diff options
Diffstat (limited to 'src/video_core')
27 files changed, 318 insertions, 82 deletions
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp index 324dafdcd..16311f05e 100644 --- a/src/video_core/dma_pusher.cpp +++ b/src/video_core/dma_pusher.cpp @@ -71,16 +71,22 @@ bool DmaPusher::Step() { gpu.MemoryManager().ReadBlockUnsafe(dma_get, command_headers.data(), command_list_header.size * sizeof(u32)); - for (const CommandHeader& command_header : command_headers) { - - // now, see if we're in the middle of a command - if (dma_state.length_pending) { - // Second word of long non-inc methods command - method count - dma_state.length_pending = 0; - dma_state.method_count = command_header.method_count_; - } else if (dma_state.method_count) { + for (std::size_t index = 0; index < command_headers.size();) { + const CommandHeader& command_header = command_headers[index]; + + if (dma_state.method_count) { // Data word of methods command - CallMethod(command_header.argument); + if (dma_state.non_incrementing) { + const u32 max_write = static_cast<u32>( + std::min<std::size_t>(index + dma_state.method_count, command_headers.size()) - + index); + CallMultiMethod(&command_header.argument, max_write); + dma_state.method_count -= max_write; + index += max_write; + continue; + } else { + CallMethod(command_header.argument); + } if (!dma_state.non_incrementing) { dma_state.method++; @@ -120,6 +126,7 @@ bool DmaPusher::Step() { break; } } + index++; } if (!non_main) { @@ -140,4 +147,9 @@ void DmaPusher::CallMethod(u32 argument) const { gpu.CallMethod({dma_state.method, argument, dma_state.subchannel, dma_state.method_count}); } +void DmaPusher::CallMultiMethod(const u32* base_start, u32 num_methods) const { + gpu.CallMultiMethod(dma_state.method, dma_state.subchannel, base_start, num_methods, + dma_state.method_count); +} + } // namespace Tegra diff --git a/src/video_core/dma_pusher.h b/src/video_core/dma_pusher.h index d6188614a..6cef71306 100644 --- a/src/video_core/dma_pusher.h +++ b/src/video_core/dma_pusher.h @@ -75,6 +75,7 @@ private: void SetState(const CommandHeader& command_header); void CallMethod(u32 argument) const; + void CallMultiMethod(const u32* base_start, u32 num_methods) const; std::vector<CommandHeader> command_headers; ///< Buffer for list of commands fetched at once diff --git a/src/video_core/engines/fermi_2d.cpp b/src/video_core/engines/fermi_2d.cpp index bace6affb..8a47614d2 100644 --- a/src/video_core/engines/fermi_2d.cpp +++ b/src/video_core/engines/fermi_2d.cpp @@ -28,6 +28,12 @@ void Fermi2D::CallMethod(const GPU::MethodCall& method_call) { } } +void Fermi2D::CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending) { + for (std::size_t i = 0; i < amount; i++) { + CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)}); + } +} + static std::pair<u32, u32> DelimitLine(u32 src_1, u32 src_2, u32 dst_1, u32 dst_2, u32 src_line) { const u32 line_a = src_2 - src_1; const u32 line_b = dst_2 - dst_1; diff --git a/src/video_core/engines/fermi_2d.h b/src/video_core/engines/fermi_2d.h index dba342c70..939a5966d 100644 --- a/src/video_core/engines/fermi_2d.h +++ b/src/video_core/engines/fermi_2d.h @@ -39,6 +39,9 @@ public: /// Write the value to the register identified by method. void CallMethod(const GPU::MethodCall& method_call); + /// Write multiple values to the register identified by method. + void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending); + enum class Origin : u32 { Center = 0, Corner = 1, diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp index 368c75a66..00a12175f 100644 --- a/src/video_core/engines/kepler_compute.cpp +++ b/src/video_core/engines/kepler_compute.cpp @@ -51,6 +51,13 @@ void KeplerCompute::CallMethod(const GPU::MethodCall& method_call) { } } +void KeplerCompute::CallMultiMethod(u32 method, const u32* base_start, u32 amount, + u32 methods_pending) { + for (std::size_t i = 0; i < amount; i++) { + CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)}); + } +} + Texture::FullTextureInfo KeplerCompute::GetTexture(std::size_t offset) const { const std::bitset<8> cbuf_mask = launch_description.const_buffer_enable_mask.Value(); ASSERT(cbuf_mask[regs.tex_cb_index]); diff --git a/src/video_core/engines/kepler_compute.h b/src/video_core/engines/kepler_compute.h index eeb79c56f..fe55fdfd0 100644 --- a/src/video_core/engines/kepler_compute.h +++ b/src/video_core/engines/kepler_compute.h @@ -202,6 +202,9 @@ public: /// Write the value to the register identified by method. void CallMethod(const GPU::MethodCall& method_call); + /// Write multiple values to the register identified by method. + void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending); + Texture::FullTextureInfo GetTexture(std::size_t offset) const; /// Given a texture handle, returns the TSC and TIC entries. diff --git a/src/video_core/engines/kepler_memory.cpp b/src/video_core/engines/kepler_memory.cpp index 597872e43..586ff15dc 100644 --- a/src/video_core/engines/kepler_memory.cpp +++ b/src/video_core/engines/kepler_memory.cpp @@ -41,4 +41,11 @@ void KeplerMemory::CallMethod(const GPU::MethodCall& method_call) { } } +void KeplerMemory::CallMultiMethod(u32 method, const u32* base_start, u32 amount, + u32 methods_pending) { + for (std::size_t i = 0; i < amount; i++) { + CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)}); + } +} + } // namespace Tegra::Engines diff --git a/src/video_core/engines/kepler_memory.h b/src/video_core/engines/kepler_memory.h index 396fb6e86..bb26fb030 100644 --- a/src/video_core/engines/kepler_memory.h +++ b/src/video_core/engines/kepler_memory.h @@ -40,6 +40,9 @@ public: /// Write the value to the register identified by method. void CallMethod(const GPU::MethodCall& method_call); + /// Write multiple values to the register identified by method. + void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending); + struct Regs { static constexpr size_t NUM_REGS = 0x7F; diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp index 2824ed707..39e3b66a2 100644 --- a/src/video_core/engines/maxwell_3d.cpp +++ b/src/video_core/engines/maxwell_3d.cpp @@ -280,6 +280,58 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) { } } +void Maxwell3D::CallMultiMethod(u32 method, const u32* base_start, u32 amount, + u32 methods_pending) { + // Methods after 0xE00 are special, they're actually triggers for some microcode that was + // uploaded to the GPU during initialization. + if (method >= MacroRegistersStart) { + // We're trying to execute a macro + if (executing_macro == 0) { + // A macro call must begin by writing the macro method's register, not its argument. + ASSERT_MSG((method % 2) == 0, + "Can't start macro execution by writing to the ARGS register"); + executing_macro = method; + } + + for (std::size_t i = 0; i < amount; i++) { + macro_params.push_back(base_start[i]); + } + + // Call the macro when there are no more parameters in the command buffer + if (amount == methods_pending) { + CallMacroMethod(executing_macro, macro_params.size(), macro_params.data()); + macro_params.clear(); + } + return; + } + switch (method) { + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[1]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[2]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[3]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[4]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[5]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[6]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[7]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[8]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[9]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[10]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[11]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[12]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[13]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[14]): + case MAXWELL3D_REG_INDEX(const_buffer.cb_data[15]): { + ProcessCBMultiData(method, base_start, amount); + break; + } + default: { + for (std::size_t i = 0; i < amount; i++) { + CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)}); + } + } + } +} + void Maxwell3D::StepInstance(const MMEDrawMode expected_mode, const u32 count) { if (mme_draw.current_mode == MMEDrawMode::Undefined) { if (mme_draw.gl_begin_consume) { @@ -570,6 +622,28 @@ void Maxwell3D::StartCBData(u32 