vulkan: Fix blit engine transfer to ARGB8 render target memory

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -33,6 +33,8 @@ namespace rsx
 		u32 dst_gcm_format = 0;
 		f32 src_scaling_hint = 1.f;
 		f32 dst_scaling_hint = 1.f;
+		texture_upload_context src_context = texture_upload_context::blit_engine_src;
+		texture_upload_context dst_context = texture_upload_context::blit_engine_dst;
 
 		void analyse()
 		{
@@ -2125,6 +2127,7 @@ namespace rsx
 				if (cached_dest)
 				{
 					dest_texture = cached_dest->get_raw_texture();
+					typeless_info.dst_context = cached_dest->get_context();
 
 					max_dst_width = cached_dest->get_width();
 					max_dst_height = cached_dest->get_height();
@@ -2144,6 +2147,7 @@ namespace rsx
 				dst_area.y2 += dst_subres.y;
 
 				dest_texture = dst_subres.surface->get_surface();
+				typeless_info.dst_context = texture_upload_context::framebuffer_storage;
 
 				max_dst_width = (u16)(dst_subres.surface->get_surface_width() * typeless_info.dst_scaling_hint);
 				max_dst_height = dst_subres.surface->get_surface_height();
@@ -2179,6 +2183,7 @@ namespace rsx
 						src_area.y2 <= surface->get_height())
 					{
 						vram_texture = surface->get_raw_texture();
+						typeless_info.src_context = surface->get_context();
 						break;
 					}
 
@@ -2206,6 +2211,7 @@ namespace rsx
 						subresource_layout, rsx::texture_dimension_extended::texture_dimension_2d, dst.swizzled)->get_raw_texture();
 
 					m_texture_memory_in_use += src.pitch * src.slice_h;
+					typeless_info.src_context = texture_upload_context::blit_engine_src;
 				}
 			}
 			else
@@ -2232,6 +2238,7 @@ namespace rsx
 				src_area.y2 += src_subres.y;
 
 				vram_texture = src_subres.surface->get_surface();
+				typeless_info.src_context = texture_upload_context::framebuffer_storage;
 			}
 
 			const bool src_is_depth = src_subres.is_depth_surface;
@@ -2342,6 +2349,7 @@ namespace rsx
 					channel_order);
 
 				dest_texture = cached_dest->get_raw_texture();
+				typeless_info.dst_context = texture_upload_context::blit_engine_dst;
 				m_texture_memory_in_use += dst.pitch * dst_dimensions.height;
 			}
 

rpcs3/Emu/RSX/VK/VKHelpers.cpp

@@ -195,7 +195,7 @@ namespace vk
 		if (!g_scratch_buffer)
 		{
 			// 32M disposable scratch memory
-			g_scratch_buffer = std::make_unique<vk::buffer>(*g_current_renderer, 32 * 0x100000,
+			g_scratch_buffer = std::make_unique<vk::buffer>(*g_current_renderer, 64 * 0x100000,
 				g_current_renderer->get_memory_mapping().device_local, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
 				VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 0);
 		}

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -1129,6 +1129,47 @@ namespace vk
 						vk::copy_image_typeless(*commands, dst, real_dst, { 0, 0, (s32)dst->width(), (s32)dst->height() }, { 0, 0, (s32)internal_width, (s32)dst->height() }, 1,
 							vk::get_aspect_flags(dst->info.format), vk::get_aspect_flags(format));
 					}
+					else if (xfer_info.dst_context == rsx::texture_upload_context::framebuffer_storage)
+					{
+						if (xfer_info.src_context != rsx::texture_upload_context::blit_engine_dst &&
+							xfer_info.src_context != rsx::texture_upload_context::framebuffer_storage)
+						{
+							// Data moving to rendertarget, where byte ordering has to be preserved
+							// NOTE: This is a workaround, true accuracy would require all RTT<->cache transfers to invoke this step but thats too slow
+							// Sampling is ok; image view swizzle will work around it
+							if (dst->info.format == VK_FORMAT_B8G8R8A8_UNORM)
+							{
+								// For this specific format, channel ordering is faked via custom remap, undo this before transfer
+								VkBufferImageCopy copy{};
+								copy.imageExtent = src->info.extent;
+								copy.imageOffset = { 0, 0, 0 };
+								copy.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 };
+
+								const auto scratch_buf = vk::get_scratch_buffer();
+								const auto data_length = src->info.extent.width * src->info.extent.height * 4;
+
+								const auto current_layout = src->current_layout;
+								vk::change_image_layout(*commands, real_src, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 });
+								vkCmdCopyImageToBuffer(*commands, src->value, src->current_layout, scratch_buf->value, 1, &copy);
+								vk::change_image_layout(*commands, real_src, current_layout, { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 });
+
+								vk::insert_buffer_memory_barrier(*commands, scratch_buf->value, 0, data_length,
+									VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+									VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT);
+
+								vk::get_compute_task<vk::cs_shuffle_32>()->run(*commands, scratch_buf, data_length);
+
+								vk::insert_buffer_memory_barrier(*commands, scratch_buf->value, 0, data_length,
+									VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+									VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_TRANSFER_READ_BIT);
+
+								real_src = vk::get_typeless_helper(src->info.format);
+								vk::change_image_layout(*commands, real_src, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1});
+
+								vkCmdCopyBufferToImage(*commands, scratch_buf->value, real_src->value, real_src->current_layout, 1, &copy);
+							}
+						}
+					}
 
 					// Checks
 					if (src_area.x2 <= src_area.x1 || src_area.y2 <= src_area.y1 || dst_area.x2 <= dst_area.x1 || dst_area.y2 <= dst_area.y1)