mirrors/shadPS4
1
0
Fork 0
mirror of https://github.com/shadps4-emu/shadPS4.git synced 2025-12-10 21:58:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

video_core: Rework tile manager (#3374)

Browse Source 
* video_core: Rework detiling

* video_core: Support tiling and macrotile detiling

* clang format

* image_info: Cleanups

* resource: Revert some changes

* texture_cache: Fix small error

* image_info: Set depth flag on depth promote

* buffer_cache: Remove level check

* tile_manager: Handle case of staging buffer causing flush

* image_info: Add 2D thick array mode

* image_info: Add slices to mip size

* tile_manager: Set bank swizzle

* buffer_cache: Support image copies from DmaData

* vk_rasterizer: Accelerate trivial render target copies with compute

Before tiling PR compute image copies were done with the following sequence

vkCmdCopyImageToBuffer (in SynchronizeBufferFromImage)  -> vkCmdDispatch (copy) -> vkCmdCopyBufferToImage (in RefreshImage)

With the tiling PR it added extra tiling/detiling steps

vkCmdCopyImageToBuffer -> vkCmdDispatch (tiling) -> vkCmdDispatch (copy) -> vkCmdDispatch (detiling) -> vkCmdCopyBufferToImage

This is quite a bit of overhead for a simple image copy. This commit tries to detect trivial image copies i.e cs shaders that copy the full source image to all of the destination.
So now all this sequence is just a vkCmdCopyImage. How much it triggers depends on the guest

* texture_cache: Fix build

* image: Copy all subresources with buffer too
This commit is contained in:
TheTurtle
2025-08-08 15:27:11 +03:00
committed by GitHub
parent befc5ec17b
commit d9108cd39a
30 changed files with 1999 additions and 797 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/video_core/amdgpu/liverpool.h
View File

@@ -22,7 +22,7 @@
#include "common/unique_function.h"
#include "shader_recompiler/params.h"
#include "video_core/amdgpu/pixel_format.h"
#include "video_core/amdgpu/resource.h"
#include "video_core/amdgpu/tiling.h"
#include "video_core/amdgpu/types.h"
namespace Vulkan {
@@ -426,7 +426,7 @@ struct Liverpool {
BitField<0, 2, ZFormat> format;
BitField<2, 2, u32> num_samples;
BitField<13, 3, u32> tile_split;
BitField<20, 3, u32> tile_mode_index;
BitField<20, 3, TileMode> tile_mode_index;
BitField<23, 4, u32> decompress_on_n_zplanes;
BitField<27, 1, u32> allow_expclear;
BitField<28, 1, u32> read_size;
@@ -502,6 +502,14 @@ struct Liverpool {
const auto bpe = NumBits() >> 3; // in bytes
return (depth_slice.tile_max + 1) * 64 * bpe * NumSamples();
}
TileMode GetTileMode() const {
return z_info.tile_mode_index.Value();
}
bool IsTiled() const {
return GetTileMode() != TileMode::DisplayLinearAligned;
}
};
enum class ClipSpace : u32 {
@@ -888,7 +896,7 @@ struct Liverpool {
u32 u32all;
} info;
union Color0Attrib {
BitField<0, 5, TilingMode> tile_mode_index;
BitField<0, 5, TileMode> tile_mode_index;
BitField<5, 5, u32> fmask_tile_mode_index;
BitField<10, 2, u32> fmask_bank_height;
BitField<12, 3, u32> num_samples_log2;
@@ -949,13 +957,13 @@ struct Liverpool {
return slice_size;
}
TilingMode GetTilingMode() const {
return info.linear_general ? TilingMode::Display_Linear
TileMode GetTileMode() const {
return info.linear_general ? TileMode::DisplayLinearAligned
: attrib.tile_mode_index.Value();
}
bool IsTiled() const {
return GetTilingMode() != TilingMode::Display_Linear;
return GetTileMode() != TileMode::DisplayLinearAligned;
}
[[nodiscard]] DataFormat GetDataFmt() const {