gpu: store large compute shared memory in device buffers

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -68,7 +68,8 @@ vk::DescriptorType NativeDescriptorType(BindingKind kind) {
 		case BindingKind::BdaPagetable:
 		case BindingKind::FaultBuffer:
 		case BindingKind::FlattenedSrt:
-		case BindingKind::ShaderData: return vk::DescriptorType::eStorageBuffer;
+		case BindingKind::ShaderData:
+		case BindingKind::SharedMemory: return vk::DescriptorType::eStorageBuffer;
 		case BindingKind::Count: EXIT("invalid native descriptor binding kind");
 	}
 	EXIT("invalid native descriptor binding kind");
@@ -779,6 +780,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
 	prepared.flattened_srt = {};
 	prepared.shader_data_buffer = {};
+	prepared.shared_memory = {};
 	prepared.images.resize(program.info.images.size());
 	prepared.samplers.clear();
 	prepared.shader_data.clear();
@@ -1120,12 +1122,15 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 					}
 					case BindingKind::FlattenedSrt:
 					case BindingKind::ShaderData:
+					case BindingKind::SharedMemory:
 					case BindingKind::Gds: {
 						const vk::DescriptorBufferInfo* view = &descriptors.gds;
 						if (binding.kind == BindingKind::FlattenedSrt) {
 							view = &descriptors.flattened_srt;
 						} else if (binding.kind == BindingKind::ShaderData) {
 							view = &descriptors.shader_data_buffer;
+						} else if (binding.kind == BindingKind::SharedMemory) {
+							view = &descriptors.shared_memory;
 						}
 						EXIT_IF(view->buffer == nullptr);
 						m_descriptor_buffers.push_back(*view);

src/graphics/host_gpu/renderer/pipeline/descriptors.h

@@ -44,6 +44,7 @@ struct PreparedBindings {
 	vk::DescriptorBufferInfo              gds {nullptr, 0, VK_WHOLE_SIZE};
 	vk::DescriptorBufferInfo              flattened_srt;
 	vk::DescriptorBufferInfo              shader_data_buffer;
+	vk::DescriptorBufferInfo              shared_memory;
 	std::vector<uint32_t>                 shader_data;
 };
 

src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp

@@ -680,19 +680,8 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
                                                ShaderComputeInputInfo&      input_info) {
 	input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
 	const auto        params      = PrepareProgram(regs, sh, input_info);
-	// Use one effective size for the cache key, LDS declaration, and access bounds.
-	const auto max_lds_dwords =
-	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize / 4u;
-	if (input_info.lds_size_dwords > max_lds_dwords) {
-		static std::atomic_bool warned = false;
-		if (!warned.exchange(true, std::memory_order_relaxed)) {
-			PipelineCacheLog("GPU warning: game compute shader requests {} bytes of LDS, but "
-			                 "the Vulkan device limit is {} bytes. Clamping LDS; rendering may "
-			                 "be incorrect.",
-			                 input_info.lds_size_dwords * 4u, max_lds_dwords * 4u);
-		}
-	}
-	input_info.lds_size_dwords = std::min(input_info.lds_size_dwords, max_lds_dwords);
+	input_info.lds_storage = input_info.lds_size_dwords * 4u >
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize;
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -20,6 +20,7 @@
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 #include "graphics/shader/shader.h"
 #include "kernel/eventQueue.h"
 #include "kernel/pthread.h"
@@ -195,6 +196,35 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	return true;
 }
 
+static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
+                             PreparedBindings& bindings, uint64_t indirect_args = 0) {
+	if (ShaderRecompiler::IR::FindBinding(input.stage.program->bindings,
+	        ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory) == nullptr) {
+		return;
+	}
+	auto& cache = context.GetBufferCache();
+	if (indirect_args != 0) {
+		cache.ReadMemory(indirect_args, sizeof(vk::DispatchIndirectCommand));
+		std::memcpy(input.workgroup_counts, reinterpret_cast<const void*>(indirect_args),
+		            sizeof(input.workgroup_counts));
+	}
+	// LDS has no contents to preserve between dispatches. The existing shader hazard
+	// barriers also order other users of this GPU-only utility buffer.
+	auto& storage = cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
+	const auto limit = std::min<uint64_t>(storage.Size(),
+	    context.GetGraphics().GetPhysicalDeviceProperties().limits.maxStorageBufferRange);
+	uint64_t size = sizeof(uint32_t);
+	if (std::ranges::find(input.workgroup_counts, 0u) == std::end(input.workgroup_counts)) {
+		size = uint64_t {input.lds_size_dwords} * sizeof(uint32_t);
+		EXIT_IF(size == 0 || size > limit);
+		for (const auto count: input.workgroup_counts) {
+			EXIT_IF(size > limit / count);
+			size *= count;
+		}
+	}
+	bindings.shared_memory = {storage.Handle(), 0, size};
+}
+
 void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_x, uint32_t thread_group_y,
                                     uint32_t thread_group_z, uint32_t mode) {
@@ -317,13 +347,15 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		m_context.PrepareBda();
 	}
 	RebindImages(bindings);
+	BindSharedMemory(m_context, input_info, bindings);
 	RebindBuffers(bindings);
 
 	auto              vk_buffer        = buffer.Handle();
 	PreparedBindings* descriptor_stage = &bindings;
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
-	bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
+	bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage);
 	has_storage_writes =
 	    std::any_of(program.info.images.begin(), program.info.images.end(),
 	                [](const auto& image) {
@@ -430,6 +462,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
+	BindSharedMemory(m_context, input_info, bindings, args_addr);
 	RebindImages(bindings);
 	// Acquiring arguments can merge cache buffers; finalize shader bindings afterward.
 	const auto [args_buffer, args_offset] = m_context.GetBufferCache().ObtainBuffer(
@@ -440,7 +473,8 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
 	const auto vk_buffer = buffer.Handle();
-	const bool has_storage_writes = HasShaderBufferWrites(input_info.stage) ||
+	const bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage) ||
 	    std::any_of(program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 		    return image.written && image.resource_class ==
 		                                ShaderRecompiler::IR::ImageResourceClass::Storage;