From 8beda6a2bf4da0c9673a0532d8e80c9e5e534ced Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sat, 26 Aug 2023 22:00:43 +0200
Subject: [PATCH 1/3] DMA Pusher: Fix regression caused by guest memory
optimizations
---
src/video_core/dma_pusher.cpp | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 9f1b340a9..ab28951b6 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -83,6 +83,14 @@ bool DmaPusher::Step() {
dma_state.dma_get, command_list_header.size * sizeof(u32));
}
}
+ if (Settings::IsGPULevelHigh() && dma_state.method < MacroRegistersStart) {
+ Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+ Core::Memory::GuestMemoryFlags::SafeRead>
+ headers(memory_manager, dma_state.dma_get, command_list_header.size,
+ &command_headers);
+ ProcessCommands(headers);
+ return true;
+ }
Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
Core::Memory::GuestMemoryFlags::UnsafeRead>
headers(memory_manager, dma_state.dma_get, command_list_header.size, &command_headers);
From 710ca3ca494e1af4ce5481f650e67c75c235be83 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 27 Aug 2023 02:00:48 +0200
Subject: [PATCH 2/3] Shader Recompiler: Auto stub special registers and dump
pipelines on exception.
---
.../translate/impl/move_special_register.cpp | 3 ++-
.../renderer_vulkan/vk_pipeline_cache.cpp | 13 +++++++++++++
2 files changed, 15 insertions(+), 1 deletion(-)
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
index 753c62098..e593132e6 100644
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/move_special_register.cpp
@@ -161,7 +161,8 @@ enum class SpecialRegister : u64 {
LOG_WARNING(Shader, "(STUBBED) SR_AFFINITY");
return ir.Imm32(0); // This is the default value hardware returns.
default:
- throw NotImplementedException("S2R special register {}", special_register);
+ LOG_CRITICAL(Shader, "(STUBBED) Special register {}", special_register);
+ return ir.Imm32(0); // This is the default value hardware returns.
}
}
} // Anonymous namespace
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index c1314ca99..b1730a170 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -664,6 +664,19 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline(
std::move(modules), infos);
} catch (const Shader::Exception& exception) {
+ auto hash = key.Hash();
+ size_t env_index{0};
+ for (size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
+ if (key.unique_hashes[index] == 0) {
+ continue;
+ }
+ Shader::Environment& env{*envs[env_index]};
+ ++env_index;
+
+ const u32 cfg_offset{static_cast<u32>(env.StartAddress() + sizeof(Shader::ProgramHeader))};
+ Shader::Maxwell::Flow::CFG cfg(env, pools.flow_block, cfg_offset, index == 0);
+ env.Dump(hash, key.unique_hashes[index]);
+ }
LOG_ERROR(Render_Vulkan, "{}", exception.what());
return nullptr;
}
From 115792158d3ac4ca746d1775f2381e8f8dd18582 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 27 Aug 2023 02:58:00 +0200
Subject: [PATCH 3/3] VideoCore: Implement DispatchIndirect
---
src/video_core/dma_pusher.cpp | 28 +++++++++++++++----
src/video_core/dma_pusher.h | 5 +++-
src/video_core/engines/engine_interface.h | 8 ++++++
src/video_core/engines/engine_upload.h | 8 ++++++
src/video_core/engines/kepler_compute.cpp | 20 ++++++++++++-
src/video_core/engines/kepler_compute.h | 17 +++++++++++
src/video_core/engines/puller.cpp | 15 ++++++----
.../renderer_opengl/gl_rasterizer.cpp | 11 ++++++++
.../renderer_vulkan/vk_rasterizer.cpp | 14 ++++++++++
.../vulkan_common/vulkan_wrapper.cpp | 1 +
src/video_core/vulkan_common/vulkan_wrapper.h | 5 ++++
11 files changed, 119 insertions(+), 13 deletions(-)
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index ab28951b6..58ce0d8c2 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -14,6 +14,7 @@
namespace Tegra {
constexpr u32 MacroRegistersStart = 0xE00;
+constexpr u32 ComputeInline = 0x6D;
