fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[D3D12] Use typed buffers for texture untiling, calculate guest address once per thread

Browse Source
This commit is contained in:
Triang3l
2020-07-03 20:03:53 +03:00
parent 1d35ba9ee7
commit 713f26b5c8
169 changed files with 49015 additions and 22864 deletions
Download Patch File Download Diff File Expand all files Collapse all files

119
src/xenia/gpu/d3d12/shaders/texture_load_ctx1.cs.hlsl
View File

@@ -1,92 +1,71 @@
#include "pixel_formats.hlsli"
#include "texture_load.hlsli"
Buffer<uint4> xe_texture_load_source : register(t0);
RWBuffer<uint4> xe_texture_load_dest : register(u0);
// http://fileadmin.cs.lth.se/cs/Personal/Michael_Doggett/talks/unc-xenos-doggett.pdf
// CXT1 is like DXT3/5 color, but 2-component and with 8:8 endpoints rather than
// 5:6:5.
//
// Dword 1:
// rrrrrrrr gggggggg
// RRRRRRRR GGGGGGGG
// rrrrrrrrgggggggg
// RRRRRRRRGGGGGGGG
// (R is in the higher bits, according to how this format is used in Halo 3).
// Dword 2:
// AA BB CC DD
// EE FF GG HH
// II JJ KK LL
// MM NN OO PP
void XeCTX1FourBlocksRowToR8G8(uint4 end_low_rr00gg00, uint4 end_high_rr00gg00,
uint4 weights_high, out uint4 row_01,
out uint4 row_23) {
uint4 weights_low = ~weights_high;
uint4 row_3aaaa = (weights_low & 3u) * end_low_rr00gg00 +
(weights_high & 3u) * end_high_rr00gg00;
uint4 row_3bbbb = ((weights_low >> 2u) & 3u) * end_low_rr00gg00 +
((weights_high >> 2u) & 3u) * end_high_rr00gg00;
uint4 row_3cccc = ((weights_low >> 4u) & 3u) * end_low_rr00gg00 +
((weights_high >> 4u) & 3u) * end_high_rr00gg00;
uint4 row_3dddd = ((weights_low >> 6u) & 3u) * end_low_rr00gg00 +
((weights_high >> 6u) & 3u) * end_high_rr00gg00;
uint4 row_half_3acac = uint4(row_3aaaa.xy, row_3cccc.xy).xzyw;
uint4 row_half_3bdbd = uint4(row_3bbbb.xy, row_3dddd.xy).xzyw;
// R0A G0A R0B G0B | R0C G0C R0D G0D | R1A G1A R1B G1B | R1C G1C R1D G1D
row_01 = ((row_half_3acac & 0xFFFFu) / 3u) |
(((row_half_3acac >> 16u) / 3u) << 8u) |
(((row_half_3bdbd & 0xFFFFu) / 3u) << 16u) |
(((row_half_3bdbd >> 16u) / 3u) << 24u);
row_half_3acac = uint4(row_3aaaa.zw, row_3cccc.zw).xzyw;
row_half_3bdbd = uint4(row_3bbbb.zw, row_3dddd.zw).xzyw;
// R2A G2A R2B G2B | R2C G2C R2D G2D | R3A G3A R3B G3B | R3C G3C R3D G3D
row_23 = ((row_half_3acac & 0xFFFFu) / 3u) |
(((row_half_3acac >> 16u) / 3u) << 8u) |
(((row_half_3bdbd & 0xFFFFu) / 3u) << 16u) |
(((row_half_3bdbd >> 16u) / 3u) << 24u);
}
[numthreads(8, 32, 1)]
void main(uint3 xe_thread_id : SV_DispatchThreadID) {
// 1 thread = 4 CTX1 (8bpb) blocks to 16x4 R8G8 texels.
uint3 block_index = xe_thread_id;
block_index.x <<= 2u;
// 1 thread = 4 CTX1 blocks to 16x4 R8G8 texels.
uint3 block_index = xe_thread_id << uint3(2, 0, 0);
[branch] if (any(block_index >= xe_texture_load_size_blocks)) {
return;
}
uint4 block_offsets_guest =
XeTextureLoadGuestBlockOffsets(block_index, 8u, 3u);
uint4 blocks_01 = uint4(xe_texture_load_source.Load2(block_offsets_guest.x),
xe_texture_load_source.Load2(block_offsets_guest.y));
uint4 blocks_23 = uint4(xe_texture_load_source.Load2(block_offsets_guest.z),
xe_texture_load_source.Load2(block_offsets_guest.w));