method) { ProcessCBData(regs.const_buffer.cb_data[cb_data_state.id]); } +void Maxwell3D::ProcessCBMultiData(u32 method, const u32* start_base, u32 amount) { + if (cb_data_state.current != method) { + if (cb_data_state.current != null_cb_data) { + FinishCBData(); + } + constexpr u32 first_cb_data = MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]); + cb_data_state.start_pos = regs.const_buffer.cb_pos; + cb_data_state.id = method - first_cb_data; + cb_data_state.current = method; + cb_data_state.counter = 0; + } + const std::size_t id = cb_data_state.id; + const std::size_t size = amount; + std::size_t i = 0; + for (; i < size; i++) { + cb_data_state.buffer[id][cb_data_state.counter] = start_base[i]; + cb_data_state.counter++; + } + // Increment the current buffer position. + regs.const_buffer.cb_pos = regs.const_buffer.cb_pos + 4 * amount; +} + void Maxwell3D::FinishCBData() { // Write the input value to the current const buffer at the current position. const GPUVAddr buffer_address = regs.const_buffer.BufferAddress(); diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h index 59d5752d2..3dfba8197 100644 --- a/src/video_core/engines/maxwell_3d.h +++ b/src/video_core/engines/maxwell_3d.h @@ -1259,7 +1259,8 @@ public: GPUVAddr LimitAddress() const { return static_cast<GPUVAddr>((static_cast<GPUVAddr>(limit_high) << 32) | - limit_low); + limit_low) + + 1; } } vertex_array_limit[NumVertexArrays]; @@ -1358,6 +1359,9 @@ public: /// Write the value to the register identified by method. void CallMethod(const GPU::MethodCall& method_call); + /// Write multiple values to the register identified by method. + void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending); + /// Write the value to the register identified by method. void CallMethodFromMME(const GPU::MethodCall& method_call); @@ -1511,6 +1515,7 @@ private: /// Handles a write to the CB_DATA[i] register. void StartCBData(u32 method); void ProcessCBData(u32 value); + void ProcessCBMultiData(u32 method, const u32* start_base, u32 amount); void FinishCBData(); /// Handles a write to the CB_BIND register. diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp index 3bfed6ab8..6630005b0 100644 --- a/src/video_core/engines/maxwell_dma.cpp +++ b/src/video_core/engines/maxwell_dma.cpp @@ -36,6 +36,13 @@ void MaxwellDMA::CallMethod(const GPU::MethodCall& method_call) { #undef MAXWELLDMA_REG_INDEX } +void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount, + u32 methods_pending) { + for (std::size_t i = 0; i < amount; i++) { + CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)}); + } +} + void MaxwellDMA::HandleCopy() { LOG_TRACE(HW_GPU, "Requested a DMA copy"); diff --git a/src/video_core/engines/maxwell_dma.h b/src/video_core/engines/maxwell_dma.h index 4f40d1d1f..c43ed8194 100644 --- a/src/video_core/engines/maxwell_dma.h +++ b/src/video_core/engines/maxwell_dma.h @@ -35,6 +35,9 @@ public: /// Write the value to the register identified by method. void CallMethod(const GPU::MethodCall& method_call); + /// Write multiple values to the register identified by method. + void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending); + struct Regs { static constexpr std::size_t NUM_REGS = 0x1D6; diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h index 7231597d4..cde3a26b9 100644 --- a/src/video_core/engines/shader_bytecode.h +++ b/src/video_core/engines/shader_bytecode.h @@ -655,6 +655,7 @@ union Instruction { } constexpr Instruction(u64 value) : value{value} {} + constexpr Instruction(const Instruction& instr) : value(instr.value) {} BitField<0, 8, Register> gpr0; BitField<8, 8, Register> gpr8; @@ -817,11 +818,9 @@ union Instruction { BitField<32, 1, u64> saturate; BitField<49, 2, HalfMerge> merge; - BitField<43, 1, u64> negate_a; BitField<44, 1, u64> abs_a; BitField<47, 2, HalfType> type_a; - BitField<31, 1, u64> negate_b; BitField<30, 1, u64> abs_b; BitField<28, 2, HalfType> type_b; diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp index 3b7572d61..b87fd873d 100644 --- a/src/video_core/gpu.cpp +++ b/src/video_core/gpu.cpp @@ -9,6 +9,7 @@ #include "core/core_timing_util.h" #include "core/frontend/emu_window.h" #include "core/memory.h" +#include "core/settings.h" #include "video_core/engines/fermi_2d.h" #include "video_core/engines/kepler_compute.h" #include "video_core/engines/kepler_memory.h" @@ -154,7 +155,10 @@ u64 GPU::GetTicks() const { constexpr u64 gpu_ticks_den = 625; const u64 cpu_ticks = system.CoreTiming().GetTicks(); - const u64 nanoseconds = Core::Timing::CyclesToNs(cpu_ticks).count(); + u64 nanoseconds = Core::Timing::CyclesToNs(cpu_ticks).count(); + if (Settings::values.use_fast_gpu_time) { + nanoseconds /= 256; + } const u64 nanoseconds_num = nanoseconds / gpu_ticks_den; const u64 nanoseconds_rem = nanoseconds % gpu_ticks_den; return nanoseconds_num * gpu_ticks_num + (nanoseconds_rem * gpu_ticks_num) / gpu_ticks_den; @@ -209,16 +213,32 @@ void GPU::CallMethod(const MethodCall& method_call) { ASSERT(method_call.subchannel < bound_engines.size()); - if (ExecuteMethodOnEngine(method_call)) { + if (ExecuteMethodOnEngine(method_call.method)) { CallEngineMethod(method_call); } else { CallPullerMethod(method_call); } } -bool GPU::ExecuteMethodOnEngine(const MethodCall& method_call) { - const auto method = static_cast<BufferMethods>(method_call.method); - return method >= BufferMethods::NonPullerMethods; +void GPU::CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount, + u32 methods_pending) { + LOG_TRACE(HW_GPU, "Processing method {:08X} on subchannel {}", method, subchannel); + + ASSERT(subchannel < bound_engines.size()); + + if (ExecuteMethodOnEngine(method)) { + CallEngineMultiMethod(method, subchannel, base_start, amount, methods_pending); + } else { + for (std::size_t i = 0; i < amount; i++) { + CallPullerMethod( + {method, base_start[i], subchannel, methods_pending - static_cast<u32>(i)}); + } + } +} + +bool GPU::ExecuteMethodOnEngine(u32 method) { + const auto buffer_method = static_cast<BufferMethods>(method); + return buffer_method >= BufferMethods::NonPullerMethods; } void GPU::CallPullerMethod(const MethodCall& method_call) { @@ -298,6 +318,31 @@ void GPU::CallEngineMethod(const MethodCall& method_call) { } } +void GPU::CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount, + u32 methods_pending) { + const EngineID engine = bound_engines[subchannel]; + + switch (engine) { + case EngineID::FERMI_TWOD_A: + fermi_2d->CallMultiMethod(method, base_start, amount, methods_pending); + break; + case EngineID::MAXWELL_B: + maxwell_3d->CallMultiMethod(method, base_start, amount, methods_pending); + break; + case EngineID::KEPLER_COMPUTE_B: + kepler_compute->CallMultiMethod(method, base_start, amount, methods_pending); + break; + case EngineID::MAXWELL_DMA_COPY_A: + maxwell_dma->CallMultiMethod(method, base_start, amount, methods_pending); + break; + case EngineID::KEPLER_INLINE_TO_MEMORY_B: + kepler_memory->CallMultiMethod(method, base_start, amount, methods_pending); + break; + default: + UNIMPLEMENTED_MSG("Unimplemented engine"); + } +} + void GPU::ProcessBindMethod(const MethodCall& method_call) { // Bind the current subchannel to the desired engine id. LOG_DEBUG(HW_GPU, "Binding subchannel {} to engine {}", method_call.subchannel, diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h index 5e3eb94e9..dd51c95b7 100644 --- a/src/video_core/gpu.h +++ b/src/video_core/gpu.h @@ -155,6 +155,10 @@ public: /// Calls a GPU method. void CallMethod(const MethodCall& method_call); + /// Calls a GPU multivalue method. + void CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount, + u32 methods_pending); + /// Flush all current written commands into the host GPU for execution. void FlushCommands(); /// Synchronizes CPU writes with Host GPU memory. @@ -309,8 +313,12 @@ private: /// Calls a GPU engine method. void CallEngineMethod(const MethodCall& method_call); + /// Calls a GPU engine multivalue method. + void CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount, + u32 methods_pending); + /// Determines where the method should be executed. - bool ExecuteMethodOnEngine(const MethodCall& method_call); + bool ExecuteMethodOnEngine(u32 method); protected: std::unique_ptr<Tegra::DmaPusher> dma_pusher; diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index 4c16c89d2..6fe155bcc 