vk: Update binding model for compute jobs

rpcs3/Emu/RSX/Common/simple_array.hpp

@@ -305,6 +305,11 @@ namespace rsx
 			return _size * sizeof(Ty);
 		}
 
+		u32 size_bytes32() const
+		{
+			return _size * sizeof(Ty);
+		}
+
 		u32 capacity() const
 		{
 			return _capacity;

rpcs3/Emu/RSX/VK/VKCompute.cpp

@@ -132,7 +132,7 @@ namespace vk
 			m_program = compiler->compile(create_info, vk::pipe_compiler::COMPILE_INLINE, {}, get_inputs());
 		}
 
-		bind_resources();
+		bind_resources(cmd);
 		m_program->bind(cmd, VK_PIPELINE_BIND_POINT_COMPUTE);
 	}
 
@@ -243,20 +243,19 @@ namespace vk
 		m_src += suffix;
 	}
 
-	void cs_shuffle_base::bind_resources()
+	void cs_shuffle_base::bind_resources(const vk::command_buffer& cmd)
 	{
+		set_parameters(cmd);
 		m_program->bind_uniform({ m_data->value, m_data_offset, m_data_length }, 0, 0);
 	}
 
-	void cs_shuffle_base::set_parameters(const vk::command_buffer& cmd, const u32* params, u8 count)
+	void cs_shuffle_base::set_parameters(const vk::command_buffer& cmd)
 	{
-		if (!m_program)
+		if (!m_params.empty())
 		{
-			load_program(cmd);
+			ensure(use_push_constants);
+			vkCmdPushConstants(cmd, m_program->layout(), VK_SHADER_STAGE_COMPUTE_BIT, 0, m_params.size_bytes32(), m_params.data());
 		}
-
-		ensure(use_push_constants);
-		vkCmdPushConstants(cmd, m_program->layout(), VK_SHADER_STAGE_COMPUTE_BIT, 0, count * 4, params);
 	}
 
 	void cs_shuffle_base::run(const vk::command_buffer& cmd, const vk::buffer* data, u32 data_length, u32 data_offset)
@@ -294,15 +293,15 @@ namespace vk
 			"	uint stencil_offset;\n";
 	}
 
-	void cs_interleave_task::bind_resources()
+	void cs_interleave_task::bind_resources(const vk::command_buffer& cmd)
 	{
+		set_parameters(cmd);
 		m_program->bind_uniform({ m_data->value, m_data_offset, m_ssbo_length }, 0, 0);
 	}
 
 	void cs_interleave_task::run(const vk::command_buffer& cmd, const vk::buffer* data, u32 data_offset, u32 data_length, u32 zeta_offset, u32 stencil_offset)
 	{
-		u32 parameters[4] = { data_length, zeta_offset - data_offset, stencil_offset - data_offset, 0 };
+		m_params = { data_length, zeta_offset - data_offset, stencil_offset - data_offset, 0 };
-		set_parameters(cmd, parameters, 4);
 
 		ensure(stencil_offset > data_offset);
 		m_ssbo_length = stencil_offset + (data_length / 4) - data_offset;
@@ -354,7 +353,7 @@ namespace vk
 		m_src = fmt::replace_all(m_src, syntax_replace);
 	}
 
-	void cs_aggregator::bind_resources()
+	void cs_aggregator::bind_resources(const vk::command_buffer& cmd)
 	{
 		m_program->bind_uniform({ src->value, 0, block_length }, 0, 0);
 		m_program->bind_uniform({ dst->value, 0, 4 }, 0, 1);

rpcs3/Emu/RSX/VK/VKCompute.h

@@ -35,7 +35,7 @@ namespace vk
 		void destroy();
 
 		virtual std::vector<glsl::program_input> get_inputs();
-		virtual void bind_resources() {}
+		virtual void bind_resources(const vk::command_buffer& cmd) {}
 
 		void load_program(const vk::command_buffer& cmd);
 
@@ -50,6 +50,8 @@ namespace vk
 		u32 m_data_length = 0;
 		u32 kernel_size = 1;
 
