp_h, nh2 = nominal_h * samp_v2 / max_samp_v; + + unsigned stride0 = (nw0 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned stride1 = (nw1 + DCTSIZE-1) & ~(DCTSIZE-1); + unsigned stride2 = (nw2 + DCTSIZE-1) & ~(DCTSIZE-1); + + struct jpeg_decompress_struct dinfo; + struct jpeg_error_mgr jerr; + dinfo.err = jpeg_std_error(&jerr); + jpeg_create_decompress(&dinfo); + jpeg_stdio_src(&dinfo, stdin); + jpeg_read_header(&dinfo, TRUE); + 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); + + 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; + + JSAMPROW y_row_ptrs[max_lines]; + JSAMPROW cb_row_ptrs[max_lines]; + JSAMPROW cr_row_ptrs[max_lines]; + JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs }; + int i; + + for (i = 0; i < max_lines; ++i) { + 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; + } + + total_lines += max_lines; + ++blocks; - if (sscanf(buf, "%u %u", &w, &h) != 2) { - fprintf(stderr, "Error: Expected width/height\n"); - exit(1); + if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0) + break; } - myreadline(buf); - if (strcmp(buf, "255") != 0) { - fprintf(stderr, "Error: Expected 8-bit\n"); - exit(1); + { + 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); } - - pix = (unsigned char *)malloc(w * h); - if (pix == NULL) { - perror("malloc"); - exit(1); + { + 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); } - - if (fread(pix, w*h, 1, stdin) != 1) { - fprintf(stderr, "Error: Short read\n"); - exit(1); + { + 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); - npix = (float *)malloc(sizeof(float) * nw * h); - if (npix == NULL) { - perror("malloc"); - exit(1); - } - npix2 = (unsigned char *)malloc(nw * nh); - if (npix2 == NULL) { - perror("malloc"); - exit(1); - } - - vscale(pix, npix, w, h, nh); - hscale(npix, npix2, w, nh, nw); - - printf("P5\n# COMMENT: Made by qscale\n%u %u\n255\n", nw, nh); - { - int y, x; - for (y = 0; y < nh; ++y) { - for (x = 0; x < nw; ++x) { - putchar(npix2[y * nw + x]); - } + struct jpeg_compress_struct cinfo; + cinfo.err = jpeg_std_error(&jerr); + jpeg_create_compress(&cinfo); + jpeg_stdio_dest(&cinfo, stdout); + cinfo.input_components = 3; + jpeg_set_defaults(&cinfo); + jpeg_set_quality(&cinfo, 85, FALSE); + cinfo.image_width = nominal_w; + cinfo.image_height = nominal_h; + cinfo.raw_data_in = TRUE; + jpeg_set_colorspace(&cinfo, JCS_YCbCr); + cinfo.comp_info[0].h_samp_factor = samp_h0; + cinfo.comp_info[0].v_samp_factor = samp_v0; + cinfo.comp_info[1].h_samp_factor = samp_h1; + cinfo.comp_info[1].v_samp_factor = samp_v1; + cinfo.comp_info[2].h_samp_factor = samp_h2; + cinfo.comp_info[2].v_samp_factor = samp_v2; + jpeg_start_compress(&cinfo, TRUE); + + total_lines = 0; + blocks = 0; + while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) { + unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE; + + JSAMPROW y_row_ptrs[max_lines]; + JSAMPROW cb_row_ptrs[max_lines]; + JSAMPROW cr_row_ptrs[max_lines]; + JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs }; + int i; + + for (i = 0; i < max_lines; ++i) { + // simple edge extension + int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor; + if (yline > nh0 - 1) + yline = nh0 - 1; + + int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor; + if (cbline > nh1 - 1) + cbline = nh1 - 1; + + int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor; + if (crline > nh2 - 1) + crline = nh2 - 1; + + 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; + ++blocks; + + jpeg_write_raw_data(&cinfo, ptrs, max_lines); } + jpeg_finish_compress(&cinfo); + jpeg_destroy_compress(&cinfo); return 0; } +