Stop hardcoding blocks per row in the shader.

[narabu] / qdc.cpp

diff --git a/qdc.cpp b/qdc.cpp
index db3fe275a97c18e9a386aca0332cabcd90ef0038..5a4a9d506195586ec9c56d2893621c0f69379233 100644 (file)
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -18,6 +18,11 @@
 
 #define WIDTH 1280
 #define HEIGHT 720
+#define WIDTH_BLOCKS (WIDTH/8)
+#define WIDTH_BLOCKS_CHROMA (WIDTH/16)
+#define HEIGHT_BLOCKS (HEIGHT/8)
+#define NUM_BLOCKS (WIDTH_BLOCKS * HEIGHT_BLOCKS)
+#define NUM_BLOCKS_CHROMA (WIDTH_BLOCKS_CHROMA * HEIGHT_BLOCKS)
 #define NUM_SYMS 256
 #define ESCAPE_LIMIT (NUM_SYMS - 1)
 
@@ -714,16 +719,35 @@ int main(int argc, char **argv)
                chroma_energy / (WIDTH * HEIGHT), chroma_energy_pred / (WIDTH * HEIGHT));
 #endif
 
-       // DC coefficient pred from the right to left
-       for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-               for (unsigned xb = 0; xb < WIDTH - 8; xb += 8) {
-                       coeff_y[yb * WIDTH + xb] -= coeff_y[yb * WIDTH + (xb + 8)];
+       // DC coefficient pred from the right to left (within each slice)
+       for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += 320) {
+               int prev_k = 128;
+
+               for (unsigned subblock_idx = 320; subblock_idx --> 0; ) {
+                       unsigned yb = (block_idx + subblock_idx) / WIDTH_BLOCKS;
+                       unsigned xb = (block_idx + subblock_idx) % WIDTH_BLOCKS;
+                       int k = coeff_y[(yb * 8) * WIDTH + (xb * 8)];
+
+                       coeff_y[(yb * 8) * WIDTH + (xb * 8)] = k - prev_k;
+
+                       prev_k = k;
                }
        }
-       for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-               for (unsigned xb = 0; xb < WIDTH/2 - 8; xb += 8) {
-                       coeff_cb[yb * WIDTH/2 + xb] -= coeff_cb[yb * WIDTH/2 + (xb + 8)];
-                       coeff_cr[yb * WIDTH/2 + xb] -= coeff_cr[yb * WIDTH/2 + (xb + 8)];
+       for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += 320) {
+               int prev_k_cb = 0;
+               int prev_k_cr = 0;
+
+               for (unsigned subblock_idx = 320; subblock_idx --> 0; ) {
+                       unsigned yb = (block_idx + subblock_idx) / WIDTH_BLOCKS_CHROMA;
+                       unsigned xb = (block_idx + subblock_idx) % WIDTH_BLOCKS_CHROMA;
+                       int k_cb = coeff_cb[(yb * 8) * WIDTH/2 + (xb * 8)];
+                       int k_cr = coeff_cr[(yb * 8) * WIDTH/2 + (xb * 8)];
+
+                       coeff_cb[(yb * 8) * WIDTH/2 + (xb * 8)] = k_cb - prev_k_cb;
+                       coeff_cr[(yb * 8) * WIDTH/2 + (xb * 8)] = k_cr - prev_k_cr;
+
+                       prev_k_cb = k_cb;
+                       prev_k_cr = k_cr;
                }
        }
 
@@ -848,21 +872,20 @@ int main(int argc, char **argv)
                        // need to reverse later
                        rans_encoder.clear();
                        size_t num_bytes = 0;
-                       for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-                               for (unsigned xb = 0; xb < WIDTH; xb += 8) {
-                                       int k = coeff_y[(yb + y) * WIDTH + (xb + x)];
-                                       //printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
-                                       rans_encoder.encode_coeff(k);
-                               }
-                               if (yb % 16 == 8) {
+                       for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; ++block_idx) {
+                               unsigned yb = block_idx / WIDTH_BLOCKS;
+                               unsigned xb = block_idx % WIDTH_BLOCKS;
+
+                               int k = coeff_y[(yb * 8 + y) * WIDTH + (xb * 8 + x)];
+                               //printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
+                               rans_encoder.encode_coeff(k);
+
+                               if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
                                        int l = rans_encoder.save_block(codedfp);
                                        num_bytes += l;
                                        lens.push_back(l);
                                }
                        }
-                       if (HEIGHT % 16 != 0) {
-                               num_bytes += rans_encoder.save_block(codedfp);
-                       }
                        tot_bytes += num_bytes;
                        printf("coeff %d Y': %ld bytes\n", y * 8 + x, num_bytes);
 
@@ -889,18 +912,18 @@ int main(int argc, char **argv)
 
                        rans_encoder.clear();
                        size_t num_bytes = 0;
-                       for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-                               for (unsigned xb = 0; xb < WIDTH/2; xb += 8) {
-                                       int k = coeff_cb[(yb + y) * WIDTH/2 + (xb + x)];
-                                       rans_encoder.encode_coeff(k);
-                               }
-                               if (yb % 16 == 8) {
+                       for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; ++block_idx) {
+                               unsigned yb = block_idx / WIDTH_BLOCKS_CHROMA;
+                               unsigned xb = block_idx % WIDTH_BLOCKS_CHROMA;
+
+                               int k = coeff_cb[(yb * 8 + y) * WIDTH/2 + (xb * 8 + x)];
+                               //printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
+                               rans_encoder.encode_coeff(k);
+
+                               if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
                                        num_bytes += rans_encoder.save_block(codedfp);
                                }
                        }
-                       if (HEIGHT % 16 != 0) {
-                               num_bytes += rans_encoder.save_block(codedfp);
-                       }
                        tot_bytes += num_bytes;
                        printf("coeff %d Cb: %ld bytes\n", y * 8 + x, num_bytes);
                }
@@ -914,18 +937,18 @@ int main(int argc, char **argv)
 
                        rans_encoder.clear();
                        size_t num_bytes = 0;
-                       for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
-                               for (unsigned xb = 0; xb < WIDTH/2; xb += 8) {
-                                       int k = coeff_cr[(yb + y) * WIDTH/2 + (xb + x)];
-                                       rans_encoder.encode_coeff(k);
-                               }
-                               if (yb % 16 == 8) {
+                       for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; ++block_idx) {
+                               unsigned yb = block_idx / WIDTH_BLOCKS_CHROMA;
+                               unsigned xb = block_idx % WIDTH_BLOCKS_CHROMA;
+
+                               int k = coeff_cr[(yb * 8 + y) * WIDTH/2 + (xb * 8 + x)];
+                               //printf("encoding coeff %d xb,yb=%d,%d: %d\n", y*8+x, xb, yb, k);
+                               rans_encoder.encode_coeff(k);
+
+                               if (block_idx % 320 == 319 || block_idx == NUM_BLOCKS - 1) {
                                        num_bytes += rans_encoder.save_block(codedfp);
                                }
                        }
-                       if (HEIGHT % 16 != 0) {
-                               num_bytes += rans_encoder.save_block(codedfp);
-                       }
                        tot_bytes += num_bytes;
                        printf("coeff %d Cr: %ld bytes\n", y * 8 + x, num_bytes);
                }