#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)
+#define BLOCKS_PER_STREAM 320
// If you set this to 1, the program will try to optimize the placement
// of coefficients to rANS probability distributions. This is randomized,
void init_prob(SymbolStats &s)
{
for (int i = 0; i < NUM_SYMS; i++) {
- printf("%d: cumfreqs=%d freqs=%d prob_bits=%d\n", i, s.cum_freqs[i], s.freqs[i], prob_bits + 1);
+ //printf("%d: cumfreqs=%d freqs=%d prob_bits=%d\n", i, s.cum_freqs[i], s.freqs[i], prob_bits + 1);
RansEncSymbolInit(&esyms[i], s.cum_freqs[i], s.freqs[i], prob_bits + 1);
}
sign_bias = s.cum_freqs[NUM_SYMS];
//printf("post-flush = %08x\n", rans);
uint32_t num_rans_bytes = out_end - ptr;
-#if 0
- if (num_rans_bytes == 4) {
- uint32_t block;
- memcpy(&block, ptr, 4);
-
- if (block == last_block) {
- write_varint(0, codedfp);
- clear();
- return 1;
- }
-
- last_block = block;
+ if (num_rans_bytes == last_block.size() &&
+ memcmp(last_block.data(), ptr, last_block.size()) == 0) {
+ write_varint(0, codedfp);
+ clear();
+ return 1;
} else {
- last_block = 0;
+ last_block = string((const char *)ptr, num_rans_bytes);
}
-#endif
write_varint(num_rans_bytes, codedfp);
//fwrite(&num_rans_bytes, 1, 4, codedfp);
clear();
- printf("Saving block: %d rANS bytes\n", num_rans_bytes);
+ //printf("Saving block: %d rANS bytes\n", num_rans_bytes);
return num_rans_bytes;
//return num_rans_bytes;
}
RansEncSymbol esyms[NUM_SYMS];
uint32_t sign_bias;
- uint32_t last_block = 0; // Not a valid 4-byte rANS block (?)
+ std::string last_block;
};
static constexpr int dc_scalefac = 8; // Matches the FDCT's gain.
//double last_cb_cfl_fac = 0.0;
//double last_cr_cfl_fac = 0.0;
+ int max_val_x[8] = {0}, min_val_x[8] = {0};
+ int max_val_y[8] = {0}, min_val_y[8] = {0};
+
// DCT and quantize luma
for (unsigned yb = 0; yb < HEIGHT; yb += 8) {
for (unsigned xb = 0; xb < WIDTH; xb += 8) {
int k = quantize(in_y[coeff_idx], coeff_idx);
coeff_y[(yb + y) * WIDTH + (xb + x)] = k;
+ max_val_x[x] = std::max(max_val_x[x], k);
+ min_val_x[x] = std::min(min_val_x[x], k);
+ max_val_y[y] = std::max(max_val_y[y], k);
+ min_val_y[y] = std::min(min_val_y[y], k);
+
// Store back for reconstruction / PSNR calculation
in_y[coeff_idx] = unquantize(k, coeff_idx);
}
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 += BLOCKS_PER_STREAM) {
+ int prev_k = 128;
+
+ for (unsigned subblock_idx = BLOCKS_PER_STREAM; 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 block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += BLOCKS_PER_STREAM) {
+ int prev_k_cb = 0;
+ int prev_k_cr = 0;
+
+ for (unsigned subblock_idx = BLOCKS_PER_STREAM; 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;
}
}
// 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 % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || 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);
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 % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || 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);
}
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 % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || 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);
}
extra_bits,
extra_bits / 8,
tot_bytes);
+
+#if 0
+ printf("Max coefficient ranges (as a function of x):\n\n");
+ for (unsigned x = 0; x < 8; ++x) {
+ int range = std::max(max_val_x[x], -min_val_x[x]);
+ printf(" [%4d, %4d] (%.2f bits)\n", min_val_x[x], max_val_x[x], log2(range * 2 + 1));
+ }
+
+ printf("Max coefficient ranges (as a function of y):\n\n");
+ for (unsigned y = 0; y < 8; ++y) {
+ int range = std::max(max_val_y[y], -min_val_y[y]);
+ printf(" [%4d, %4d] (%.2f bits)\n", min_val_y[y], max_val_y[y], log2(range * 2 + 1));
+ }
+#endif
}
projects / narabu / blobdiff