X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=narabu-encoder.cpp;h=ca285282110480de405eadd04acf495898d2be6c;hb=f3eac72e679d3e0ec7ae1d4484736cd552c344dd;hp=b1b86d847a3f2cc128430c22a975a36adc986618;hpb=4e20a14f8ca0bc3259fa2be5bbbd4057080ce62c;p=narabu

diff --git a/narabu-encoder.cpp b/narabu-encoder.cpp
index b1b86d8..ca28528 100644
--- a/narabu-encoder.cpp
+++ b/narabu-encoder.cpp
@@ -19,8 +19,6 @@
 
 #include <movit/util.h>
 
-#include "ryg_rans/rans_byte.h"
-#include "ryg_rans/renormalize.h"
 #include "util.h"
 
 #define WIDTH 1280
@@ -45,7 +43,11 @@ unsigned char pix_cr[(WIDTH/2) * HEIGHT];
 
 struct RansDistSSBO {
 	unsigned dist[4 * 256];
-	std::pair<unsigned, unsigned> ransdist[4 * 256];
+	unsigned ransfreq[4 * 256];
+	struct {
+		uint32_t x_max, rcp_freq, bias, rcp_shift_and_cmpl_freq;
+	} ransdist[4 * 256];
+	unsigned sign_biases[4];
 };
 
 using namespace std;
@@ -196,7 +198,7 @@ int main(int argc, char **argv)
 	GLuint ssbo;
 	glGenBuffers(1, &ssbo);
 	glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
-	glNamedBufferStorage(ssbo, 256 * 16 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+	glNamedBufferStorage(ssbo, sizeof(RansDistSSBO), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 	check_error();
 
 	// SSBOs for the rANS output (data and offsets).
@@ -263,11 +265,11 @@ int main(int argc, char **argv)
 		check_error();
 	}
 
-	glBindImageTexture(0, dc_ac7_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R16UI);
-	glBindImageTexture(1, ac1_ac6_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R16UI);
-	glBindImageTexture(2, ac2_ac5_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R16UI);
-	glBindImageTexture(3, ac3_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R8I);
-	glBindImageTexture(4, ac4_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R8I);
+	glBindImageTexture(0, dc_ac7_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
+	glBindImageTexture(1, ac1_ac6_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
+	glBindImageTexture(2, ac2_ac5_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
+	glBindImageTexture(3, ac3_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R8I);
+	glBindImageTexture(4, ac4_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R8I);
 	glBindImageTexture(5, y_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8UI);
 	check_error();
 
@@ -302,7 +304,7 @@ int main(int argc, char **argv)
 	steady_clock::time_point start = steady_clock::now();
 	unsigned num_iterations = 100;
 	for (unsigned i = 0; i < num_iterations; ++i) {
-		glClearNamedBufferSubData(ssbo, GL_R8, 0, 256 * 16 * sizeof(uint32_t), GL_RED, GL_UNSIGNED_BYTE, nullptr);
+		glClearNamedBufferSubData(ssbo, GL_R8, 0, sizeof(RansDistSSBO), GL_RED, GL_UNSIGNED_BYTE, nullptr);
 		glUseProgram(glsl_program_num);
 		glDispatchCompute(WIDTH_BLOCKS / 16, HEIGHT_BLOCKS, 1);
 		glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
@@ -338,14 +340,20 @@ int main(int argc, char **argv)
 	}
 
 	// Write out the distributions.
-	const RansDistSSBO *rans_dist = (const RansDistSSBO *)glMapNamedBufferRange(ssbo, 0, 256 * 16 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+	const RansDistSSBO *rans_dist = (const RansDistSSBO *)glMapNamedBufferRange(ssbo, 0, sizeof(RansDistSSBO), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 	for (unsigned r = 0; r < 2; ++r) {  // Hack to write fake chroma tables.
 		// TODO: rather gamma-k or something
 		for (unsigned i = 0; i < 4; ++i) {
 			printf("writing table %d\n", i);
 			for (unsigned j = 0; j < NUM_SYMS; ++j) {
-				printf("%d,%d: start=%d freq=%d\n", i, j, rans_dist->ransdist[i * 256 + j].first, rans_dist->ransdist[i * 256 + j].second);
-				write_varint(rans_dist->ransdist[i * 256 + j].second, codedfp);
+				printf("%d,%d: freq=%d  x_max=%d, rcp_freq=%08x, bias=%d, rcp_shift=%d, cmpl_freq=%d\n",
+					i, j, rans_dist->ransfreq[i * 256 + j],
+					rans_dist->ransdist[i * 256 + j].x_max,
+					rans_dist->ransdist[i * 256 + j].rcp_freq,
+					rans_dist->ransdist[i * 256 + j].bias,
+					rans_dist->ransdist[i * 256 + j].rcp_shift_and_cmpl_freq & 0xffff,
+					rans_dist->ransdist[i * 256 + j].rcp_shift_and_cmpl_freq >> 16);
+				write_varint(rans_dist->ransfreq[i * 256 + j], codedfp);
 			}
 		}
 	}