raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

graphics: write proven compute buffer fills to guest memory (#1628)

Browse Source 
Avoid readbacks for compute metadata fills

Co-authored-by: nmzik <Nmzik@mail.ru>
This commit is contained in:
Santiago Paquete
2026-10-18 04:42:14 +01:00
committed by GitHub
parent fe3857a171
commit 4059179e9d
3 changed files with 67 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/host_gpu/renderer/render.h
+3 -3
View File
@@ -203,9 +203,9 @@ private:
void ResetBindings();
void UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
const CommandBuffer& buffer);
[[nodiscard]] bool TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
CommandBuffer& command, uint32_t group_x,
uint32_t group_y, uint32_t group_z, uint32_t mode);
[[nodiscard]] bool TryConsumeComputeFill(const ShaderComputeInputInfo& input,
CommandBuffer& command, uint32_t group_x,
uint32_t group_y, uint32_t group_z, uint32_t mode);
RenderContext& m_context;
GraphicsBindings m_graphics_bindings;
src/graphics/host_gpu/renderer/renderCompute.cpp
+21 -9
View File
@@ -128,9 +128,9 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
return true;
}
bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
CommandBuffer& command, uint32_t group_x,
uint32_t group_y, uint32_t group_z, uint32_t mode) {
bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
CommandBuffer& command, uint32_t group_x,
uint32_t group_y, uint32_t group_z, uint32_t mode) {
const auto& program = *input.stage.program;
const auto& resources = *input.stage.resources;
const auto& fill = resources.uniform_fill;
@@ -186,14 +186,26 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
size)) {
return false;
}
if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
return false;
const bool image_clear =
cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear);
if (!image_clear) {
// CPU fills invalidate overlapping cached images. Preserve the native dispatch for
// partial or ambiguous image backings, including pooled metadata aliases.
if ((descriptor.Base48() & 3u) != 0 ||
!m_context.IsMapped(descriptor.Base48(), size) ||
cache.IsRegionRegistered(descriptor.Base48(), size)) {
return false;
}
// Metadata need not have an image backing yet. Keep proven fills in guest memory
// when CPU-owned, so color-clear discovery can read them without a GPU download.
m_context.GetBufferCache().FillBuffer(descriptor.Base48(), size, packed_clear, false);
}
static std::atomic<uint32_t> logged_clears {0};
if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
LOGF("GraphicsRenderDispatchDirect: compute image clear shader=0x%016" PRIx64
LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
" addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
descriptor.Base48(), size, packed_clear);
}
return true;
}
@@ -308,8 +320,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
if (resources.specialization_reads.empty()) {
// Cached metadata addresses can be reused as ordinary buffers; preserve their guest writes.
UpdateComputeMetaClears(input_info, buffer);
if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
thread_group_z, mode)) {
if (TryConsumeComputeFill(input_info, buffer, thread_group_x, thread_group_y,
thread_group_z, mode)) {
ResetBindings();
return;
}