Fix a number of bugs and stuff affecting Vulkan on Mali

Common/Vulkan/VulkanContext.cpp

@@ -86,6 +86,7 @@ VulkanContext::VulkanContext(const char *app_name, int app_ver, uint32_t flags)
 	instance_extension_names.push_back(VK_KHR_ANDROID_SURFACE_EXTENSION_NAME);
 #endif
 	device_extension_names.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
+	// device_extension_names.push_back(VK_NV_DEDICATED_ALLOCATION_EXTENSION_NAME);
 
 	if (flags & VULKAN_FLAG_VALIDATE) {
 		for (size_t i = 0; i < ARRAY_SIZE(validationLayers); i++) {
@@ -217,11 +218,9 @@ void VulkanContext::QueueBeforeSurfaceRender(VkCommandBuffer cmd) {
 
 VkCommandBuffer VulkanContext::BeginFrame() {
 	FrameData *frame = &frame_[curFrame_];
-
 	// Get the index of the next available swapchain image, and a semaphore to block command buffer execution on.
 	// Now, I wonder if we should do this early in the frame or late? Right now we do it early, which should be fine.
 	VkResult res = vkAcquireNextImageKHR(device_, swap_chain_, UINT64_MAX, acquireSemaphore, VK_NULL_HANDLE, &current_buffer);
-
 	// TODO: Deal with the VK_SUBOPTIMAL_KHR and VK_ERROR_OUT_OF_DATE_KHR
 	// return codes
 	assert(res == VK_SUCCESS);
@@ -252,16 +251,13 @@ VkCommandBuffer VulkanContext::BeginSurfaceRenderPass(VkClearValue clear_values[
 	rp_begin.renderArea.extent.height = height_;
 	rp_begin.clearValueCount = 2;
 	rp_begin.pClearValues = clear_values;
-
-	// We don't really need to record this at this point in time, but hey, at some point we'll start this
-	// pass anyway so might as well do it now (although you can imagine getting away with just a stretchblt and not
-	// even starting a final render pass if there's nothing to overlay... hm. Uncommon though on mobile).
 	vkCmdBeginRenderPass(frame->cmdBuf, &rp_begin, VK_SUBPASS_CONTENTS_INLINE);
 	return frame->cmdBuf;
 }
 
 void VulkanContext::EndSurfaceRenderPass() {
 	FrameData *frame = &frame_[curFrame_];
+	ILOG("VulkanContext::EndSurfaceRenderPass");
 	vkCmdEndRenderPass(frame->cmdBuf);
 }
 
@@ -292,21 +288,21 @@ void VulkanContext::EndFrame() {
 	VkSubmitInfo submit_info = { VK_STRUCTURE_TYPE_SUBMIT_INFO };
 	submit_info.waitSemaphoreCount = 1;
 	submit_info.pWaitSemaphores = &acquireSemaphore;
-	VkPipelineStageFlags waitStage[1] = { VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT };
+	VkPipelineStageFlags waitStage[1] = { VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT };
 	submit_info.pWaitDstStageMask = waitStage;
 	submit_info.commandBufferCount = (uint32_t)cmdBufs.size();
 	submit_info.pCommandBuffers = cmdBufs.data();
-	submit_info.signalSemaphoreCount = 0;
-	submit_info.pSignalSemaphores = NULL;
+	submit_info.signalSemaphoreCount = 1;
+	submit_info.pSignalSemaphores = &renderingCompleteSemaphore;
 	res = vkQueueSubmit(gfx_queue_, 1, &submit_info, frame->fence);
 	assert(res == VK_SUCCESS);
 
 	VkPresentInfoKHR present = { VK_STRUCTURE_TYPE_PRESENT_INFO_KHR };
 	present.swapchainCount = 1;
 	present.pSwapchains = &swap_chain_;
 	present.pImageIndices = &current_buffer;
-	present.pWaitSemaphores = NULL;
-	present.waitSemaphoreCount = 0;
+	present.pWaitSemaphores = &renderingCompleteSemaphore;
+	present.waitSemaphoreCount = 1;
 	present.pResults = NULL;
 
 	res = vkQueuePresentKHR(gfx_queue_, &present);
@@ -949,13 +945,11 @@ void VulkanContext::InitQueue() {
 	vkGetDeviceQueue(device_, graphics_queue_family_index_, 0, &gfx_queue_);
 	ILOG("gfx_queue_: %p", gfx_queue_);
 
