shared float temp[64];
+layout(std430, binding = 9) buffer layoutName
+{
+ uint dist[4][256];
+};
+
+#define MAPPING(s0, s1, s2, s3, s4, s5, s6, s7) ((s0) | (s1 << 2) | (s2 << 4) | (s3 << 6) | (s4 << 8) | (s5 << 10) | (s6 << 12) | (s7 << 14))
+
+const uint luma_mapping[8] = {
+ MAPPING(0, 0, 1, 1, 2, 2, 3, 3),
+ MAPPING(0, 0, 1, 2, 2, 2, 3, 3),
+ MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
+ MAPPING(1, 1, 2, 2, 2, 3, 3, 3),
+ MAPPING(1, 2, 2, 2, 2, 3, 3, 3),
+ MAPPING(2, 2, 2, 2, 3, 3, 3, 3),
+ MAPPING(2, 2, 3, 3, 3, 3, 3, 3),
+ MAPPING(3, 3, 3, 3, 3, 3, 3, 3),
+};
+
// Scale factors; 1.0 / (sqrt(2.0) * cos(k * M_PI / 16.0)), except for the first which is 1.
const float sf[8] = {
1.0, 0.7209598220069479, 0.765366864730180, 0.8504300947672564,
imageStore(ac2_ac5_tex, ivec2(sx, y + n), uvec4(pack_9_7(c2, c5), 0, 0, 0));
imageStore(ac3_tex, ivec2(sx, y + n), ivec4(c3, 0, 0, 0));
imageStore(ac4_tex, ivec2(sx, y + n), ivec4(c4, 0, 0, 0));
+
+ // Count frequencies, but only for every 8th block or so, randomly selected.
+ uint wg_index = gl_WorkGroupID.y * gl_WorkGroupSize.x + gl_WorkGroupID.x;
+ if ((wg_index * 0x9E3779B9u) >> 29 == 0) { // Fibonacci hashing, essentially a PRNG in this context.
+ c0 = min(abs(c0), 255);
+ c1 = min(abs(c1), 255);
+ c2 = min(abs(c2), 255);
+ c3 = min(abs(c3), 255);
+ c4 = min(abs(c4), 255);
+ c5 = min(abs(c5), 255);
+ c6 = min(abs(c6), 255);
+ c7 = min(abs(c7), 255);
+
+ // Spread out the most popular elements among the cache lines by reversing the bits
+ // of the index, reducing false sharing.
+ c0 = bitfieldReverse(c0) >> 24;
+ c1 = bitfieldReverse(c1) >> 24;
+ c2 = bitfieldReverse(c2) >> 24;
+ c3 = bitfieldReverse(c3) >> 24;
+ c4 = bitfieldReverse(c4) >> 24;
+ c5 = bitfieldReverse(c5) >> 24;
+ c6 = bitfieldReverse(c6) >> 24;
+ c7 = bitfieldReverse(c7) >> 24;
+
+ uint m = luma_mapping[n];
+ atomicAdd(dist[bitfieldExtract(m, 0, 2)][c0], 1);
+ atomicAdd(dist[bitfieldExtract(m, 2, 2)][c1], 1);
+ atomicAdd(dist[bitfieldExtract(m, 4, 2)][c2], 1);
+ atomicAdd(dist[bitfieldExtract(m, 6, 2)][c3], 1);
+ atomicAdd(dist[bitfieldExtract(m, 8, 2)][c4], 1);
+ atomicAdd(dist[bitfieldExtract(m, 10, 2)][c5], 1);
+ atomicAdd(dist[bitfieldExtract(m, 12, 2)][c6], 1);
+ atomicAdd(dist[bitfieldExtract(m, 14, 2)][c7], 1);
+ }
}
glUseProgram(glsl_program_num);
+ // An SSBO for the rANS distributions.
+ GLuint ssbo;
+ glGenBuffers(1, &ssbo);
+ glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
+ glBufferData(GL_SHADER_STORAGE_BUFFER, 65536 * 4 * sizeof(uint32_t), nullptr, GL_DYNAMIC_COPY);
+ glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 9, ssbo);
+
// Upload luma.
GLuint y_tex;
glGenTextures(1, &y_tex);
check_error();
steady_clock::time_point start = steady_clock::now();
- unsigned num_iterations = 1000;
+ unsigned num_iterations = 100;
for (unsigned i = 0; i < num_iterations; ++i) {
glDispatchCompute(WIDTH_BLOCKS, HEIGHT_BLOCKS, 1);
}
printf("\n");
printf("Each iteration took %.3f ms (but note that is DCT only, no rANS).\n", 1e3 * duration<double>(now - start).count() / num_iterations);
+#if 1
+ glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
+ const uint32_t *dist = (const uint32_t *)glMapBuffer(GL_SHADER_STORAGE_BUFFER, GL_READ_ONLY);
+ for (int i = 0; i < 1024; ++i) {
+ printf("%d,%d: %u\n", i / 256, i % 256, dist[i] / num_iterations);
+ }
+#endif
}