vk: Per work-queue scratch resources

- Avoids parallel tasks from trampling over each other's data

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1994,7 +1994,7 @@ void VKGSRender::load_program_env()
 
 	if (vk::emulate_conditional_rendering())
 	{
-		auto predicate = m_cond_render_buffer ? m_cond_render_buffer->value : vk::get_scratch_buffer(4)->value;
+		auto predicate = m_cond_render_buffer ? m_cond_render_buffer->value : vk::get_scratch_buffer(*m_current_command_buffer, 4)->value;
 		m_program->bind_buffer({ predicate, 0, 4 }, binding_table.conditional_render_predicate_slot, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, m_current_frame->descriptor_set);
 	}
 
@@ -2640,7 +2640,7 @@ void VKGSRender::begin_conditional_rendering(const std::vector<rsx::reports::occ
 		}
 	}
 
-	auto scratch = vk::get_scratch_buffer(OCCLUSION_MAX_POOL_SIZE * 4);
+	auto scratch = vk::get_scratch_buffer(*m_current_command_buffer, OCCLUSION_MAX_POOL_SIZE * 4);
 	u32 dst_offset = 0;
 	usz first = 0;
 	usz last;

rpcs3/Emu/RSX/VK/VKTexture.cpp

@@ -347,7 +347,7 @@ namespace vk
 		const auto min_scratch_size = calculate_working_buffer_size(src_length, src->aspect() | dst->aspect());
 
 		// Initialize scratch memory
-		auto scratch_buf = vk::get_scratch_buffer(min_scratch_size);
+		auto scratch_buf = vk::get_scratch_buffer(cmd, min_scratch_size);
 
 		for (u32 mip_level = 0; mip_level < mipmaps; ++mip_level)
 		{
@@ -552,7 +552,7 @@ namespace vk
 					const auto dst_w = dst_rect.width();
 					const auto dst_h = dst_rect.height();
 
-					auto scratch_buf = vk::get_scratch_buffer(std::max(src_w, dst_w) * std::max(src_h, dst_h) * 4);
+					auto scratch_buf = vk::get_scratch_buffer(cmd, std::max(src_w, dst_w) * std::max(src_h, dst_h) * 4);
 
 					//1. Copy unscaled to typeless surface
 					VkBufferImageCopy info{};
@@ -1014,7 +1014,13 @@ namespace vk
 						scratch_buf_size += dst_image->width() * dst_image->height() * 5;
 					}
 
-					scratch_buf = vk::get_scratch_buffer(scratch_buf_size);
+					// Must acquire scratch buffer owned by the processing command queue!
+					auto pdev = vk::get_current_renderer();
+					const u32 queue_family = (image_setup_flags & vk::upload_contents_async) ?
+						pdev->get_transfer_queue_family() :
+						pdev->get_graphics_queue_family();
+
+					scratch_buf = vk::get_scratch_buffer(queue_family, scratch_buf_size);
 					buffer_copies.reserve(subresource_layout.size());
 				}
 

rpcs3/Emu/RSX/VK/VKTextureCache.cpp

@@ -86,7 +86,7 @@ namespace vk
 			const auto task_length = transfer_pitch * src_area.height();
 			const auto working_buffer_length = calculate_working_buffer_size(task_length, src->aspect());
 
-			auto working_buffer = vk::get_scratch_buffer(working_buffer_length);
+			auto working_buffer = vk::get_scratch_buffer(cmd, working_buffer_length);
 			auto final_mapping = vk::map_dma(valid_range.start, section_length);
 
 			VkBufferImageCopy region = {};
@@ -358,7 +358,7 @@ namespace vk
 					copy.imageSubresource = { src_image->aspect(), 0, 0, 1 };
 
 					const auto mem_length = src_w * src_h * dst_bpp;
-					auto scratch_buf = vk::get_scratch_buffer(mem_length);
+					auto scratch_buf = vk::get_scratch_buffer(cmd, mem_length);
 					vkCmdCopyImageToBuffer(cmd, src_image->value, src_image->current_layout, scratch_buf->value, 1, &copy);
 
 					vk::insert_buffer_memory_barrier(cmd, scratch_buf->value, 0, mem_length, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,

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

@@ -55,7 +55,7 @@ namespace vk
 		return g_null_sampler;
 	}
 
-	vk::image_view* null_image_view(vk::command_buffer& cmd, VkImageViewType type)
+	vk::image_view* null_image_view(const vk::command_buffer& cmd, VkImageViewType type)
 	{
 		if (auto found = g_null_image_views.find(type);
 			found != g_null_image_views.end())
@@ -151,9 +151,9 @@ namespace vk
 		return ptr.get();
 	}
 
-	vk::buffer* get_scratch_buffer(u64 min_required_size)
+	vk::buffer* get_scratch_buffer(u32 queue_family, u64 min_required_size)
 	{
-		auto& scratch_buffer = g_scratch_buffers_pool[0 /*TODO: Replace with Queue Family ID*/].get_buf();
+		auto& scratch_buffer = g_scratch_buffers_pool[queue_family].get_buf();
 
 		if (scratch_buffer && scratch_buffer->size() < min_required_size)
 		{
@@ -174,6 +174,11 @@ namespace vk
 		return scratch_buffer.get();
 	}
 
+	vk::buffer* get_scratch_buffer(const vk::command_buffer& cmd, u64 min_required_size)
+	{
+		return get_scratch_buffer(cmd.get_queue_family(), min_required_size);
+	}
+
 	void clear_scratch_resources()
 	{
 		g_null_image_views.clear();

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

@@ -4,9 +4,11 @@
 namespace vk
 {
 	VkSampler null_sampler();
-	image_view* null_image_view(command_buffer&, VkImageViewType type);
+	image_view* null_image_view(const command_buffer& cmd, VkImageViewType type);
 	image* get_typeless_helper(VkFormat format, rsx::format_class format_class, u32 requested_width, u32 requested_height);
-	buffer* get_scratch_buffer(u64 min_required_size);
+
+	buffer* get_scratch_buffer(u32 queue_family, u64 min_required_size);
+	buffer* get_scratch_buffer(const command_buffer& cmd, u64 min_required_size);
 
 	void clear_scratch_resources();
 }