DmaPusher::DmaPusher(Core::System& system_, GPU& gpu_, MemoryManager& memory_manager_,
Control::ChannelState& channel_state_)
@@ -83,20 +84,35 @@ bool DmaPusher::Step() {
dma_state.dma_get, command_list_header.size * sizeof(u32));
}
}
- if (Settings::IsGPULevelHigh() && dma_state.method < MacroRegistersStart) {
+ const auto safe_process = [&] {
Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
Core::Memory::GuestMemoryFlags::SafeRead>
headers(memory_manager, dma_state.dma_get, command_list_header.size,
&command_headers);
ProcessCommands(headers);
+ };
+ const auto unsafe_process = [&] {
+ Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+ Core::Memory::GuestMemoryFlags::UnsafeRead>
+ headers(memory_manager, dma_state.dma_get, command_list_header.size,
+ &command_headers);
+ ProcessCommands(headers);
+ };
+ if (Settings::IsGPULevelHigh()) {
+ if (dma_state.method >= MacroRegistersStart) {
+ unsafe_process();
+ return true;
+ }
+ if (subchannel_type[dma_state.subchannel] == Engines::EngineTypes::KeplerCompute &&
+ dma_state.method == ComputeInline) {
+ unsafe_process();
+ return true;
+ }
+ safe_process();
return true;
}
- Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
- Core::Memory::GuestMemoryFlags::UnsafeRead>
- headers(memory_manager, dma_state.dma_get, command_list_header.size, &command_headers);
- ProcessCommands(headers);
+ unsafe_process();
}
-
return true;
}
diff --git a/src/video_core/dma_pusher.h b/src/video_core/dma_pusher.h
index 8a2784cdc..c9fab2d90 100644
--- a/src/video_core/dma_pusher.h
+++ b/src/video_core/dma_pusher.h
@@ -130,8 +130,10 @@ public:
void DispatchCalls();
- void BindSubchannel(Engines::EngineInterface* engine, u32 subchannel_id) {
+ void BindSubchannel(Engines::EngineInterface* engine, u32 subchannel_id,
+ Engines::EngineTypes engine_type) {
subchannels[subchannel_id] = engine;
+ subchannel_type[subchannel_id] = engine_type;
}
void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
@@ -170,6 +172,7 @@ private:
const bool ib_enable{true}; ///< IB mode enabled
std::array<Engines::EngineInterface*, max_subchannels> subchannels{};
+ std::array<Engines::EngineTypes, max_subchannels> subchannel_type;
GPU& gpu;
Core::System& system;
diff --git a/src/video_core/engines/engine_interface.h b/src/video_core/engines/engine_interface.h
index 392322358..54631ee6c 100644
--- a/src/video_core/engines/engine_interface.h
+++ b/src/video_core/engines/engine_interface.h
@@ -11,6 +11,14 @@
namespace Tegra::Engines {
+enum class EngineTypes : u32 {
+ KeplerCompute,
+ Maxwell3D,
+ Fermi2D,
+ MaxwellDMA,
+ KeplerMemory,
+};
+
class EngineInterface {
public:
virtual ~EngineInterface() = default;
diff --git a/src/video_core/engines/engine_upload.h b/src/video_core/engines/engine_upload.h
index 7242d2529..21bf8aeb4 100644
--- a/src/video_core/engines/engine_upload.h
+++ b/src/video_core/engines/engine_upload.h
@@ -69,6 +69,14 @@ public:
/// Binds a rasterizer to this engine.
void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
+ GPUVAddr ExecTargetAddress() const {
+ return regs.dest.Address();
+ }
+
+ u32 GetUploadSize() const {
+ return copy_size;
+ }
+
private:
void ProcessData(std::span<const u8> read_buffer);
diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp
index a38d9528a..cd61ab222 100644
--- a/src/video_core/engines/kepler_compute.cpp
+++ b/src/video_core/engines/kepler_compute.cpp
@@ -43,16 +43,33 @@ void KeplerCompute::CallMethod(u32 method, u32 method_argument, bool is_last_cal
switch (method) {
case KEPLER_COMPUTE_REG_INDEX(exec_upload): {
+ UploadInfo info{.upload_address = upload_address,
+ .exec_address = upload_state.ExecTargetAddress(),
+ .copy_size = upload_state.GetUploadSize()};
+ uploads.push_back(info);
upload_state.ProcessExec(regs.exec_upload.linear != 0);
break;
}
case KEPLER_COMPUTE_REG_INDEX(data_upload): {
