raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-23 04:16:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

graphics: draw indexed indirect multi-draws from GPU-read arguments

Browse Source
This commit is contained in:
nmzik
2026-10-22 14:01:25 +02:00
parent c513a4a13a
commit 949de19bb0
5 changed files with 48 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+14 -9
View File
@@ -18659,9 +18659,9 @@ public:
}
}
// Embedded fetches take indirect arguments from the GPU copy, whose vertex offset 7
// selects the covering triangle; the CPU backing holds zeros. Other vertex shaders read
// the arguments on the CPU.
// Embedded fetches take indirect records from the GPU copy: the first draws degenerate
// vertices 0-2, the second's vertex offset 7 selects the covering triangle. The CPU
// backing holds zeros. Other vertex shaders read the records on the CPU.
{
constexpr uint64_t vertices_address = depth_address + 0x38000;
constexpr uint64_t args_address = vertices_address + 0x100;
@@ -18670,14 +18670,17 @@ public:
std::ranges::copy(triangle, positions.begin() + 28);
std::memcpy(reinterpret_cast<void *>(vertices_address), positions.data(),
sizeof(positions));
const vk::DrawIndexedIndirectCommand args{3, 1, 0, 7, 5};
std::memcpy(reinterpret_cast<void *>(args_address), &args, sizeof(args));
constexpr uint32_t stride = 32;
const vk::DrawIndexedIndirectCommand degenerate{3, 1, 0, 0, 5};
const vk::DrawIndexedIndirectCommand covering{3, 1, 0, 7, 5};
std::memcpy(reinterpret_cast<void *>(args_address), &degenerate, sizeof(degenerate));
std::memcpy(reinterpret_cast<void *>(args_address + stride), &covering, sizeof(covering));
// Larger than a caching page, so the arguments are cached instead of streamed.
(void)context.GetBufferCache().ObtainBuffer(
vertices_address, BufferCache::CACHING_PAGESIZE * 2, false);
const vk::DrawIndexedIndirectCommand zero_args{};
const std::array<u32, 2 * stride / 4> zeros{};
Require("GpuIndirectDraw", "CPU backing",
LibKernel::Memory::TryWriteBacking(args_address, &zero_args, sizeof(zero_args)),
LibKernel::Memory::TryWriteBacking(args_address, zeros.data(), sizeof(zeros)),
"the CPU copy of the arguments could not be cleared");
static const auto fetch_vertex = [] {
std::vector<u32> code{EncodeSMovB32(0, InlineU32(0)), EncodeSmem0(0x02u, 20, 4),
@@ -18718,7 +18721,7 @@ public:
}
const DrawIndexArgs indirect{
.index_count = 3, .index_addr = reinterpret_cast<const void *>(index_address),
.index_type_and_size = 1, .indirect_args = args_address};
.index_type_and_size = 1, .indirect = {args_address, 2, stride}};
shaders.SetEsShaderBase(fetch_vs);
TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), sparse_color.image_id,
{vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, {});
@@ -18730,7 +18733,7 @@ public:
for (size_t component = 0; component < pixels.size(); component++) {
Require("GpuIndirectDraw", "vertex offset",
pixels[component] == (component % 4 == 3 ? 0x3f800000u : 0x3e800000u),
"the draw did not take its arguments, vertex offset included, from the GPU");
"the draw did not take every record, vertex offset included, from the GPU");
}
shaders.SetEsShaderBase(indirect_vs);
Require("GpuIndirectDraw", "CPU fallback",
@@ -20773,6 +20776,7 @@ private:
m_rasterization_supported = available_features.fillModeNonSolid &&
available_features.tessellationShader &&
available_features.drawIndirectFirstInstance &&
available_features.multiDrawIndirect &&
available_features.depthBounds &&
available_depth_clip.depthClipEnable &&
available_clip_control.depthClipControl &&
@@ -20866,6 +20870,7 @@ private:
device_features.fillModeNonSolid = m_rasterization_supported;
device_features.tessellationShader = m_rasterization_supported;
device_features.drawIndirectFirstInstance = m_rasterization_supported;
device_features.multiDrawIndirect = m_rasterization_supported;
device_features.depthBounds = m_rasterization_supported;
device_info.pEnabledFeatures = &device_features;
std::vector<const char *> device_extensions{