X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=encoder.shader;h=163f8fa0110aed25ff9654156f115b54167d95e5;hb=4ec1e96782990c7fad993a4edaa24e071f7dcc85;hp=54103a093d374cca0fb2fcda37a7add308574268;hpb=20ada0b5ea3436888d737e18dc5d867942d5581e;p=narabu diff --git a/encoder.shader b/encoder.shader index 54103a0..163f8fa 100644 --- a/encoder.shader +++ b/encoder.shader @@ -12,6 +12,30 @@ layout(r8ui) uniform restrict readonly uimage2D image_tex; 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, + 1.0, 1.2727585805728336, 1.847759065022573, 3.6245097854115502 +}; + const float W[64] = { 8, 16, 19, 22, 26, 27, 29, 34, 16, 16, 22, 24, 27, 29, 34, 37, @@ -22,18 +46,19 @@ const float W[64] = { 26, 27, 29, 34, 38, 46, 56, 69, 27, 29, 35, 38, 46, 56, 69, 83 }; -const float S = 4.0; // whatever? +const float S = 4.0 * 0.5; // whatever? // NOTE: Contains factors to counteract the scaling in the DCT implementation. +#define QM(x, y) (sf[x] * sf[y] / (W[y*8 + x] * S)) const float quant_matrix[64] = { - 1.0 / 64.0, 1.0 / (W[ 1] * S), 1.0 / (W[ 2] * S), 1.0 / (W[ 3] * S), 1.0 / (W[ 4] * S), 1.0 / (W[ 5] * S), 1.0 / (W[ 6] * S), 1.0 / (W[ 7] * S), - 1.0 / (W[ 8] * S), 2.0 / (W[ 9] * S), 2.0 / (W[10] * S), 2.0 / (W[11] * S), 2.0 / (W[12] * S), 2.0 / (W[13] * S), 2.0 / (W[14] * S), 2.0 / (W[15] * S), - 1.0 / (W[16] * S), 2.0 / (W[17] * S), 2.0 / (W[18] * S), 2.0 / (W[19] * S), 2.0 / (W[20] * S), 2.0 / (W[21] * S), 2.0 / (W[22] * S), 2.0 / (W[23] * S), - 1.0 / (W[24] * S), 2.0 / (W[25] * S), 2.0 / (W[26] * S), 2.0 / (W[27] * S), 2.0 / (W[28] * S), 2.0 / (W[29] * S), 2.0 / (W[30] * S), 2.0 / (W[31] * S), - 1.0 / (W[32] * S), 2.0 / (W[33] * S), 2.0 / (W[34] * S), 2.0 / (W[35] * S), 2.0 / (W[36] * S), 2.0 / (W[37] * S), 2.0 / (W[38] * S), 2.0 / (W[39] * S), - 1.0 / (W[40] * S), 2.0 / (W[41] * S), 2.0 / (W[42] * S), 2.0 / (W[43] * S), 2.0 / (W[44] * S), 2.0 / (W[45] * S), 2.0 / (W[46] * S), 2.0 / (W[47] * S), - 1.0 / (W[48] * S), 2.0 / (W[49] * S), 2.0 / (W[50] * S), 2.0 / (W[51] * S), 2.0 / (W[52] * S), 2.0 / (W[53] * S), 2.0 / (W[54] * S), 2.0 / (W[55] * S), - 1.0 / (W[56] * S), 2.0 / (W[57] * S), 2.0 / (W[58] * S), 2.0 / (W[59] * S), 2.0 / (W[60] * S), 2.0 / (W[61] * S), 2.0 / (W[62] * S), 2.0 / (W[63] * S) + 1.0 / 64.0, QM(1, 0), QM(2, 0), QM(3, 0), QM(4, 0), QM(5, 0), QM(6, 0), QM(7, 0), + QM(0, 1), QM(1, 1), QM(2, 1), QM(3, 1), QM(4, 1), QM(5, 1), QM(6, 1), QM(7, 1), + QM(0, 2), QM(1, 2), QM(2, 2), QM(3, 2), QM(4, 2), QM(5, 2), QM(6, 2), QM(7, 2), + QM(0, 3), QM(1, 3), QM(2, 3), QM(3, 3), QM(4, 3), QM(5, 3), QM(6, 3), QM(7, 3), + QM(0, 4), QM(1, 4), QM(2, 4), QM(3, 4), QM(4, 4), QM(5, 4), QM(6, 4), QM(7, 4), + QM(0, 5), QM(1, 5), QM(2, 5), QM(3, 5), QM(4, 5), QM(5, 5), QM(6, 5), QM(7, 5), + QM(0, 6), QM(1, 6), QM(2, 6), QM(3, 6), QM(4, 6), QM(5, 6), QM(6, 6), QM(7, 6), + QM(0, 7), QM(1, 7), QM(2, 7), QM(3, 7), QM(4, 7), QM(5, 7), QM(6, 7), QM(7, 7) }; // Clamp and pack a 9-bit and a 7-bit signed value into a 16-bit word. @@ -42,7 +67,7 @@ uint pack_9_7(int v9, int v7) return (uint(clamp(v9, -256, 255)) & 0x1ffu) | ((uint(clamp(v7, -64, 63)) & 0x7fu) << 9); } -// Scaled 1D DCT. y0 output is scaled by 8, everything else is scaled by 16. +// Scaled 1D DCT (AA&N). y0 is correctly scaled, all other y_k are scaled by sqrt(2) cos(k * Pi / 16). void dct_1d(inout float y0, inout float y1, inout float y2, inout float y3, inout float y4, inout float y5, inout float y6, inout float y7) { const float a1 = 0.7071067811865474; // sqrt(2) @@ -97,6 +122,7 @@ void dct_1d(inout float y0, inout float y1, inout float y2, inout float y3, inou y7 = p5_5 - p4_6; y3 = p5_7 - p4_4; } + void main() { uint x = 8 * gl_WorkGroupID.x; @@ -159,5 +185,39 @@ void main() 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); + } }