Make the encoder 100% GPU. Not working yet, though.

[narabu] / decoder.shader

diff --git a/decoder.shader b/decoder.shader
index 724a2f5f38362538adab6750d83297f96c440d24..616c3e2dfd0896e5cd035bacc42ec0720c5174a1 100644 (file)
--- a/decoder.shader
+++ b/decoder.shader
@@ -16,6 +16,7 @@ const uint prob_bits = 12;
 const uint prob_scale = 1 << prob_bits;
 const uint NUM_SYMS = 256;
 const uint ESCAPE_LIMIT = NUM_SYMS - 1;
+const uint BLOCKS_PER_STREAM = 320;
 
 // These need to be folded into quant_matrix.
 const float dc_scalefac = 8.0;
@@ -161,8 +162,8 @@ void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, ino
        y0 = p6_0 + p6_7;
        y1 = p6_1 + p6_6;
        y2 = p6_2 + p6_5;
-       y3 = p6_3 - p4_4;
-       y4 = p6_3 + p4_4;
+       y3 = p6_3 + p4_4;
+       y4 = p6_3 - p4_4;
        y5 = p6_2 - p6_5;
        y6 = p6_1 - p6_6;
        y7 = p6_0 - p6_7;
@@ -230,7 +231,7 @@ void main()
 
        pick_timer(start, local_timing[0]);
 
-       for (uint block_idx = 40; block_idx --> 0; ) {
+       for (uint block_idx = BLOCKS_PER_STREAM / 8; block_idx --> 0; ) {
                pick_timer(start, local_timing[1]);
 
                // rANS decode one coefficient across eight blocks (so 64x8 coefficients).