-	VkSemaphoreCreateInfo acquireSemaphoreCreateInfo = { VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO };
-	acquireSemaphoreCreateInfo.flags = 0;
-
-	res = vkCreateSemaphore(device_,
-		&acquireSemaphoreCreateInfo,
-		NULL,
-		&acquireSemaphore);
+	VkSemaphoreCreateInfo semaphoreCreateInfo = { VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO };
+	semaphoreCreateInfo.flags = 0;
+	res = vkCreateSemaphore(device_, &semaphoreCreateInfo, NULL, &acquireSemaphore);
+	assert(res == VK_SUCCESS);
+	res = vkCreateSemaphore(device_, &semaphoreCreateInfo, NULL, &renderingCompleteSemaphore);
 	assert(res == VK_SUCCESS);
 }
 
@@ -1070,7 +1064,9 @@ bool VulkanContext::InitSwapchain(VkCommandBuffer cmd) {
 		&swapchainImageCount, NULL);
 	assert(res == VK_SUCCESS);
 
-	VkImage* swapchainImages = (VkImage*)malloc(swapchainImageCount * sizeof(VkImage));
+	ILOG("Vulkan swapchain image count: %d", swapchainImageCount);
+
+	VkImage* swapchainImages = new VkImage[swapchainImageCount];
 	assert(swapchainImages);
 	res = vkGetSwapchainImagesKHR(device_, swap_chain_, &swapchainImageCount, swapchainImages);
 	assert(res == VK_SUCCESS);
@@ -1108,8 +1104,7 @@ bool VulkanContext::InitSwapchain(VkCommandBuffer cmd) {
 		swapChainBuffers.push_back(sc_buffer);
 		assert(res == VK_SUCCESS);
 	}
-	free(swapchainImages);
-
+	delete[] swapchainImages;
 	current_buffer = 0;
 
 	return true;
@@ -1249,6 +1244,7 @@ void VulkanContext::DestroySwapChain() {
 	swap_chain_ = VK_NULL_HANDLE;
 	swapChainBuffers.clear();
 	vkDestroySemaphore(device_, acquireSemaphore, NULL);
+	vkDestroySemaphore(device_, renderingCompleteSemaphore, NULL);
 }
 
 void VulkanContext::DestroyFramebuffers() {

Common/Vulkan/VulkanContext.h

@@ -317,6 +317,7 @@ class VulkanContext {
 
 private:
 	VkSemaphore acquireSemaphore;
+	VkSemaphore renderingCompleteSemaphore;
 
 #ifdef _WIN32
 	HINSTANCE connection;        // hInstance - Windows Instance

Common/Vulkan/VulkanMemory.h

@@ -51,6 +51,7 @@ class VulkanPushBuffer {
 	void Unmap() {
 		assert(writePtr_);
 		/*
+		// Should not need this since we use coherent memory.
 		VkMappedMemoryRange range = { VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE };
 		range.offset = 0;
 		range.size = offset_;

GPU/GLES/DrawEngineGLES.cpp

@@ -645,6 +645,8 @@ void DrawEngineGLES::DoFlush() {
 	PROFILE_THIS_SCOPE("flush");
 	CHECK_GL_ERROR_IF_DEBUG();
 
+
+
 	gpuStats.numFlushes++;
 	gpuStats.numTrackedVertexArrays = (int)vai_.size();
 

GPU/GLES/FramebufferManagerGLES.cpp

@@ -1089,7 +1089,7 @@ void FramebufferManagerGLES::PackFramebufferSync_(VirtualFramebuffer *vfb, int x
 	if (gl_extensions.GLES3 && glInvalidateFramebuffer != nullptr) {
 #ifdef USING_GLES2
 		// GLES3 doesn't support using GL_READ_FRAMEBUFFER here.
-		draw_->BindFramebufferAsRenderTarget(vfb->fbo);
+		draw_->BindFramebufferAsRenderTarget(vfb->fbo, { Draw::RPAction::DONT_CARE, Draw::RPAction::DONT_CARE });
 		const GLenum target = GL_FRAMEBUFFER;
 #else
 		const GLenum target = GL_READ_FRAMEBUFFER;

GPU/GLES/GPU_GLES.cpp

@@ -660,7 +660,7 @@ void GPU_GLES::Execute_Prim(u32 op, u32 diff) {
 			return;
 	}
 
-	// This also makes skipping drawing very effective.
+	// This also makes skipping drawing very effective. This function can change the framebffer.
 	framebufferManagerGL_->SetRenderFrameBuffer(gstate_c.IsDirty(DIRTY_FRAMEBUF), gstate_c.skipDrawReason);
 	if (gstate_c.skipDrawReason & (SKIPDRAW_SKIPFRAME | SKIPDRAW_NON_DISPLAYED_FB))	{
 		drawEngine_.SetupVertexDecoder(gstate.vertType);

GPU/Software/SoftGpu.cpp

@@ -205,7 +205,7 @@ void SoftGPU::CopyToCurrentFboFromDisplayRam(int srcwidth, int srcheight) {
 		u1 = 1.0f;
 	}
 	if (!hasImage) {
-		draw_->Clear(Draw::FB_COLOR_BIT, 0, 0, 0);
+		draw_->BindFramebufferAsRenderTarget(nullptr, { Draw::RPAction::CLEAR, Draw::RPAction::DONT_CARE });
 		return;
 	}
 

