X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=qdc.cpp;h=a0919bd750d476e0e6aace471532fafedf4df94c;hb=7417c69ce221dac7361f5982809a3833c75b5f6b;hp=951915225f8784ff1ede25c8978db48ea6395a2a;hpb=0d22d7fb73e0c14431a15c7bb0522d0691452a31;p=narabu

diff --git a/qdc.cpp b/qdc.cpp
index 9519152..a0919bd 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -18,8 +18,15 @@
 
 #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,
@@ -212,7 +219,7 @@ public:
 			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[256];
+		sign_bias = s.cum_freqs[NUM_SYMS];
 	}
 
 	void clear()
@@ -270,7 +277,7 @@ public:
 			RansEncPut(&rans, &ptr, k, 1, prob_bits);
 			k = ESCAPE_LIMIT;
 		}
-		RansEncPutSymbol(&rans, &ptr, &esyms[(k - 1) & 255]);
+		RansEncPutSymbol(&rans, &ptr, &esyms[(k - 1) & (NUM_SYMS - 1)]);
 		if (signed_k < 0) {
 			rans += sign_bias;
 		}
@@ -404,7 +411,7 @@ void find_optimal_stream_assignment(int base)
 	double inv_sum[64];
 	for (unsigned i = 0; i < 64; ++i) {
 		double s = 0.0;
-		for (unsigned k = 0; k < 256; ++k) {
+		for (unsigned k = 0; k < NUM_SYMS; ++k) {
 			s += stats[i + base].freqs[k] + 0.5;
 		}
 		inv_sum[i] = 1.0 / s;
@@ -413,7 +420,7 @@ void find_optimal_stream_assignment(int base)
 	for (unsigned i = 0; i < 64; ++i) {
 		for (unsigned j = 0; j < 64; ++j) {
 			double d = 0.0;
-			for (unsigned k = 0; k < 256; ++k) {
+			for (unsigned k = 0; k < NUM_SYMS; ++k) {
 				double p1 = (stats[i + base].freqs[k] + 0.5) * inv_sum[i];
 				double p2 = (stats[j + base].freqs[k] + 0.5) * inv_sum[j];
 
@@ -714,10 +721,35 @@ 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 += 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;
 		}
 	}
 
@@ -771,7 +803,7 @@ int main(int argc, char **argv)
 						k = ESCAPE_LIMIT;
 						extra_bits += 12;  // escape this one
 					}
-					++s_luma.freqs[(k - 1) & 255];
+					++s_luma.freqs[(k - 1) & (NUM_SYMS - 1)];
 				}
 			}
 			// Chroma
@@ -787,8 +819,8 @@ int main(int argc, char **argv)
 						k_cr = ESCAPE_LIMIT;
 						extra_bits += 12;  // escape this one
 					}
-					++s_chroma.freqs[(k_cb - 1) & 255];
-					++s_chroma.freqs[(k_cr - 1) & 255];
+					++s_chroma.freqs[(k_cb - 1) & (NUM_SYMS - 1)];
+					++s_chroma.freqs[(k_cr - 1) & (NUM_SYMS - 1)];
 				}
 			}
 		}
@@ -803,10 +835,10 @@ int main(int argc, char **argv)
 #endif
 
 	for (unsigned i = 0; i < 64; ++i) {
-		stats[i].freqs[255] /= 2;  // zero, has no sign bits (yes, this is trickery)
+		stats[i].freqs[NUM_SYMS - 1] /= 2;  // zero, has no sign bits (yes, this is trickery)
 		stats[i].normalize_freqs(prob_scale);
-		stats[i].cum_freqs[256] += stats[i].freqs[255];
-		stats[i].freqs[255] *= 2;
+		stats[i].cum_freqs[NUM_SYMS] += stats[i].freqs[NUM_SYMS - 1];
+		stats[i].freqs[NUM_SYMS - 1] *= 2;
 	}
 
 	FILE *codedfp = fopen("coded.dat", "wb");
@@ -842,21 +874,20 @@ 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 % 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);
 
@@ -883,18 +914,18 @@ 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 % 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);
 		}
@@ -908,18 +939,18 @@ 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 % 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);
 		}