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

[D3D12] Experimental 2x resolution scale

Browse Source
This commit is contained in:
Triang3l
2018-12-06 10:19:07 +03:00
parent 132af3e266
commit 9427667a27
173 changed files with 16680 additions and 4985 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
src/xenia/gpu/d3d12/shaders/edram_clear_32bpp.cs.hlsl
View File

@@ -1,25 +1,41 @@
#define XE_EDRAM_WRITE_ONLY
#include "edram_load_store.hlsli"
// Load4/Store4 aren't needed here, but 80x16 threads is over the limit.
[numthreads(40, 16, 1)]
void main(uint3 xe_group_id : SV_GroupID,
uint3 xe_group_thread_id : SV_GroupThreadID,
uint3 xe_thread_id : SV_DispatchThreadID) {
// 1 thread = 2 guest samples (all the pixels of it, no matter whether the
// resolution scale is 1x or 2x).
uint4 clear_rect;
clear_rect.xz = xe_edram_clear_rect & 0xFFFFu;
clear_rect.yw = xe_edram_clear_rect >> 16u;
uint2 sample_index = xe_thread_id.xy;
sample_index.x *= 2u;
[branch] if (any(sample_index < clear_rect.xy) ||
any(sample_index >= clear_rect.zw)) {
bool4 x_in_rect =
sample_index.x >= clear_rect.x && sample_index.x < clear_rect.z;
[branch] if (sample_index.y < clear_rect.y ||
sample_index.y >= clear_rect.w || !any(x_in_rect)) {
return;
}
uint2 tile_sample_index = xe_group_thread_id.xy;
tile_sample_index.x *= 2u;
uint edram_offset = XeEDRAMOffset32bpp(xe_group_id.xy, tile_sample_index);
xe_edram_load_store_dest.Store(edram_offset, xe_edram_clear_color32);
if (sample_index.x + 1u < clear_rect.z) {
xe_edram_load_store_dest.Store(edram_offset + 4u, xe_edram_clear_color32);
[branch] if (XeEDRAMScaleLog2() != 0u) {
[branch] if (x_in_rect.x) {
xe_edram_load_store_dest.Store4(edram_offset,
xe_edram_clear_color32.xxxx);
}
[branch] if (x_in_rect.y) {
xe_edram_load_store_dest.Store4(edram_offset + 16u,
xe_edram_clear_color32.xxxx);
}
} else {
[branch] if (x_in_rect.x) {
xe_edram_load_store_dest.Store(edram_offset, xe_edram_clear_color32);
}
[branch] if (x_in_rect.y) {
xe_edram_load_store_dest.Store(edram_offset + 4u, xe_edram_clear_color32);
}
}
}