From f762285b9c4896b0ee63d13488659c6bc0ffa782 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Tue, 23 May 2017 10:26:49 +0200
Subject: [PATCH] Vulkan: Some barrier optimization

---
 ext/native/thin3d/thin3d_vulkan.cpp | 48 ++++++++++++++++++-----------
 1 file changed, 30 insertions(+), 18 deletions(-)

diff --git a/ext/native/thin3d/thin3d_vulkan.cpp b/ext/native/thin3d/thin3d_vulkan.cpp
index cf8db31a78..cb9459bc2b 100644
--- a/ext/native/thin3d/thin3d_vulkan.cpp
+++ b/ext/native/thin3d/thin3d_vulkan.cpp
@@ -1666,23 +1666,29 @@ void VKContext::BindFramebufferAsRenderTarget(Framebuffer *fbo, const RenderPass
 		assert(res == VK_SUCCESS);
 		// Now, if the image needs transitioning, let's transition.
 		// The backbuffer does not, that's handled by VulkanContext.
-		VkImageMemoryBarrier barrier{};
-		barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
-		barrier.oldLayout = fb->color.layout;
-		barrier.subresourceRange.layerCount = 1;
-		barrier.subresourceRange.levelCount = 1;
-		barrier.image = fb->color.image;
-		barrier.srcAccessMask = 0;
-		switch (fb->color.layout) {
-		case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
-			barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
-			break;
+		if (fb->color.layout != VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL) {
+			VkImageMemoryBarrier barrier{};
+			barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
+			barrier.oldLayout = fb->color.layout;
+			barrier.subresourceRange.layerCount = 1;
+			barrier.subresourceRange.levelCount = 1;
+			barrier.image = fb->color.image;
+			barrier.srcAccessMask = 0;
+			switch (fb->color.layout) {
+			case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
+				barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
+				break;
+			case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL:
+				barrier.srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
+				break;
+			}
+			barrier.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+			barrier.newLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
+			barrier.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+			// TODO: Optimize these flags. 
+			vkCmdPipelineBarrier(cmd_, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, 0, 0, 0, 0, 0, 1, &barrier);
+			fb->color.layout = barrier.newLayout;
 		}
-		barrier.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-		barrier.newLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
-		barrier.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
-		vkCmdPipelineBarrier(cmd_, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0, 0, 0, 0, 0, 1, &barrier);
-		fb->color.layout = barrier.newLayout;
 		renderPass = renderPasses_[RPIndex(rp.color, rp.depth)];
 		// ILOG("Switching framebuffer to FBO (fc=%d, cmd=%x, rp=%x)", frameNum_, (int)(uintptr_t)cmd_, (int)(uintptr_t)renderPass);
 		if (rp.color == RPAction::CLEAR) {
@@ -1719,6 +1725,11 @@ void VKContext::BindFramebufferAsRenderTarget(Framebuffer *fbo, const RenderPass
 // color must be 0, for now.
 void VKContext::BindFramebufferAsTexture(Framebuffer *fbo, int binding, FBChannel channelBit, int attachment) {
 	VKFramebuffer *fb = (VKFramebuffer *)fbo;
+	boundImageView_[0] = fb->color.view;
+	// If we already have the right layout, nothing else to do.
+	if (fb->color.layout == VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL)
+		return;
+
 	VkCommandBuffer transitionCmdBuf;
 	if (fb->cmdBuf && fb->frameCount == frameNum_) {
 		// If the framebuffer has a "live" command buffer, we can directly use it to transition it for sampling.
@@ -1746,10 +1757,11 @@ void VKContext::BindFramebufferAsTexture(Framebuffer *fbo, int binding, FBChanne
 	barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
 	barrier.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 	barrier.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+
 	// we're between passes so it's OK.
-	vkCmdPipelineBarrier(transitionCmdBuf, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0, 0, 0, 0, 0, 1, &barrier);
+	// ARM Best Practices guide recommends these stage bits.
+	vkCmdPipelineBarrier(transitionCmdBuf, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, 0, 0, 0, 1, &barrier);
 	fb->color.layout = barrier.newLayout;
-	boundImageView_[0] = fb->color.view;
 }
 
 void VKContext::BindFramebufferForRead(Framebuffer *fbo) { /* noop */ }