VideoCore: Implement DispatchIndirect

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);
-            return true;
+        };
-        }
+        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);
+                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;
+        }
+        unsafe_process();
     }
-
     return true;
 }
 

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;

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;

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);
 

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:

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)                                                  \

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);

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();

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]); });

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);

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,