Go down to 4 rANS streams instead of 8.

Costs approx 0.8% bitrate, but reduces GPU cost from 1,3 to 1,2 ms
(~8%) due to less L1 cache pressure.

diff --git a/coded.dat b/coded.dat
index 02a6c03e2bb4bc3f8364cd645f468d77baae7e4a..4ecb22c35d8ccb259ab48b7ed930cf5d14aaeb0f 100644 (file)
Binary files a/coded.dat and b/coded.dat differ

diff --git a/decoder.shader b/decoder.shader
index 8b772c70fbefdd228ce6be25c6faff7ae1cea045..09027915633af93b69a883039836bdecc16f8df2 100644 (file)
--- a/decoder.shader
+++ b/decoder.shader
@@ -40,6 +40,16 @@ const uint ff_zigzag_direct[64] = {
     58, 59, 52, 45, 38, 31, 39, 46,
     53, 60, 61, 54, 47, 55, 62, 63
 };
+const uint stream_mapping[64] = {
+       0, 0, 1, 1, 2, 2, 3, 3,
+       0, 0, 1, 2, 2, 2, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 2, 2, 2, 2, 3, 3, 3,
+       2, 2, 2, 2, 3, 3, 3, 3,
+       2, 2, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+};
 
 layout(std430, binding = 9) buffer layoutName
 {
@@ -204,7 +214,7 @@ void main()
        //const uint coeff_num = ff_zigzag_direct[thread_num];
        const uint coeff_num = thread_num;
        const uint stream_num = coeff_num * num_blocks + block_row;
-       const uint model_num = min((coeff_num % 8) + (coeff_num / 8), 7);
+       const uint model_num = stream_mapping[coeff_num];
        const uint sign_bias = sign_bias_per_model[model_num];
 
        // Initialize rANS decoder.

diff --git a/narabu.cpp b/narabu.cpp
index ba1e967e8c95884d36b93e16f6370ad921c1c2f6..1419ff0bc42a57c64600983002dab6136dcec466 100644 (file)
--- a/narabu.cpp
+++ b/narabu.cpp
@@ -23,7 +23,7 @@ using namespace std::chrono;
 const unsigned prob_bits = 12;
 const unsigned prob_scale = 1 << prob_bits;
 const unsigned NUM_SYMS = 256;
-const unsigned NUM_TABLES = 16;
+const unsigned NUM_TABLES = 8;
 
 struct RansDecSymbol {
         unsigned sym_start;

diff --git a/qdc.cpp b/qdc.cpp
index b5594b76856918d9b14bee3f55dd5bcec24fdd7a..951915225f8784ff1ede25c8978db48ea6395a2a 100644 (file)
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -25,7 +25,7 @@
 // of coefficients to rANS probability distributions. This is randomized,
 // so you might want to run it a few times.
 #define FIND_OPTIMAL_STREAM_ASSIGNMENT 0
-#define NUM_CLUSTERS 8
+#define NUM_CLUSTERS 4
 
 static constexpr uint32_t prob_bits = 12;
 static constexpr uint32_t prob_scale = 1 << prob_bits;
@@ -154,12 +154,37 @@ SymbolStats stats[128];
 float kl_dist[64][64];
 #endif
 
+const int luma_mapping[64] = {
+       0, 0, 1, 1, 2, 2, 3, 3,
+       0, 0, 1, 2, 2, 2, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 2, 2, 2, 2, 3, 3, 3,
+       2, 2, 2, 2, 3, 3, 3, 3,
+       2, 2, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+};
+const int chroma_mapping[64] = {
+       0, 1, 1, 2, 2, 2, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       2, 2, 2, 2, 3, 3, 3, 3,
+       2, 2, 2, 3, 3, 3, 3, 3,
+       2, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+};
+
 int pick_stats_for(int x, int y, bool is_chroma)
 {
 #if FIND_OPTIMAL_STREAM_ASSIGNMENT
        return y * 8 + x + is_chroma * 64;
 #else
-       return std::min<int>(x + y, 7) + is_chroma * 8;
+       if (is_chroma) {
+               return chroma_mapping[y * 8 + x] + 4;
+       } else {
+               return luma_mapping[y * 8 + x];
+       }
 #endif
 }
                

diff --git a/qdd.cpp b/qdd.cpp
index a4fa907a297549f03d0860bb016d41fb7bf83420..522957d3f8de83afd17a291d8fb4f4eb460af866 100644 (file)
--- a/qdd.cpp
+++ b/qdd.cpp
@@ -29,7 +29,8 @@ struct RansDecodeTable {
        int cum2sym[prob_scale];
        RansDecSymbol dsyms[NUM_SYMS];
 };
-RansDecodeTable decode_tables[16];
+#define NUM_TABLES 8
+RansDecodeTable decode_tables[NUM_TABLES];
 
 static const unsigned char std_luminance_quant_tbl[64] = {
 #if 0
@@ -55,10 +56,34 @@ static const unsigned char std_luminance_quant_tbl[64] = {
 };
 
 
-int pick_stats_for(int y, int x)
+const int luma_mapping[64] = {
+       0, 0, 1, 1, 2, 2, 3, 3,
+       0, 0, 1, 2, 2, 2, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       1, 2, 2, 2, 2, 3, 3, 3,
+       2, 2, 2, 2, 3, 3, 3, 3,
+       2, 2, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+};
+const int chroma_mapping[64] = {
+       0, 1, 1, 2, 2, 2, 3, 3,
+       1, 1, 2, 2, 2, 3, 3, 3,
+       2, 2, 2, 2, 3, 3, 3, 3,
+       2, 2, 2, 3, 3, 3, 3, 3,
+       2, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+       3, 3, 3, 3, 3, 3, 3, 3,
+};
+
+int pick_stats_for(int x, int y, bool is_chroma)
 {
-       if (x + y >= 7) return 7;
-       return x + y;
+       if (is_chroma) {
+               return chroma_mapping[y * 8 + x] + 4;
+       } else {
+               return luma_mapping[y * 8 + x];
+       }
 }
 
 uint32_t read_varint(FILE *fp)
@@ -103,8 +128,8 @@ int main(void)
                exit(1);
        }
 
-       uint32_t sign_bias[16];
-       for (unsigned table = 0; table < 16; ++table) {
+       uint32_t sign_bias[NUM_TABLES];
+       for (unsigned table = 0; table < NUM_TABLES; ++table) {
                uint32_t cum_freq = 0;
                for (unsigned sym = 0; sym < NUM_SYMS; ++sym) {
                        uint32_t freq = read_varint(fp);
@@ -123,7 +148,7 @@ int main(void)
        // loop over all coefficients
        for (unsigned y = 0; y < 8; ++y) {
                for (unsigned x = 0; x < 8; ++x) {
-                       unsigned tbl = pick_stats_for(x, y);
+                       unsigned tbl = pick_stats_for(x, y, false);
                
                        RansState rans = 0;