vk: Implement EXT_multi_draw support

rpcs3/Emu/RSX/NV47/FW/draw_call.cpp

@@ -138,6 +138,47 @@ namespace rsx
 		is_disjoint_primitive = is_primitive_disjointed(primitive);
 	}
 
+	simple_array<draw_range_t> draw_clause::get_subranges() const
+	{
+		ensure(!is_single_draw());
+
+		const auto range = get_range();
+		const auto limit = range.first + range.count;
+		const auto _pass_count = pass_count();
+
+		simple_array<draw_range_t> ret;
+		ret.reserve(_pass_count);
+
+		u32 previous_barrier = range.first;
+		u32 vertex_counter = 0;
+
+		for (auto it = current_barrier_it;
+			it != draw_command_barriers.end() && it->draw_id == current_range_index;
+			it++)
+		{
+			const auto& barrier = *it;
+			if (barrier.type != primitive_restart_barrier)
+				continue;
+
+			if (barrier.address <= range.first)
+				continue;
+
+			if (barrier.address >= limit)
+				break;
+
+			const u32 count = barrier.address - previous_barrier;
+			ret.push_back({ 0, vertex_counter, count });
+			previous_barrier = barrier.address;
+			vertex_counter += count;
+		}
+
+		ensure(!ret.empty());
+		ensure(previous_barrier < limit);
+		ret.push_back({ 0, vertex_counter, limit - previous_barrier });
+
+		return ret;
+	}
+
 	u32 draw_clause::execute_pipeline_dependencies(context* ctx, instanced_draw_config_t* instance_config) const
 	{
 		u32 result = 0u;

rpcs3/Emu/RSX/NV47/FW/draw_call.hpp

@@ -285,48 +285,19 @@ namespace rsx
 		 */
 		u32 execute_pipeline_dependencies(struct context* ctx, instanced_draw_config_t* instance_config = nullptr) const;
 
+		/**
+		 * Returns the first-count data for the current subdraw
+		 */
 		const draw_range_t& get_range() const
 		{
 			ensure(current_range_index < draw_command_ranges.size());
 			return draw_command_ranges[current_range_index];
 		}
 
-		simple_array<draw_range_t> get_subranges() const
-		{
-			ensure(!is_single_draw());
-
-			const auto range = get_range();
-			const auto limit = range.first + range.count;
-
-			simple_array<draw_range_t> ret;
-			u32 previous_barrier = range.first;
-			u32 vertex_counter = 0;
-
-			for (const auto& barrier : draw_command_barriers)
-			{
-				if (barrier.draw_id != current_range_index)
-					continue;
-
-				if (barrier.type != primitive_restart_barrier)
-					continue;
-
-				if (barrier.address <= range.first)
-					continue;
-
-				if (barrier.address >= limit)
-					break;
-
-				const u32 count = barrier.address - previous_barrier;
-				ret.push_back({ 0, vertex_counter, count });
-				previous_barrier = barrier.address;
-				vertex_counter += count;
-			}
-
-			ensure(!ret.empty());
-			ensure(previous_barrier < limit);
-			ret.push_back({ 0, vertex_counter, limit - previous_barrier });
-
-			return ret;
-		}
+		/*
+		 * Returns a compiled list of all subdraws.
+		 * NOTE: This is a non-trivial operation as it takes disjoint primitive boundaries into account.
+		 */
+		simple_array<draw_range_t> get_subranges() const;
 	};
 }

rpcs3/Emu/RSX/VK/VKDraw.cpp

@@ -6,6 +6,7 @@
 #include "VKGSRender.h"
 #include "vkutils/buffer_object.h"
 #include "vkutils/chip_class.h"
+#include <vulkan/vulkan_core.h>
 