+		rsx::simple_array<u32> m_params;
+
 		std::string variables, work_kernel, loop_advance, suffix;
 		std::string method_declarations;
 
@@ -57,9 +59,9 @@ namespace vk
 
 		void build(const char* function_name, u32 _kernel_size = 0);
 
-		void bind_resources() override;
+		void bind_resources(const vk::command_buffer& cmd) override;
 
-		void set_parameters(const vk::command_buffer& cmd, const u32* params, u8 count);
+		void set_parameters(const vk::command_buffer& cmd);
 
 		void run(const vk::command_buffer& cmd, const vk::buffer* data, u32 data_length, u32 data_offset = 0);
 	};
@@ -125,7 +127,7 @@ namespace vk
 
 		cs_interleave_task();
 
-		void bind_resources() override;
+		void bind_resources(const vk::command_buffer& cmd) override;
 
 		void run(const vk::command_buffer& cmd, const vk::buffer* data, u32 data_offset, u32 data_length, u32 zeta_offset, u32 stencil_offset);
 	};
@@ -342,8 +344,9 @@ namespace vk
 			cs_shuffle_base::build("");
 		}
 
-		void bind_resources() override
+		void bind_resources(const vk::command_buffer& cmd) override
 		{
+			set_parameters(cmd);
 			m_program->bind_uniform({ m_data->value, m_data_offset, m_ssbo_length }, 0, 0);
 		}
 
@@ -361,8 +364,7 @@ namespace vk
 				data_offset = src_offset;
 			}
 
-			u32 parameters[4] = { src_length, src_offset - data_offset, dst_offset - data_offset, 0 };
+			m_params = { src_length, src_offset - data_offset, dst_offset - data_offset, 0 };
-			set_parameters(cmd, parameters, 4);
 			cs_shuffle_base::run(cmd, data, src_length, data_offset);
 		}
 	};
@@ -443,19 +445,16 @@ namespace vk
 			m_src = fmt::replace_all(m_src, syntax_replace);
 		}
 
-		void bind_resources() override
+		void bind_resources(const vk::command_buffer& cmd) override
 		{
+			set_parameters(cmd);
+
 			m_program->bind_uniform({ src_buffer->value, in_offset, block_length }, 0, 0);
 			m_program->bind_uniform({ dst_buffer->value, out_offset, block_length }, 0, 1);
 		}
 
 		void set_parameters(const vk::command_buffer& cmd)
 		{
-			if (!m_program)
-			{
-				load_program(cmd);
-			}
-
 			vkCmdPushConstants(cmd, m_program->layout(), VK_SHADER_STAGE_COMPUTE_BIT, 0, push_constants_size, params.data);
 		}
 
@@ -475,7 +474,6 @@ namespace vk
 			params.logw = rsx::ceil_log2(width);
 			params.logh = rsx::ceil_log2(height);
 			params.logd = rsx::ceil_log2(depth);
-			set_parameters(cmd);
 
 			const u32 num_bytes_per_invocation = (sizeof(_BlockType) * optimal_group_size);
 			const u32 linear_invocations = utils::aligned_div(data_length, num_bytes_per_invocation);
@@ -492,7 +490,7 @@ namespace vk
 
 		cs_aggregator();
 
-		void bind_resources() override;
+		void bind_resources(const vk::command_buffer& cmd) override;
 
 		void run(const vk::command_buffer& cmd, const vk::buffer* dst, const vk::buffer* src, u32 num_words);
 	};
@@ -576,8 +574,10 @@ namespace vk
 			m_src = fmt::replace_all(m_src, syntax_replace);
 		}
 
