-#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(r32i) uniform restrict writeonly iimage2D coeff_tex;
+layout(r32i) uniform restrict writeonly iimage2D coeff2_tex;
+uniform int num_blocks;
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;
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
{
};
struct CoeffStream {
- uint src_offset, src_len, sign_offset, sign_len, extra_bits;
+ uint src_offset, src_len;
};
layout(std430, binding = 0) buffer whatever3
{
CoeffStream streams[];
};
+uniform uint sign_bias_per_model[16];
-uniform uint src_offset, src_len, sign_offset, sign_len, extra_bits;
-
-const uint RANS_BYTE_L = (1u << 23); // lower bound of our normalization interval
+struct myuint64 {
+ uint high, low;
+};
-uint last_offset = -1, ransbuf;
+const uint RANS64_L = (1u << 31); // lower bound of our normalization interval
-uint get_rans_byte(uint offset)
+myuint64 RansDecInit(inout uint offset)
{
- if (last_offset != (offset >> 2)) {
- last_offset = offset >> 2;
- ransbuf = data_SSBO[offset >> 2];
- }
- return bitfieldExtract(ransbuf, 8 * int(offset & 3u), 8);
+ myuint64 x;
+ x.low = data_SSBO[offset++];
+ x.high = data_SSBO[offset++];
+ return x;
+}
- // We assume little endian.
-// return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
+uint RansDecGet(myuint64 r, uint scale_bits)
+{
+ return r.low & ((1u << scale_bits) - 1);
}
-void RansDecInit(out uint r, inout uint offset)
+void RansDecAdvance(inout myuint64 rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
{
- uint x;
+ const uint mask = (1u << prob_bits) - 1;
+ const uint recovered_lowbits = (rans.low & mask) - start;
- x = get_rans_byte(offset);
- x |= get_rans_byte(offset + 1) << 8;
- x |= get_rans_byte(offset + 2) << 16;
- x |= get_rans_byte(offset + 3) << 24;
- offset += 4;
+ // rans >>= prob_bits;
+ rans.low = (rans.low >> prob_bits) | ((rans.high & mask) << (32 - prob_bits));
+ rans.high >>= prob_bits;
- r = x;
-}
+ // rans *= freq;
+ uint h1, l1, h2, l2;
+ umulExtended(rans.low, freq, h1, l1);
+ umulExtended(rans.high, freq, h2, l2);
+ rans.low = l1;
+ rans.high = l2 + h1;
-uint RansDecGet(uint r, uint scale_bits)
-{
- return r & ((1u << scale_bits) - 1);
-}
+ // rans += recovered_lowbits;
+ uint carry;
+ rans.low = uaddCarry(rans.low, recovered_lowbits, carry);
+ rans.high += carry;
-void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
-{
- const uint mask = (1u << prob_bits) - 1;
- rans = freq * (rans >> prob_bits) + (rans & mask) - start;
-
// renormalize
- while (rans < RANS_BYTE_L) {
- rans = (rans << 8) | get_rans_byte(offset++);
+ if (rans.high == 0 && rans.low < RANS64_L) {
+ rans.high = rans.low;
+ rans.low = data_SSBO[offset++];
}
}
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)
{
}
uvec2 start = clock2x32ARB();
#else
- uvec2 start;
+ uvec2 start = uvec2(0, 0);
+ local_timing[0] = start;
#endif
- const uint num_blocks = 720 / 16; // FIXME: make a uniform
- const uint thread_num = gl_LocalInvocationID.y * 8 + gl_LocalInvocationID.x;
+ 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;
+
+ 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.
- uint offset = streams[stream_num].src_offset;
- uint rans;
- RansDecInit(rans, offset);
-
- // Initialize sign bit decoder. TODO: this ought to be 32-bit-aligned instead!
- uint soffset = streams[stream_num].sign_offset;
- uint sign_buf = get_rans_byte(soffset++) >> streams[stream_num].extra_bits;
- uint sign_bits_left = 8 - streams[stream_num].extra_bits;
+ uint offset = streams[stream_num].src_offset >> 2;
+ myuint64 rans = RansDecInit(offset);
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).
for (uint subblock_idx = 8; subblock_idx --> 0; ) {
// Read a symbol.
- int k = int(cum2sym(RansDecGet(rans, prob_bits), model_num));
+ uint bottom_bits = RansDecGet(rans, prob_bits + 1);
+ bool sign = false;
+ if (bottom_bits >= sign_bias) {
+ bottom_bits -= sign_bias;
+ rans.low -= sign_bias;
+ sign = true;
+ }
+ int k = int(cum2sym(bottom_bits, model_num)); // Can go out-of-bounds; that will return zero.
uvec2 sym = get_dsym(k, model_num);
- RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits);
+ RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits + 1);
if (k == ESCAPE_LIMIT) {
k = int(RansDecGet(rans, prob_bits));
RansDecAdvance(rans, offset, k, 1, prob_bits);
}
- if (k != 0) {
- if (sign_bits_left == 0) {
- sign_buf = get_rans_byte(soffset++);
- sign_bits_left = 8;
- }
- if ((sign_buf & 1u) == 1u) k = -k;
- --sign_bits_left;
- sign_buf >>= 1;
+ if (sign) {
+ k = -k;
}
+#if 0
+ if (coeff_num == 0) {
+ //imageStore(coeff_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(k, 0,0,0));
+ imageStore(coeff_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(rans.low, 0,0,0));
+ imageStore(coeff2_tex, ivec2((block_row * 40 + block_idx) * 8 + subblock_idx, 0), ivec4(rans.high, 0,0,0));
+ }
+#endif
if (coeff_num == 0) {
k += last_k;
last_k = k;
}
- temp[subblock_idx * 64 + coeff_num] = k * q;
+
+ 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
}
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]);
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],
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 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));
}
projects / narabu / blobdiff