 namespace vk
 {
@@ -937,6 +938,20 @@ void VKGSRender::emit_geometry(u32 sub_index)
 		{
 			vkCmdDraw(*m_current_command_buffer, upload_info.vertex_draw_count, 1, 0, 0);
 		}
+		else if (m_device->get_multidraw_support())
+		{
+			const auto subranges = draw_call.get_subranges();
+			const auto subranges_count = ::size32(subranges);
+			auto [offset, ptr] = m_draw_indirect_count_ring_info.alloc_and_map<4, VkMultiDrawInfoEXT>(subranges_count);
+
+			auto _ptr = ptr;
+			for (const auto& range : subranges)
+			{
+				_ptr->firstVertex = range.first;
+				_ptr->vertexCount = range.count;
+			}
+			vkCmdDrawMultiEXT(*m_current_command_buffer, subranges_count, ptr, 1, 0, sizeof(VkMultiDrawInfoEXT));
+		}
 		else
 		{
 			u32 vertex_offset = 0;
@@ -963,6 +978,24 @@ void VKGSRender::emit_geometry(u32 sub_index)
 		{
 			vkCmdDrawIndexed(*m_current_command_buffer, upload_info.vertex_draw_count, 1, 0, 0, 0);
 		}
+		else if (m_device->get_multidraw_support())
+		{
+			const auto subranges = draw_call.get_subranges();
+			const auto subranges_count = ::size32(subranges);
+			auto [offset, ptr] = m_draw_indirect_count_ring_info.alloc_and_map<4, VkMultiDrawIndexedInfoEXT>(subranges_count);
+
+			auto _ptr = ptr;
+			u32 vertex_offset = 0;
+			for (const auto& range : subranges)
+			{
+				const auto count = get_index_count(draw_call.primitive, range.count);
+				_ptr->vertexOffset = 0;
+				_ptr->firstIndex = vertex_offset;
+				_ptr->indexCount = count;
+				vertex_offset += count;
+			}
+			vkCmdDrawMultiIndexedEXT(*m_current_command_buffer, subranges_count, ptr, 1, 0, sizeof(VkMultiDrawIndexedInfoEXT), nullptr);
+		}
 		else
 		{
 			u32 vertex_offset = 0;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1,3 +1,4 @@
+#include "Emu/RSX/VK/VKDataHeapManager.h"
 #include "stdafx.h"
 #include "../Overlays/overlay_compile_notification.h"
 #include "../Overlays/Shaders/shader_loading_dialog_native.h"
@@ -23,6 +24,7 @@
 #include "../Program/SPIRVCommon.h"
 
 #include "util/asm.hpp"
+#include <vulkan/vulkan_core.h>
 
 namespace vk
 {
@@ -556,6 +558,12 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 		});
 	}
 
+	if (m_device->get_multidraw_support().supported)
+	{
+		m_draw_indirect_count_ring_info.create(VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT, 16 * 0x100000, "multidraw indirect buffer", 1024);
+		vk::data_heap_manager::register_ring_buffer(m_draw_indirect_count_ring_info);
+	}
+
 	// Initialize optional allocation information with placeholders
 	m_vertex_env_buffer_info = { m_vertex_env_ring_info.heap->value, 0, 16 };
 	m_vertex_constants_buffer_info = { m_transform_constants_ring_info.heap->value, 0, 16 };
@@ -1165,7 +1173,11 @@ void VKGSRender::check_heap_status(u32 flags)
 				heap_critical = m_texture_upload_buffer_ring_info.is_critical();
 				break;
 			case VK_HEAP_CHECK_VERTEX_STORAGE:
-				heap_critical = m_attrib_ring_info.is_critical() || m_index_buffer_ring_info.is_critical();
+				heap_critical = m_attrib_ring_info.is_critical() ||
+				m_index_buffer_ring_info.is_critical() ||
+				(m_draw_indirect_count_ring_info.heap
+					? m_draw_indirect_count_ring_info.is_critical()
+					: false);
 				break;
 			case VK_HEAP_CHECK_VERTEX_ENV_STORAGE:
 				heap_critical = m_vertex_env_ring_info.is_critical();

rpcs3/Emu/RSX/VK/vkutils/data_heap.h

@@ -6,6 +6,7 @@
 #include "commands.h"
 