+ upload_address = current_dma_segment;
upload_state.ProcessData(method_argument, is_last_call);
break;
}
- case KEPLER_COMPUTE_REG_INDEX(launch):
+ case KEPLER_COMPUTE_REG_INDEX(launch): {
+ const GPUVAddr launch_desc_loc = regs.launch_desc_loc.Address();
+
+ for (auto& data : uploads) {
+ const GPUVAddr offset = data.exec_address - launch_desc_loc;
+ if (offset / sizeof(u32) == LAUNCH_REG_INDEX(grid_dim_x) &&
+ memory_manager.IsMemoryDirty(data.upload_address, data.copy_size)) {
+ indirect_compute = {data.upload_address};
+ }
+ }
+ uploads.clear();
ProcessLaunch();
+ indirect_compute = std::nullopt;
break;
+ }
default:
break;
}
@@ -62,6 +79,7 @@ void KeplerCompute::CallMultiMethod(u32 method, const u32* base_start, u32 amoun
u32 methods_pending) {
switch (method) {
case KEPLER_COMPUTE_REG_INDEX(data_upload):
+ upload_address = current_dma_segment;
upload_state.ProcessData(base_start, amount);
return;
default:
diff --git a/src/video_core/engines/kepler_compute.h b/src/video_core/engines/kepler_compute.h
index 2092e685f..735e05fb4 100644
--- a/src/video_core/engines/kepler_compute.h
+++ b/src/video_core/engines/kepler_compute.h
@@ -5,6 +5,7 @@
#include <array>
#include <cstddef>
+#include <optional>
#include <vector>
#include "common/bit_field.h"
#include "common/common_funcs.h"
@@ -36,6 +37,9 @@ namespace Tegra::Engines {
#define KEPLER_COMPUTE_REG_INDEX(field_name) \
(offsetof(Tegra::Engines::KeplerCompute::Regs, field_name) / sizeof(u32))
+#define LAUNCH_REG_INDEX(field_name) \
+ (offsetof(Tegra::Engines::KeplerCompute::LaunchParams, field_name) / sizeof(u32))
+
class KeplerCompute final : public EngineInterface {
public:
explicit KeplerCompute(Core::System& system, MemoryManager& memory_manager);
@@ -201,6 +205,10 @@ public:
void CallMultiMethod(u32 method, const u32* base_start, u32 amount,
u32 methods_pending) override;
+ std::optional<GPUVAddr> GetIndirectComputeAddress() const {
+ return indirect_compute;
+ }
+
private:
void ProcessLaunch();
@@ -216,6 +224,15 @@ private:
MemoryManager& memory_manager;
VideoCore::RasterizerInterface* rasterizer = nullptr;
Upload::State upload_state;
+ GPUVAddr upload_address;
+
+ struct UploadInfo {
+ GPUVAddr upload_address;
+ GPUVAddr exec_address;
+ u32 copy_size;
+ };
+ std::vector<UploadInfo> uploads;
+ std::optional<GPUVAddr> indirect_compute{};
};
#define ASSERT_REG_POSITION(field_name, position) \
diff --git a/src/video_core/engines/puller.cpp b/src/video_core/engines/puller.cpp
index 7718a09b3..6de2543b7 100644
--- a/src/video_core/engines/puller.cpp
+++ b/src/video_core/engines/puller.cpp
@@ -34,19 +34,24 @@ void Puller::ProcessBindMethod(const MethodCall& method_call) {
bound_engines[method_call.subchannel] = engine_id;
switch (engine_id) {
case EngineID::FERMI_TWOD_A:
- dma_pusher.BindSubchannel(channel_state.fermi_2d.get(), method_call.subchannel);
+ dma_pusher.BindSubchannel(channel_state.fermi_2d.get(), method_call.subchannel,
+ EngineTypes::Fermi2D);
break;
case EngineID::MAXWELL_B:
- dma_pusher.BindSubchannel(channel_state.maxwell_3d.get(), method_call.subchannel);
+ dma_pusher.BindSubchannel(channel_state.maxwell_3d.get(), method_call.subchannel,
+ EngineTypes::Maxwell3D);
break;
case EngineID::KEPLER_COMPUTE_B:
- dma_pusher.BindSubchannel(channel_state.kepler_compute.get(), method_call.subchannel);
+ dma_pusher.BindSubchannel(channel_state.kepler_compute.get(), method_call.subchannel,
+ EngineTypes::KeplerCompute);
break;
case EngineID::MAXWELL_DMA_COPY_A:
- dma_pusher.BindSubchannel(channel_state.maxwell_dma.get(), method_call.subchannel);
+ dma_pusher.BindSubchannel(channel_state.maxwell_dma.get(), method_call.subchannel,
+ EngineTypes::MaxwellDMA);
break;
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
- dma_pusher.BindSubchannel(channel_state.kepler_memory.get(), method_call.subchannel);
