Prepare for more flexible slices.

[narabu] / decoder.shader

diff --git a/decoder.shader b/decoder.shader
index 512d743cb1307b2932f5a198c81f2cb562a4174a..09027915633af93b69a883039836bdecc16f8df2 100644 (file)
--- a/decoder.shader
+++ b/decoder.shader
@@ -1,12 +1,15 @@
-#version 430
+#version 440
 #extension GL_ARB_shader_clock : enable
 
+#define PARALLEL_SLICES 1
+
 #define ENABLE_TIMING 0
 
-layout(local_size_x = 8, local_size_y = 8) in;
+layout(local_size_x = 64*PARALLEL_SLICES) in;
 layout(r8ui) uniform restrict readonly uimage2D cum2sym_tex;
 layout(rg16ui) uniform restrict readonly uimage2D dsyms_tex;
 layout(r8) uniform restrict writeonly image2D out_tex;
+layout(r16i) uniform restrict writeonly iimage2D coeff_tex;
 
 const uint prob_bits = 12;
 const uint prob_scale = 1 << prob_bits;
@@ -37,6 +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
 {
@@ -58,18 +71,10 @@ uniform uint sign_bias_per_model[16];
 
 const uint RANS_BYTE_L = (1u << 23);  // lower bound of our normalization interval
 
-uint last_offset = -1, ransbuf;
-
 uint get_rans_byte(uint offset)
 {
-       if (last_offset != (offset >> 2)) {
-               last_offset = offset >> 2;
-               ransbuf = data_SSBO[offset >> 2];
-       }
-       return bitfieldExtract(ransbuf, 8 * int(offset & 3u), 8);
-
        // We assume little endian.
-//     return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
+       return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
 }
 
 uint RansDecInit(inout uint offset)
@@ -162,7 +167,7 @@ void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, ino
        y7 = p6_0 - p6_7;
 }
 
-shared float temp[64 * 8];
+shared float temp[64 * 8 * PARALLEL_SLICES];
 
 void pick_timer(inout uvec2 start, inout uvec2 t)
 {
@@ -193,18 +198,23 @@ void main()
        }
        uvec2 start = clock2x32ARB();
 #else
-       uvec2 start;
+       uvec2 start = uvec2(0, 0);
+       local_timing[0] = start;
 #endif
 
+       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;
+
        const uint num_blocks = 720 / 16;  // FIXME: make a uniform
-       const uint thread_num = gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
+       const uint slice_num = local_z;
+       const uint thread_num = local_y * 8 + local_x;
 
-       const uint block_row = gl_WorkGroupID.y;
+       const uint block_row = gl_WorkGroupID.y * PARALLEL_SLICES + slice_num;
        //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 stream_num = block_row * num_blocks + coeff_num;  // HACK
-       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.
@@ -252,7 +262,13 @@ void main()
                                last_k = k;
                        }
 
-                       temp[subblock_idx * 64 + coeff_num] = k * q;
+#if 0
+                       uint y = block_row * 16 + block_y * 8 + local_y;
+                       uint x = block_x * 64 + subblock_idx * 8 + local_x;
+                       imageStore(coeff_tex, ivec2(x, y), ivec4(k, 0,0,0));
+#endif
+
+                       temp[slice_num * 64 * 8 + subblock_idx * 64 + coeff_num] = k * q;
                        //temp[subblock_idx * 64 + 8 * y + x] = (2 * k * w * 4) / 32;  // 100% matching unquant
                }
 
@@ -264,14 +280,14 @@ void main()
                pick_timer(start, local_timing[3]);
 
                // Horizontal DCT one row (so 64 rows).
-               idct_1d(temp[thread_num * 8 + 0],
-                       temp[thread_num * 8 + 1],
-                       temp[thread_num * 8 + 2],
-                       temp[thread_num * 8 + 3],
-                       temp[thread_num * 8 + 4],
-                       temp[thread_num * 8 + 5],
-                       temp[thread_num * 8 + 6],
-                       temp[thread_num * 8 + 7]);
+               idct_1d(temp[slice_num * 64 * 8 + thread_num * 8 + 0],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 1],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 2],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 3],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 4],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 5],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 6],
+                       temp[slice_num * 64 * 8 + thread_num * 8 + 7]);
 
                pick_timer(start, local_timing[4]);
 
@@ -281,7 +297,7 @@ void main()
                pick_timer(start, local_timing[5]);
 
                // Vertical DCT one row (so 64 columns).
-               uint row_offset = gl_LocalInvocationID.y * 64 + gl_LocalInvocationID.x;
+               uint row_offset = local_z * 64 * 8 + local_y * 64 + local_x;
                idct_1d(temp[row_offset + 0 * 8],
                        temp[row_offset + 1 * 8],
                        temp[row_offset + 2 * 8],
@@ -294,7 +310,7 @@ void main()
                pick_timer(start, local_timing[6]);
 
                uint y = block_row * 16 + block_y * 8;
-               uint x = block_x * 64 + gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
+               uint x = block_x * 64 + local_y * 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));
                }