 #include <memory>
+#include <type_traits>
 #include <vector>
 
 namespace vk
@@ -38,6 +39,15 @@ namespace vk
 		void* map(usz offset, usz size);
 		void unmap(bool force = false);
 
+		template<int Alignment, typename T = char>
+			requires std::is_trivially_destructible_v<T>
+		std::pair<usz, T*> alloc_and_map(usz count)
+		{
+			const auto size_bytes = count * sizeof(T);
+			const auto addr = alloc<Alignment>(size_bytes);
+			return { addr, reinterpret_cast<T*>(	map(addr, size_bytes)) };
+		}
+
 		void sync(const vk::command_buffer& cmd);
 
 		// Properties

rpcs3/Emu/RSX/VK/vkutils/device.cpp

@@ -2,6 +2,7 @@
 #include "instance.h"
 #include "util/logs.hpp"
 #include "Emu/system_config.h"
+#include <vulkan/vulkan_core.h>
 
 namespace vk
 {
@@ -36,6 +37,7 @@ namespace vk
 			VkPhysicalDeviceCustomBorderColorFeaturesEXT custom_border_color_info{};
 			VkPhysicalDeviceBorderColorSwizzleFeaturesEXT border_color_swizzle_info{};
 			VkPhysicalDeviceFaultFeaturesEXT device_fault_info{};
+			VkPhysicalDeviceMultiDrawFeaturesEXT multidraw_info{};
 
 			if (device_extensions.is_supported(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME))
 			{
@@ -86,6 +88,13 @@ namespace vk
 				features2.pNext         = &device_fault_info;
 			}
 
+			if (device_extensions.is_supported(VK_EXT_MULTI_DRAW_EXTENSION_NAME))
+			{
+				multidraw_info.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_FEATURES_EXT;
+				multidraw_info.pNext = features2.pNext;
+				features2.pNext      = &multidraw_info;
+			}
+
 			auto _vkGetPhysicalDeviceFeatures2KHR = reinterpret_cast<PFN_vkGetPhysicalDeviceFeatures2KHR>(vkGetInstanceProcAddr(parent, "vkGetPhysicalDeviceFeatures2KHR"));
 			ensure(_vkGetPhysicalDeviceFeatures2KHR); // "vkGetInstanceProcAddress failed to find entry point!"
 			_vkGetPhysicalDeviceFeatures2KHR(dev, &features2);
@@ -98,6 +107,9 @@ namespace vk
 			custom_border_color_support.swizzle_extension_supported = border_color_swizzle_info.sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BORDER_COLOR_SWIZZLE_FEATURES_EXT;
 			custom_border_color_support.require_border_color_remap = !border_color_swizzle_info.borderColorSwizzleFromImage;
 
+			multidraw_support.supported = !!multidraw_info.multiDraw;
+			multidraw_support.max_batch_size = 65536;
+
 			optional_features_support.barycentric_coords  = !!shader_barycentric_info.fragmentShaderBarycentric;
 			optional_features_support.framebuffer_loops   = !!fbo_loops_info.attachmentFeedbackLoopLayout;
 			optional_features_support.extended_device_fault = !!device_fault_info.deviceFault;
@@ -124,7 +136,6 @@ namespace vk
 		optional_features_support.sampler_mirror_clamped   = device_extensions.is_supported(VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME);
 		optional_features_support.synchronization_2        = device_extensions.is_supported(VK_KHR_SYNCHRONIZATION_2_EXTENSION_NAME);
 		optional_features_support.unrestricted_depth_range = device_extensions.is_supported(VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME);
-		optional_features_support.multidraw_indirect       = device_extensions.is_supported(VK_KHR_DRAW_INDIRECT_COUNT_EXTENSION_NAME);
 
 		optional_features_support.debug_utils              = instance_extensions.is_supported(VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
 		optional_features_support.surface_capabilities_2   = instance_extensions.is_supported(VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME);
@@ -165,6 +176,7 @@ namespace vk
 			properties2.pNext = nullptr;
 