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -186,8 +186,12 @@ void RasterizerOpenGL::SetupVertexBuffer() { const GPUVAddr start = vertex_array.StartAddress(); const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress(); - ASSERT(end > start); - const u64 size = end - start + 1; + ASSERT(end >= start); + const u64 size = end - start; + if (size == 0) { + glBindVertexBuffer(static_cast<GLuint>(index), 0, 0, vertex_array.stride); + continue; + } const auto [vertex_buffer, vertex_buffer_offset] = buffer_cache.UploadMemory(start, size); glBindVertexBuffer(static_cast<GLuint>(index), vertex_buffer, vertex_buffer_offset, vertex_array.stride); @@ -311,8 +315,8 @@ std::size_t RasterizerOpenGL::CalculateVertexArraysSize() const { const GPUVAddr start = regs.vertex_array[index].StartAddress(); const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress(); - ASSERT(end > start); - size += end - start + 1; + size += end - start; + ASSERT(end >= start); } return size; diff --git a/src/video_core/renderer_vulkan/fixed_pipeline_state.h b/src/video_core/renderer_vulkan/fixed_pipeline_state.h index d4fd4d3f1..77188b862 100644 --- a/src/video_core/renderer_vulkan/fixed_pipeline_state.h +++ b/src/video_core/renderer_vulkan/fixed_pipeline_state.h @@ -129,7 +129,7 @@ struct FixedPipelineState { auto& binding = bindings[index]; binding.raw = 0; binding.enabled.Assign(enabled ? 1 : 0); - binding.stride.Assign(stride); + binding.stride.Assign(static_cast<u16>(stride)); binding_divisors[index] = divisor; } diff --git a/src/video_core/renderer_vulkan/vk_device.h b/src/video_core/renderer_vulkan/vk_device.h index a4d841e26..c8640762d 100644 --- a/src/video_core/renderer_vulkan/vk_device.h +++ b/src/video_core/renderer_vulkan/vk_device.h @@ -82,11 +82,6 @@ public: return present_family; } - /// Returns true if the device is integrated with the host CPU. - bool IsIntegrated() const { - return properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU; - } - /// Returns the current Vulkan API version provided in Vulkan-formatted version numbers. u32 GetApiVersion() const { return properties.apiVersion; diff --git a/src/video_core/renderer_vulkan/vk_memory_manager.cpp b/src/video_core/renderer_vulkan/vk_memory_manager.cpp index 6a9e658bf..b4c650a63 100644 --- a/src/video_core/renderer_vulkan/vk_memory_manager.cpp +++ b/src/video_core/renderer_vulkan/vk_memory_manager.cpp @@ -118,8 +118,7 @@ private: }; VKMemoryManager::VKMemoryManager(const VKDevice& device) - : device{device}, properties{device.GetPhysical().GetMemoryProperties()}, - is_memory_unified{GetMemoryUnified(properties)} {} + : device{device}, properties{device.GetPhysical().GetMemoryProperties()} {} VKMemoryManager::~VKMemoryManager() = default; @@ -209,16 +208,6 @@ VKMemoryCommit VKMemoryManager::TryAllocCommit(const VkMemoryRequirements& requi return {}; } -bool VKMemoryManager::GetMemoryUnified(const VkPhysicalDeviceMemoryProperties& properties) { - for (u32 heap_index = 0; heap_index < properties.memoryHeapCount; ++heap_index) { - if (!(properties.memoryHeaps[heap_index].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT)) { - // Memory is considered unified when heaps are device local only. - return false; - } - } - return true; -} - VKMemoryCommitImpl::VKMemoryCommitImpl(const VKDevice& device, VKMemoryAllocation* allocation, const vk::DeviceMemory& memory, u64 begin, u64 end) : device{device}, memory{memory}, interval{begin, end}, allocation{allocation} {} diff --git a/src/video_core/renderer_vulkan/vk_memory_manager.h b/src/video_core/renderer_vulkan/vk_memory_manager.h index 5b6858e9b..1af88e3d4 100644 --- a/src/video_core/renderer_vulkan/vk_memory_manager.h +++ b/src/video_core/renderer_vulkan/vk_memory_manager.h @@ -40,11 +40,6 @@ public: /// Commits memory required by the image and binds it. VKMemoryCommit Commit(const vk::Image& image, bool host_visible); - /// Returns true if the memory allocations are done always in host visible and coherent memory. - bool IsMemoryUnified() const { - return is_memory_unified; - } - private: /// Allocates a chunk of memory. bool AllocMemory(VkMemoryPropertyFlags wanted_properties, u32 type_mask, u64 size); @@ -53,12 +48,8 @@ private: VKMemoryCommit TryAllocCommit(const VkMemoryRequirements& requirements, VkMemoryPropertyFlags wanted_properties); - /// Returns true if the device uses an unified memory model. - static bool GetMemoryUnified(const VkPhysicalDeviceMemoryProperties& properties); - - const VKDevice& device; ///< Device handler. - const VkPhysicalDeviceMemoryProperties properties; ///< Physical device properties. - const bool is_memory_unified; ///< True if memory model is unified. + const VKDevice& device; ///< Device handler. + const VkPhysicalDeviceMemoryProperties properties; ///< Physical device properties. std::vector<std::unique_ptr<VKMemoryAllocation>> allocations; ///< Current allocations. }; diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 4eafdc14d..c821b1229 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -883,8 +883,12 @@ void RasterizerVulkan::SetupVertexArrays(FixedPipelineState::VertexInput& vertex const GPUVAddr start{vertex_array.StartAddress()}; const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()}; - ASSERT(end > start); - const std::size_t size{end - start + 1}; + ASSERT(end >= start); + const std::size_t size{end - start}; + if (size == 0) { + buffer_bindings.AddVertexBinding(DefaultBuffer(), 0); + continue; + } const auto [buffer, offset] = buffer_cache.UploadMemory(start, size); buffer_bindings.AddVertexBinding(buffer, offset); } @@ -1039,8 +1043,7 @@ void RasterizerVulkan::SetupConstBuffer(const ConstBufferEntry& entry, const Tegra::Engines::ConstBufferInfo& buffer) { if (!buffer.enabled) { // Set values to zero to unbind buffers - update_descriptor_queue.AddBuffer(buffer_cache.GetEmptyBuffer(sizeof(float)), 0, - sizeof(float)); + update_descriptor_queue.AddBuffer(DefaultBuffer(), 0, DEFAULT_BUFFER_SIZE); return; } @@ -1063,7 +1066,9 @@ void RasterizerVulkan::SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAdd if (size == 0) { // Sometimes global memory pointers don't have a proper size. Upload a dummy entry // because Vulkan doesn't like empty buffers. - constexpr std::size_t dummy_size = 4; + // Note: Do *not* use DefaultBuffer() here, storage buffers can be written breaking the + // default buffer. + static constexpr std::size_t dummy_size = 4; const auto buffer = buffer_cache.GetEmptyBuffer(dummy_size); update_descriptor_queue.AddBuffer(buffer, 0, dummy_size); return; @@ -1228,7 +1233,7 @@ std::size_t RasterizerVulkan::CalculateVertexArraysSize() const { const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()}; DEBUG_ASSERT(end >= start); - size += (end - start + 1) * regs.vertex_array[index].enable; + size += (end - start) * regs.vertex_array[index].enable; } return size; } @@ -1276,4 +1281,29 @@ RenderPassParams RasterizerVulkan::GetRenderPassParams(Texceptions texceptions) return params; } +VkBuffer RasterizerVulkan::DefaultBuffer() { + if (default_buffer) { + return *default_buffer; + } + + VkBufferCreateInfo ci; + ci.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO; + ci.pNext = nullptr; + ci.flags = 0; + ci.size = DEFAULT_BUFFER_SIZE; + ci.usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | + VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT; + ci.sharingMode = VK_SHARING_MODE_EXCLUSIVE; + ci.queueFamilyIndexCount = 0; + ci.pQueueFamilyIndices = nullptr; + default_buffer = device.GetLogical().CreateBuffer(ci); + default_buffer_commit = memory_manager.Commit(default_buffer, false); + + scheduler.RequestOutsideRenderPassOperationContext(); + scheduler.Record([buffer = *default_buffer](vk::CommandBuffer cmdbuf) { + cmdbuf.FillBuffer(buffer, 0, DEFAULT_BUFFER_SIZE, 0); + }); + return *default_buffer; +} + } // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h index 2fa46b0cc..d41a7929e 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.h +++ b/src/video_core/renderer_vulkan/vk_rasterizer.h @@ -155,6 +155,7 @@ private: using Texceptions = std::bitset<Maxwell::NumRenderTargets + 1>; static constexpr std::size_t ZETA_TEXCEPTION_INDEX = 8; + static constexpr VkDeviceSize DEFAULT_BUFFER_SIZE = 4 * sizeof(float); void FlushWork(); @@ -247,6 +248,8 @@ private: RenderPassParams GetRenderPassParams(Texceptions texceptions) const; + VkBuffer