+ dma_pusher.BindSubchannel(channel_state.kepler_memory.get(), method_call.subchannel,
+ EngineTypes::KeplerMemory);
break;
default:
UNIMPLEMENTED_MSG("Unimplemented engine {:04X}", engine_id);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 1ba31be88..dd03efecd 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -380,6 +380,17 @@ void RasterizerOpenGL::DispatchCompute() {
pipeline->SetEngine(kepler_compute, gpu_memory);
pipeline->Configure();
const auto& qmd{kepler_compute->launch_description};
+ auto indirect_address = kepler_compute->GetIndirectComputeAddress();
+ if (indirect_address) {
+ // DispatchIndirect
+ static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
+ const auto post_op = VideoCommon::ObtainBufferOperation::DiscardWrite;
+ const auto [buffer, offset] =
+ buffer_cache.ObtainBuffer(*indirect_address, 12, sync_info, post_op);
+ glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, buffer->Handle());
+ glDispatchComputeIndirect(static_cast<GLintptr>(offset));
+ return;
+ }
glDispatchCompute(qmd.grid_dim_x, qmd.grid_dim_y, qmd.grid_dim_z);
++num_queued_commands;
has_written_global_memory |= pipeline->WritesGlobalMemory();
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 032f694bc..01e76a82c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -463,6 +463,20 @@ void RasterizerVulkan::DispatchCompute() {
pipeline->Configure(*kepler_compute, *gpu_memory, scheduler, buffer_cache, texture_cache);
const auto& qmd{kepler_compute->launch_description};
+ auto indirect_address = kepler_compute->GetIndirectComputeAddress();
+ if (indirect_address) {
+ // DispatchIndirect
+ static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
+ const auto post_op = VideoCommon::ObtainBufferOperation::DiscardWrite;
+ const auto [buffer, offset] =
+ buffer_cache.ObtainBuffer(*indirect_address, 12, sync_info, post_op);
+ scheduler.RequestOutsideRenderPassOperationContext();
+ scheduler.Record([indirect_buffer = buffer->Handle(),
+ indirect_offset = offset](vk::CommandBuffer cmdbuf) {
+ cmdbuf.DispatchIndirect(indirect_buffer, indirect_offset);
+ });
+ return;
+ }
const std::array<u32, 3> dim{qmd.grid_dim_x, qmd.grid_dim_y, qmd.grid_dim_z};
scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([dim](vk::CommandBuffer cmdbuf) { cmdbuf.Dispatch(dim[0], dim[1], dim[2]); });
diff --git a/src/video_core/vulkan_common/vulkan_wrapper.cpp b/src/video_core/vulkan_common/vulkan_wrapper.cpp
index 78e5a248f..c3f388d89 100644
--- a/src/video_core/vulkan_common/vulkan_wrapper.cpp
+++ b/src/video_core/vulkan_common/vulkan_wrapper.cpp
@@ -92,6 +92,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
X(vkCmdCopyImage);
X(vkCmdCopyImageToBuffer);
X(vkCmdDispatch);
+ X(vkCmdDispatchIndirect);
X(vkCmdDraw);
X(vkCmdDrawIndexed);
X(vkCmdDrawIndirect);
diff --git a/src/video_core/vulkan_common/vulkan_wrapper.h b/src/video_core/vulkan_common/vulkan_wrapper.h
index c226a2a29..049fa8038 100644
--- a/src/video_core/vulkan_common/vulkan_wrapper.h
+++ b/src/video_core/vulkan_common/vulkan_wrapper.h
@@ -203,6 +203,7 @@ struct DeviceDispatch : InstanceDispatch {
PFN_vkCmdCopyImage vkCmdCopyImage{};
PFN_vkCmdCopyImageToBuffer vkCmdCopyImageToBuffer{};
PFN_vkCmdDispatch vkCmdDispatch{};
+ PFN_vkCmdDispatchIndirect vkCmdDispatchIndirect{};
PFN_vkCmdDraw vkCmdDraw{};
PFN_vkCmdDrawIndexed vkCmdDrawIndexed{};
PFN_vkCmdDrawIndirect vkCmdDrawIndirect{};
@@ -1209,6 +1210,10 @@ public:
dld->vkCmdDispatch(handle, x, y, z);
}
+ void DispatchIndirect(VkBuffer indirect_buffer, VkDeviceSize offset) const noexcept {
+ dld->vkCmdDispatchIndirect(handle, indirect_buffer, offset);
+ }
+
void PipelineBarrier(VkPipelineStageFlags src_stage_mask, VkPipelineStageFlags dst_stage_mask,
VkDependencyFlags dependency_flags, Span<VkMemoryBarrier> memory_barriers,
Span<VkBufferMemoryBarrier> buffer_barriers,