Add an x86 optimized version of extend().

[fjl] / driver.c

diff --git a/driver.c b/driver.c
index aa803d6ed00aa67266495b65df93e6582bc89151..616b613f944ee5ae3ba12a69476eb37d9b20a336 100644 (file)
--- a/driver.c
+++ b/driver.c
@@ -15,10 +15,12 @@ struct jpeg_image {
        unsigned num_components;
        unsigned hsample[256], vsample[256], qtable[256];
        unsigned max_hsample, max_vsample;
+       unsigned stride[256];
        unsigned num_blocks_horizontal, num_blocks_vertical;
        uint32_t qvalues[256][DCTSIZE2];
        void* idct_data[256];
        uint8_t* pixel_data[256];
+       uint8_t* pixel_write_pointer[256];
 };
 
 ssize_t stdio_read(void* userdata, uint8_t* buf, size_t count) 
@@ -103,8 +105,10 @@ void read_sof(struct byte_source* source, struct jpeg_image* image)
 
                unsigned width = image->num_blocks_horizontal * image->hsample[c] * DCTSIZE;
                unsigned height = image->num_blocks_vertical * image->vsample[c] * DCTSIZE;
+               image->stride[c] = width;
                image->pixel_data[c] = (uint8_t*)malloc(width * height);
                assert(image->pixel_data[c] != NULL);
+               image->pixel_write_pointer[c] = image->pixel_data[c];
 
                fprintf(stderr, "Component %u: allocating %d x %d\n", c, width, height);
        }
@@ -151,20 +155,21 @@ void read_scan(struct byte_source* source, struct jpeg_image* image, huffman_tab
                
        unsigned mcu_x = 0, mcu_y = 0;
 
-       for ( ;; ) {
+       while (!bits.source_eof) {
                for (unsigned c = 0; c < num_components; ++c) {
                        unsigned cn = component_num[c];
-                       unsigned stride = image->num_blocks_horizontal * image->hsample[cn] * DCTSIZE;
                        assert(image->idct_data[image->qtable[cn]] != NULL);
-                       
-                       for (unsigned local_yb = 0; local_yb < image->vsample[cn]; ++local_yb) {
-                               for (unsigned local_xb = 0; local_xb < image->hsample[cn]; ++local_xb) {
+
+                       uint8_t* pixel_write_pointer_y = image->pixel_write_pointer[cn];
+                       for (unsigned local_yb = 0; local_yb < image->vsample[cn]; ++local_yb, pixel_write_pointer_y += image->stride[cn] * DCTSIZE) {
+                               uint8_t* pixel_write_pointer = pixel_write_pointer_y;
+                               for (unsigned local_xb = 0; local_xb < image->hsample[cn]; ++local_xb, pixel_write_pointer += DCTSIZE) {
                                        const struct huffman_table* dc_table = &((*tables)[DC_CLASS][dc_huffman_table[c]]);
                                        const struct huffman_table* ac_table = &((*tables)[AC_CLASS][ac_huffman_table[c]]);
 
                                        // decode DC component
                                        unsigned dc_category = read_huffman_symbol(dc_table, &bits);
-                                       possibly_refill(&bits, dc_category);
+                                       possibly_refill(&bits, dc_category + DEHUF_TABLE_BITS);
                                        last_dc[c] += extend(read_bits(&bits, dc_category), dc_category);
                                        
                                        int16_t coeff[DCTSIZE2] = { 0 };
@@ -172,7 +177,7 @@ void read_scan(struct byte_source* source, struct jpeg_image* image, huffman_tab
 
                                        // decode AC components
                                        for (unsigned i = 1; i < DCTSIZE2; ++i) {
-                                               unsigned rs = read_huffman_symbol(ac_table, &bits);
+                                               unsigned rs = read_huffman_symbol_no_refill(ac_table, &bits);
                                                unsigned r = rs >> 4;
                                                unsigned s = rs & 0xf;
 
@@ -182,40 +187,44 @@ void read_scan(struct byte_source* source, struct jpeg_image* image, huffman_tab
                                                }
                                                if (rs == 0xf0) {
                                                        /* 16 zero coefficients */
+                                                       possibly_refill(&bits, DEHUF_TABLE_BITS);
                                                        i += 15;
                                                        continue;
                                                }
                                                i += r;
 
-                                               possibly_refill(&bits, s);
+                                               possibly_refill(&bits, s + DEHUF_TABLE_BITS);
                                                coeff[unzigzag[i]] = extend(read_bits(&bits, s), s);
                                        }
                        
                                        uint8_t pixdata[DCTSIZE2];      
                                        idct_choice(coeff, image->idct_data[image->qtable[cn]], pixdata);
 
-                                       for (unsigned y = 0; y < DCTSIZE; ++y) {
-                                               unsigned real_x = (mcu_x * image->hsample[cn] + local_xb) * DCTSIZE;
-                                               unsigned real_y = (mcu_y * image->vsample[cn] + local_yb) * DCTSIZE + y;
-                                       
-                                               memcpy(image->pixel_data[cn] + real_y * stride + real_x,
-                                                      pixdata + y * DCTSIZE,
-                                                      DCTSIZE);
+                                       uint8_t* dest_pixdata = pixel_write_pointer;
+                                       for (unsigned y = 0; y < DCTSIZE; ++y, dest_pixdata += image->stride[cn]) {
+                                               memcpy(dest_pixdata, pixdata + y * DCTSIZE, DCTSIZE);
                                        }
                                }
                        }
+                       image->pixel_write_pointer[cn] += DCTSIZE * image->hsample[cn];
                }
-               
+       
                if (++mcu_x == image->num_blocks_horizontal) {
                        ++mcu_y;
                        mcu_x = 0;
+               
+                       for (unsigned c = 0; c < num_components; ++c) {
+                               unsigned cn = component_num[c];
+                               image->pixel_write_pointer[cn] += (image->vsample[cn] * DCTSIZE - 1) * image->stride[cn];
+                       }
 
                        // Some debug code.
                        const int c = 1;
                        if (mcu_y == image->num_blocks_vertical) {
                                unsigned stride = image->num_blocks_horizontal * image->hsample[c] * DCTSIZE;
-                               printf("P5\n%u %u\n255\n", stride, image->height);
-                               fwrite(image->pixel_data[c], stride * image->height, 1, stdout);
+                               unsigned height = image->num_blocks_vertical * image->vsample[c] * DCTSIZE;
+                               printf("P5\n%u %u\n255\n", stride, height);
+                               fwrite(image->pixel_data[c], stride * height, 1, stdout);
                        }
                }
        }
@@ -293,6 +302,9 @@ int main(void)
                case 0xd8:
                        /* SOI */
                        break;
+               case 0xd9:
+                       /* EOI */
+                       exit(0);
                case 0xc4:
                        /* DHT (define Huffman tables) */
                        read_huffman_tables(&tables, byte_source_input_func, &source);