DefaultBuffer(); + Core::System& system; Core::Frontend::EmuWindow& render_window; VKScreenInfo& screen_info; @@ -271,6 +274,9 @@ private: VKFenceManager fence_manager; VKQueryCache query_cache; + vk::Buffer default_buffer; + VKMemoryCommit default_buffer_commit; + std::array<View, Maxwell::NumRenderTargets> color_attachments; View zeta_attachment; diff --git a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp index 94d954d7a..45c180221 100644 --- a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp +++ b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp @@ -39,8 +39,7 @@ VKStagingBufferPool::StagingBuffer& VKStagingBufferPool::StagingBuffer::operator VKStagingBufferPool::VKStagingBufferPool(const VKDevice& device, VKMemoryManager& memory_manager, VKScheduler& scheduler) - : device{device}, memory_manager{memory_manager}, scheduler{scheduler}, - is_device_integrated{device.IsIntegrated()} {} + : device{device}, memory_manager{memory_manager}, scheduler{scheduler} {} VKStagingBufferPool::~VKStagingBufferPool() = default; @@ -56,9 +55,7 @@ void VKStagingBufferPool::TickFrame() { current_delete_level = (current_delete_level + 1) % NumLevels; ReleaseCache(true); - if (!is_device_integrated) { - ReleaseCache(false); - } + ReleaseCache(false); } VKBuffer* VKStagingBufferPool::TryGetReservedBuffer(std::size_t size, bool host_visible) { @@ -81,7 +78,7 @@ VKBuffer& VKStagingBufferPool::CreateStagingBuffer(std::size_t size, bool host_v ci.size = 1ULL << log2; ci.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | - VK_BUFFER_USAGE_INDEX_BUFFER_BIT; + VK_BUFFER_USAGE_INDEX_BUFFER_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT; ci.sharingMode = VK_SHARING_MODE_EXCLUSIVE; ci.queueFamilyIndexCount = 0; ci.pQueueFamilyIndices = nullptr; @@ -95,7 +92,7 @@ VKBuffer& VKStagingBufferPool::CreateStagingBuffer(std::size_t size, bool host_v } VKStagingBufferPool::StagingBuffersCache& VKStagingBufferPool::GetCache(bool host_visible) { - return is_device_integrated || host_visible ? host_staging_buffers : device_staging_buffers; + return host_visible ? host_staging_buffers : device_staging_buffers; } void VKStagingBufferPool::ReleaseCache(bool host_visible) { diff --git a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h index a0840ff8c..faf6418fd 100644 --- a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h +++ b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h @@ -71,7 +71,6 @@ private: const VKDevice& device; VKMemoryManager& memory_manager; VKScheduler& scheduler; - const bool is_device_integrated; StagingBuffersCache host_staging_buffers; StagingBuffersCache device_staging_buffers; diff --git a/src/video_core/renderer_vulkan/wrapper.cpp b/src/video_core/renderer_vulkan/wrapper.cpp index 539f3c974..7f5bc1404 100644 --- a/src/video_core/renderer_vulkan/wrapper.cpp +++ b/src/video_core/renderer_vulkan/wrapper.cpp @@ -2,6 +2,7 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. +#include <algorithm> #include <exception> #include <memory> #include <optional> @@ -16,6 +17,23 @@ namespace Vulkan::vk { namespace { +void SortPhysicalDevices(std::vector<VkPhysicalDevice>& devices, const InstanceDispatch& dld) { + std::stable_sort(devices.begin(), devices.end(), [&](auto lhs, auto rhs) { + // This will call Vulkan more than needed, but these calls are cheap. + const auto lhs_properties = vk::PhysicalDevice(lhs, dld).GetProperties(); + const auto rhs_properties = vk::PhysicalDevice(rhs, dld).GetProperties(); + + // Prefer discrete GPUs, Nvidia over AMD, AMD over Intel, Intel over the rest. + const bool preferred = + (lhs_properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU && + rhs_properties.deviceType != VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) || + (lhs_properties.vendorID == 0x10DE && rhs_properties.vendorID != 0x10DE) || + (lhs_properties.vendorID == 0x1002 && rhs_properties.vendorID != 0x1002) || + (lhs_properties.vendorID == 0x8086 && rhs_properties.vendorID != 0x8086); + return !preferred; + }); +} + template <typename T> bool Proc(T& result, const InstanceDispatch& dld, const char* proc_name, VkInstance instance = nullptr) noexcept { @@ -389,7 +407,8 @@ std::optional<std::vector<VkPhysicalDevice>> Instance::EnumeratePhysicalDevices( if (dld->vkEnumeratePhysicalDevices(handle, &num, physical_devices.data()) != VK_SUCCESS) { return std::nullopt; } - return physical_devices; + SortPhysicalDevices(physical_devices, *dld); + return std::make_optional(std::move(physical_devices)); } DebugCallback Instance::TryCreateDebugCallback( diff --git a/src/video_core/shader/decode/arithmetic_half.cpp b/src/video_core/shader/decode/arithmetic_half.cpp index ee7d9a29d..a276aee44 100644 --- a/src/video_core/shader/decode/arithmetic_half.cpp +++ b/src/video_core/shader/decode/arithmetic_half.cpp @@ -19,22 +19,46 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) { const Instruction instr = {program_code[pc]}; const auto opcode = OpCode::Decode(instr); - if (opcode->get().GetId() == OpCode::Id::HADD2_C || - opcode->get().GetId() == OpCode::Id::HADD2_R) { + bool negate_a = false; + bool negate_b = false; + bool absolute_a = false; + bool absolute_b = false; + + switch (opcode->get().GetId()) { + case OpCode::Id::HADD2_R: if (instr.alu_half.ftz == 0) { LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName()); } + negate_a = ((instr.value >> 43) & 1) != 0; + negate_b = ((instr.value >> 31) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 30) & 1) != 0; + break; + case OpCode::Id::HADD2_C: + if (instr.alu_half.ftz == 0) { + LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName()); + } + negate_a = ((instr.value >> 43) & 1) != 0; + negate_b = ((instr.value >> 56) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 54) & 1) != 0; + break; + case OpCode::Id::HMUL2_R: + negate_a = ((instr.value >> 43) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 30) & 1) != 0; + break; + case OpCode::Id::HMUL2_C: + negate_b = ((instr.value >> 31) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 54) & 1) != 0; + break; } - const bool negate_a = - opcode->get().GetId() != OpCode::Id::HMUL2_R && instr.alu_half.negate_a != 0; - const bool negate_b = - opcode->get().GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0; - Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half.type_a); - op_a = GetOperandAbsNegHalf(op_a, instr.alu_half.abs_a, negate_a); + op_a = GetOperandAbsNegHalf(op_a, absolute_a, negate_a); - auto [type_b, op_b] = [&]() -> std::tuple<HalfType, Node> { + auto [type_b, op_b] = [this, instr, opcode]() -> std::pair<HalfType, Node> { switch (opcode->get().GetId()) { case OpCode::Id::HADD2_C: case OpCode::Id::HMUL2_C: @@ -48,17 +72,16 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) { } }(); op_b = UnpackHalfFloat(op_b, type_b); - // redeclaration to avoid a bug in clang with reusing local bindings in lambdas - Node op_b_alt = GetOperandAbsNegHalf(op_b, instr.alu_half.abs_b, negate_b); + op_b = GetOperandAbsNegHalf(op_b, absolute_b, negate_b); - Node value = [&]() { + Node value = [this, opcode, op_a, op_b = op_b] { switch (opcode->get().GetId()) { case OpCode::Id::HADD2_C: case OpCode::Id::HADD2_R: - return Operation(OperationCode::HAdd, PRECISE, op_a, op_b_alt); + return Operation(OperationCode::HAdd, PRECISE, op_a, op_b); case OpCode::Id::HMUL2_C: case OpCode::Id::HMUL2_R: - return Operation(OperationCode::HMul, PRECISE, op_a, op_b_alt); + return Operation(OperationCode::HMul, PRECISE, op_a, op_b); default: UNIMPLEMENTED_MSG("Unhandled half float instruction: {}", opcode->get().GetName()); return Immediate(0); diff --git a/src/video_core/shader/decode/arithmetic_integer.cpp b/src/video_core/shader/decode/arithmetic_integer.cpp index 0f4c3103a..9af8c606d 100644 --- a/src/video_core/shader/decode/arithmetic_integer.cpp +++ b/src/video_core/shader/decode/arithmetic_integer.cpp @@ -249,8 +249,8 @@ u32 ShaderIR::DecodeArithmeticInteger(NodeBlock& bb, u32 pc) { } case OpCode::Id::LEA_IMM: { const bool neg = instr.lea.imm.neg != 0; - return {Immediate(static_cast<u32>(instr.lea.imm.entry_a)), - GetOperandAbsNegInteger(GetRegister(instr.gpr8), false, neg, true), + return {GetOperandAbsNegInteger(GetRegister(instr.gpr8), false, neg, true), + Immediate(static_cast<u32>(instr.lea.imm.entry_a)), Immediate(static_cast<u32>(instr.lea.imm.entry_b))}; } case OpCode::Id::LEA_RZ: { |