-		void bind_resources() override
+		void bind_resources(const vk::command_buffer& cmd) override
 		{
+			set_parameters(cmd);
+
 			const auto op = static_cast<u32>(Op);
 			m_program->bind_uniform({ src_buffer->value, in_offset, in_block_length }, 0u, 0u ^ op);
 			m_program->bind_uniform({ dst_buffer->value, out_offset, out_block_length }, 0u, 1u ^ op);
@@ -648,7 +648,6 @@ namespace vk
 			params.image_height = (Op == RSX_detiler_op::decode) ? tile_aligned_height : config.image_height;
 			params.image_pitch = config.image_pitch;
 			params.image_bpp = config.image_bpp;
-			set_parameters(cmd);
 
 			const u32 subtexels_per_invocation = (config.image_bpp < 4) ? (4 / config.image_bpp) : 1;
 			const u32 virtual_width = config.image_width / subtexels_per_invocation;

rpcs3/Emu/RSX/VK/VKResolveHelper.h

@@ -49,7 +49,7 @@ namespace vk
 			return result;
 		}
 
-		void bind_resources() override
+		void bind_resources(const vk::command_buffer& cmd) override
 		{
 			auto msaa_view = multisampled->get_view(rsx::default_remap_vector.with_encoding(VK_REMAP_VIEW_MULTISAMPLED));
 			auto resolved_view = resolve->get_view(rsx::default_remap_vector.with_encoding(VK_REMAP_IDENTITY));

rpcs3/Emu/RSX/VK/upscalers/fsr1/fsr_pass.cpp

@@ -97,7 +97,7 @@ namespace vk
 			return result;
 		}
 
-		void fsr_pass::bind_resources()
+		void fsr_pass::bind_resources(const vk::command_buffer& cmd)
 		{
 			// Bind relevant stuff
 			if (!m_sampler)

rpcs3/Emu/RSX/VK/upscalers/fsr_pass.h

@@ -20,7 +20,7 @@ namespace vk
 			u32 m_constants_buf[20];
 
 			std::vector<glsl::program_input> get_inputs() override;
-			void bind_resources() override;
+			void bind_resources(const vk::command_buffer&) override;
 
 			virtual void configure(const vk::command_buffer& cmd) = 0;
 

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

@@ -14,7 +14,7 @@ namespace vk
 		public:
 			inline void flush_all()
 			{
-				std::shared_lock lock(m_notifications_lock);
+				std::lock_guard lock(m_notifications_lock);
 
 				for (auto& set : m_notification_list)
 				{
@@ -44,7 +44,7 @@ namespace vk
 
 		private:
 			rsx::simple_array<descriptor_set*> m_notification_list;
-			std::shared_mutex m_notifications_lock;
+			std::mutex m_notifications_lock;
 
 			dispatch_manager(const dispatch_manager&) = delete;
 			dispatch_manager& operator = (const dispatch_manager&) = delete;
@@ -81,17 +81,17 @@ namespace vk
 				}
 				else
 				{
-					binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT;
+					binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT;
 				}
 			}
 
-			binding_infos.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT;
+			binding_infos.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO;
 			binding_infos.pNext = nullptr;
 			binding_infos.bindingCount = ::size32(binding_flags);
 			binding_infos.pBindingFlags = binding_flags.data();
 
 			infos.pNext = &binding_infos;
-			infos.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT;
+			infos.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT;
 
 			VkDescriptorSetLayout result;
 			CHECK_RESULT(vkCreateDescriptorSetLayout(*g_render_device, &infos, nullptr, &result));
@@ -442,6 +442,7 @@ namespace vk
 	{
 		if (!m_push_type_mask)
 		{
+			ensure(m_pending_writes.empty());
 			return;
 		}
 
@@ -450,16 +451,17 @@ namespace vk
 			(m_pending_writes.size() >= max_cache_size))
 		{
 			flush();
+			return;
 		}
-		else if (m_update_after_bind_mask)
+
-		{
+		// Register for async flush
-			// Register for async flush
+		ensure(m_update_after_bind_mask);
-			g_fxo->get<descriptors::dispatch_manager>().register_(this);
+		g_fxo->get<descriptors::dispatch_manager>().register_(this);
-		}
 	}
 
 	void descriptor_set::bind(const vk::command_buffer& cmd, VkPipelineBindPoint bind_point, VkPipelineLayout layout)
 	{
+		// Notify
 		on_bind();
 
 		vkCmdBindDescriptorSets(cmd, bind_point, layout, 0, 1, &m_handle, ::size32(m_dynamic_offsets), m_dynamic_offsets.data());