 			VkPhysicalDeviceDescriptorIndexingPropertiesEXT descriptor_indexing_props{};
+			VkPhysicalDeviceMultiDrawPropertiesEXT multidraw_props{};
 
 			if (descriptor_indexing_support)
 			{
@@ -173,6 +185,13 @@ namespace vk
 				properties2.pNext = &descriptor_indexing_props;
 			}
 
+			if (multidraw_support.supported)
+			{
+				multidraw_props.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_PROPERTIES_EXT;
+				multidraw_props.pNext = properties2.pNext;
+				properties2.pNext = &multidraw_props;
+			}
+
 			if (device_extensions.is_supported(VK_KHR_DRIVER_PROPERTIES_EXTENSION_NAME))
 			{
 				driver_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR;
@@ -199,6 +218,17 @@ namespace vk
 					descriptor_max_draw_calls = 8192;
 				}
 			}
+
+			if (multidraw_support.supported)
+			{
+				multidraw_support.max_batch_size = multidraw_props.maxMultiDrawCount;
+
+				if (!multidraw_props.maxMultiDrawCount)
+				{
+					rsx_log.error("Physical device reports 0 support maxMultiDraw count. Multidraw support will be disabled.");
+					multidraw_support.supported = false;
+				}
+			}
 		}
 	}
 

rpcs3/Emu/RSX/VK/vkutils/device.h

@@ -61,6 +61,14 @@ namespace vk
 		operator bool() const { return supported; }
 	};
 
+	struct multidraw_features
+	{
+		bool supported;
+		u32 max_batch_size;
+
+		operator bool() const { return supported; }
+	};
+
 	class physical_device
 	{
 		VkInstance parent = VK_NULL_HANDLE;
@@ -79,6 +87,8 @@ namespace vk
 
 		custom_border_color_features custom_border_color_support{};
 
+		multidraw_features multidraw_support{};
+
 		struct
 		{
 			bool barycentric_coords = false;
@@ -93,7 +103,6 @@ namespace vk
 			bool unrestricted_depth_range = false;
 			bool extended_device_fault = false;
 			bool texture_compression_bc = false;
-			bool multidraw_indirect = false;
 		} optional_features_support;
 
 		friend class render_device;
@@ -176,6 +185,7 @@ namespace vk
 		const pipeline_binding_table& get_pipeline_binding_table() const { return m_pipeline_binding_table; }
 		const gpu_shader_types_support& get_shader_types_support() const { return pgpu->shader_types_support; }
 		const custom_border_color_features& get_custom_border_color_support() const { return pgpu->custom_border_color_support; }
+		const multidraw_features get_multidraw_support() const { return pgpu->multidraw_support; }
 
 		bool get_shader_stencil_export_support() const { return pgpu->optional_features_support.shader_stencil_export; }
 		bool get_depth_bounds_support() const { return pgpu->features.depthBounds != VK_FALSE; }
@@ -193,7 +203,6 @@ namespace vk
 		bool get_synchronization2_support() const { return pgpu->optional_features_support.synchronization_2; }
 		bool get_extended_device_fault_support() const { return pgpu->optional_features_support.extended_device_fault; }
 		bool get_texture_compression_bc_support() const { return pgpu->optional_features_support.texture_compression_bc; }
-		bool get_multidraw_indirect_support() const { return pgpu->optional_features_support.multidraw_indirect; }
 
 		u64 get_descriptor_update_after_bind_support() const { return pgpu->descriptor_indexing_support.update_after_bind_mask; }
 		u32 get_descriptor_max_draw_calls() const { return pgpu->descriptor_max_draw_calls; }