X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=qdc.cpp;h=0339e12af0ad07ba2616530b30bd1b2012258745;hb=5e1d27014149311318e97b8e04a6e05ec858e57c;hp=f1ff2542575f866ae4de3c063b2c57acac1586f3;hpb=2161c821c9c65ebae24eff9c2e4809c11c67cd02;p=narabu

diff --git a/qdc.cpp b/qdc.cpp
index f1ff254..0339e12 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -23,8 +23,10 @@
 #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,
@@ -214,7 +216,7 @@ public:
 	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];
@@ -233,22 +235,14 @@ public:
 		//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);
@@ -259,7 +253,7 @@ public:
 
 		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;
 	}
@@ -292,7 +286,7 @@ private:
 	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.
@@ -505,6 +499,9 @@ int main(int argc, char **argv)
 	//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) {
@@ -525,6 +522,11 @@ int main(int argc, char **argv)
 					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);
 				}
@@ -720,10 +722,10 @@ int main(int argc, char **argv)
 #endif
 
 	// 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 = 0;
+	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS; block_idx += BLOCKS_PER_STREAM) {
+		int prev_k = 128;
 
-		for (unsigned subblock_idx = 320; subblock_idx --> 0; ) {
+		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)];
@@ -733,11 +735,11 @@ int main(int argc, char **argv)
 			prev_k = k;
 		}
 	}
-	for (unsigned block_idx = 0; block_idx < NUM_BLOCKS_CHROMA; block_idx += 320) {
+	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 = 320; subblock_idx --> 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)];
@@ -880,7 +882,7 @@ int main(int argc, char **argv)
 				//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) {
+				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);
@@ -920,7 +922,7 @@ int main(int argc, char **argv)
 				//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) {
+				if (block_idx % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || block_idx == NUM_BLOCKS - 1) {
 					num_bytes += rans_encoder.save_block(codedfp);
 				}
 			}
@@ -945,7 +947,7 @@ int main(int argc, char **argv)
 				//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) {
+				if (block_idx % BLOCKS_PER_STREAM == (BLOCKS_PER_STREAM - 1) || block_idx == NUM_BLOCKS - 1) {
 					num_bytes += rans_encoder.save_block(codedfp);
 				}
 			}
@@ -959,4 +961,18 @@ int main(int argc, char **argv)
 		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
 }