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

44
src/xenia/gpu/d3d12/shaders/texture_load_128bpb.cs.hlsl
View File

@@ -1,30 +1,28 @@
#include "texture_load.hlsli"
[numthreads(8, 32, 1)]
Buffer<uint4> xe_texture_load_source : register(t0);
RWBuffer<uint4> xe_texture_load_dest : register(u0);
[numthreads(16, 32, 1)]
void main(uint3 xe_thread_id : SV_DispatchThreadID) {
// 1 thread = 4 uint4 blocks.
uint3 block_index = xe_thread_id;
block_index.x <<= 2u;
// 1 thread = 2 blocks.
uint3 block_index = xe_thread_id << uint3(1, 0, 0);
[branch] if (any(block_index >= xe_texture_load_size_blocks)) {
return;
}
uint4 block_offsets_guest =
XeTextureLoadGuestBlockOffsets(block_index, 16u, 4u);
uint4 block_0 = xe_texture_load_source.Load4(block_offsets_guest.x);
uint4 block_1 = xe_texture_load_source.Load4(block_offsets_guest.y);
uint4 block_2 = xe_texture_load_source.Load4(block_offsets_guest.z);
uint4 block_3 = xe_texture_load_source.Load4(block_offsets_guest.w);
block_0 = XeByteSwap(block_0, xe_texture_load_endianness);
block_1 = XeByteSwap(block_1, xe_texture_load_endianness);
block_2 = XeByteSwap(block_2, xe_texture_load_endianness);
block_3 = XeByteSwap(block_3, xe_texture_load_endianness);
uint4 block_offsets_host;
block_offsets_host.x = XeTextureHostLinearOffset(
block_index, xe_texture_load_size_blocks.y, xe_texture_load_host_pitch,
16u) + xe_texture_load_host_base;
block_offsets_host.yzw = uint3(16u, 32u, 48u) + block_offsets_host.x;
xe_texture_load_dest.Store4(block_offsets_host.x, block_0);
xe_texture_load_dest.Store4(block_offsets_host.y, block_1);
xe_texture_load_dest.Store4(block_offsets_host.z, block_2);
xe_texture_load_dest.Store4(block_offsets_host.w, block_3);
int block_offset_host =
(XeTextureHostLinearOffset(int3(block_index),
xe_texture_load_size_blocks.y,
xe_texture_load_host_pitch, 16u) +
xe_texture_load_host_base) >> 4;
int block_offset_guest =
XeTextureLoadGuestBlockOffset(int3(block_index), 16u, 4u) >> 4;
uint endian = XeTextureLoadEndian();
xe_texture_load_dest[block_offset_host] =
XeByteSwap(xe_texture_load_source[block_offset_guest], endian);
++block_offset_host;
// Odd block = even block + 32 bytes when tiled.
block_offset_guest += XeTextureLoadIsTiled() ? 2 : 1;
xe_texture_load_dest[block_offset_host] =
XeByteSwap(xe_texture_load_source[block_offset_guest], endian);
}