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

[narabu] / decoder.shader

diff --git a/decoder.shader b/decoder.shader
index e7b83e0ccbb52402dcb0a827119fe272eff892fc..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;
@@ -203,6 +204,8 @@ void main()
        local_timing[0] = start;
 #endif
 
+       const uint blocks_per_row = (imageSize(out_tex).x + 7) / 8;
+
        const uint local_x = gl_LocalInvocationID.x % 8;
        const uint local_y = (gl_LocalInvocationID.x / 8) % 8;
        const uint local_z = gl_LocalInvocationID.x / 64;
@@ -224,15 +227,11 @@ void main()
        float q = (coeff_num == 0) ? 1.0 : (quant_matrix[coeff_num] * quant_scalefac / 128.0 / sqrt(2.0));  // FIXME: fold
        q *= (1.0 / 255.0);
        //int w = (coeff_num == 0) ? 32 : int(quant_matrix[coeff_num]);
-       int last_k = 0;
+       int last_k = 128;
 
        pick_timer(start, local_timing[0]);
 
-       for (uint block_idx = 40; block_idx --> 0; ) {
-               uint block_x = block_idx % 20;
-               uint block_y = block_idx / 20;
-               if (block_x == 19) last_k = 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).
@@ -309,8 +308,12 @@ void main()
 
                pick_timer(start, local_timing[6]);
 
-               uint y = block_row * 16 + block_y * 8;
-               uint x = block_x * 64 + local_y * 8 + local_x;
+               uint global_block_idx = (block_row * 40 + block_idx) * 8 + local_y;
+               uint block_x = global_block_idx % blocks_per_row;
+               uint block_y = global_block_idx / blocks_per_row;
+
+               uint y = block_y * 8;
+               uint x = block_x * 8 + local_x;
                for (uint yl = 0; yl < 8; ++yl) {
                        imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
                }