blocks_01 = XeByteSwap(blocks_01, xe_texture_load_endianness);
blocks_23 = XeByteSwap(blocks_23, xe_texture_load_endianness);
// Unpack the endpoints as:
// 0x00g000r0 0x00g100r1 0x00g200r2 0x00g300r3
// 0x00G000R0 0x00G100R1 0x00G200R2 0x00G300R3
// so they can be multiplied by their weights allowing overflow
// (R is in the higher bits, according to how this format is used in Halo 3).
uint4 end_packed = uint4(blocks_01.xz, blocks_23.xz);
uint4 end_low_rr00gg00 =
((end_packed & 0xFF00u) >> 8u) | ((end_packed & 0xFFu) << 16u);
uint4 end_high_rr00gg00 = (end_packed >> 24u) | (end_packed & 0xFF0000u);
// Sort the color indices so they can be used as weights for the second
// endpoint.
uint4 weights_high = XeDXTHighColorWeights(uint4(blocks_01.yw, blocks_23.yw));
// Uncompress and write the rows.
uint3 texel_index_host = block_index << uint3(2u, 2u, 0u);
uint texel_offset_host = XeTextureHostLinearOffset(
texel_index_host, xe_texture_load_height_texels,
xe_texture_load_host_pitch, 2u) + xe_texture_load_host_base;
for (uint i = 0u; i < 4u; ++i) {
uint4 row_01, row_23;
XeCTX1FourBlocksRowToR8G8(end_low_rr00gg00, end_high_rr00gg00,
weights_high >> (i * 8u), row_01, row_23);
xe_texture_load_dest.Store4(texel_offset_host, row_01);
xe_texture_load_dest.Store4(texel_offset_host + 16u, row_23);
if (++texel_index_host.y >= xe_texture_load_height_texels) {
return;
int3 texel_index_host = int3(block_index) << int3(2, 2, 0);
int block_offset_host =
(XeTextureHostLinearOffset(texel_index_host,
xe_texture_load_height_texels,
xe_texture_load_host_pitch, 2u) +
xe_texture_load_host_base) >> 4;
int elements_pitch_host = xe_texture_load_host_pitch >> 4;
int block_offset_guest =
XeTextureLoadGuestBlockOffset(int3(block_index), 8u, 3u) >> 4;
uint endian = XeTextureLoadEndian();
int i;
[unroll] for (i = 0; i < 2; ++i) {
if (i) {
++block_offset_host;
// Odd 2 blocks = even 2 blocks + 32 bytes when tiled.
block_offset_guest += XeTextureLoadIsTiled() ? 2 : 1;
}
// Two blocks.
uint4 blocks = XeByteSwap(xe_texture_load_source[block_offset_guest],
endian);
// Unpack the endpoints as 0x00g000r0 0x00G000R0 0x00g100r1 0x00G100R1 so
// they can be multiplied by their weights allowing overflow.
uint4 end_8in16;
end_8in16.xz = ((blocks.xz >> 8u) & 0xFFu) | ((blocks.xz & 0xFFu) << 16u);
end_8in16.yw = (blocks.xz >> 24u) | (blocks.xz & 0xFF0000u);
uint2 weights_high = XeDXTHighColorWeights(blocks.yw);
xe_texture_load_dest[block_offset_host] =
XeCTX1TwoBlocksRowToR8G8(end_8in16, weights_high);
[branch] if (texel_index_host.y + 1 < int(xe_texture_load_height_texels)) {
xe_texture_load_dest[block_offset_host + elements_pitch_host] =
XeCTX1TwoBlocksRowToR8G8(end_8in16, weights_high >> 8u);
[branch] if (texel_index_host.y + 2 <
int(xe_texture_load_height_texels)) {
xe_texture_load_dest[block_offset_host + 2 * elements_pitch_host] =
XeCTX1TwoBlocksRowToR8G8(end_8in16, weights_high >> 16u);
[branch] if (texel_index_host.y + 3 <
int(xe_texture_load_height_texels)) {
xe_texture_load_dest[block_offset_host + 3 * elements_pitch_host] =
XeCTX1TwoBlocksRowToR8G8(end_8in16, weights_high >> 24u);
}
}
}
texel_offset_host += xe_texture_load_host_pitch;
}
}