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -666,6 +666,10 @@ void DrawEngineVulkan::DirtyAllUBOs() {
 // The inline wrapper in the header checks for numDrawCalls == 0d
 void DrawEngineVulkan::DoFlush() {
 	VkCommandBuffer cmd = (VkCommandBuffer)draw_->GetNativeObject(Draw::NativeObject::RENDERPASS_COMMANDBUFFER);
+	VkRenderPass rp = (VkRenderPass)draw_->GetNativeObject(Draw::NativeObject::CURRENT_RENDERPASS);
+	if (!rp)
+		Crash();
+
 	gpuStats.numFlushes++;
 
 	FrameData *frame = &frame_[curFrame_ & 1];
@@ -847,7 +851,6 @@ void DrawEngineVulkan::DoFlush() {
 				Uint8x4ToFloat4(bc, dynState.blendColor);
 				vkCmdSetBlendConstants(cmd, bc);
 			}
-
 			dirtyUniforms_ |= shaderManager_->UpdateUniforms();
 
 			shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, useHWTransform);

GPU/Vulkan/FramebufferVulkan.cpp

@@ -461,6 +461,9 @@ VkImageView FramebufferManagerVulkan::BindFramebufferAsColorTexture(int stage, V
 	}
 	// Currently rendering to this framebuffer. Need to make a copy.
 	if (!skipCopy && framebuffer == currentRenderVfb_) {
+		// ignore this case for now, doesn't work
+		ILOG("Texturing from current render Vfb!");
+		return VK_NULL_HANDLE;
 		// TODO: Maybe merge with bvfbs_?  Not sure if those could be packing, and they're created at a different size.
 		Draw::Framebuffer *renderCopy = GetTempFBO(framebuffer->renderWidth, framebuffer->renderHeight, (Draw::FBColorDepth)framebuffer->colorDepth);
 		if (renderCopy) {
@@ -1036,6 +1039,8 @@ void FramebufferManagerVulkan::FlushBeforeCopy() {
 	// all the irrelevant state checking it'll use to decide what to do. Should
 	// do something more focused here.
 	SetRenderFrameBuffer(gstate_c.IsDirty(DIRTY_FRAMEBUF), gstate_c.skipDrawReason);
+	if (!draw_->GetNativeObject(Draw::NativeObject::CURRENT_RENDERPASS))
+		Crash();
 	drawEngine_->Flush();
 }
 

GPU/Vulkan/GPU_Vulkan.cpp

@@ -489,6 +489,8 @@ void GPU_Vulkan::Execute_Prim(u32 op, u32 diff) {
 
 	// This also makes skipping drawing very effective.
 	framebufferManager_->SetRenderFrameBuffer(gstate_c.IsDirty(DIRTY_FRAMEBUF), gstate_c.skipDrawReason);
+	if (!draw_->GetNativeObject(Draw::NativeObject::CURRENT_RENDERPASS))
+		Crash();
 
 	if (gstate_c.skipDrawReason & (SKIPDRAW_SKIPFRAME | SKIPDRAW_NON_DISPLAYED_FB)) {
 		drawEngine_.SetupVertexDecoder(gstate.vertType);

GPU/Vulkan/PipelineManagerVulkan.cpp

@@ -294,6 +294,9 @@ static VulkanPipeline *CreateVulkanPipeline(VkDevice device, VkPipelineCache pip
 
 VulkanPipeline *PipelineManagerVulkan::GetOrCreatePipeline(VkPipelineLayout layout, VkRenderPass renderPass, const VulkanPipelineRasterStateKey &rasterKey, const VertexDecoder *vtxDec, VulkanVertexShader *vs, VulkanFragmentShader *fs, bool useHwTransform) {
 	VulkanPipelineKey key;
+	if (!renderPass)
+		Crash();
+
 	key.raster = rasterKey;
 	key.renderPass = renderPass;
 	key.useHWTransform = useHwTransform;

GPU/Vulkan/ShaderManagerVulkan.cpp

@@ -16,7 +16,7 @@
 // https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
 
 #ifdef _WIN32
-#define SHADERLOG
+//#define SHADERLOG
 #endif
 
 #include <map>

GPU/Vulkan/TextureCacheVulkan.cpp

@@ -612,6 +612,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry, bool replaceIm
 			entry->vkTex = nullptr;
 		}
 	} else {
+		entry->vkTex->texture_->TransitionForUpload();
 		// TODO: If reusing an existing texture object, we must transition it into the correct layout.
 	}
 	lastBoundTexture = entry->vkTex;

UI/EmuScreen.cpp

@@ -960,7 +960,7 @@ void EmuScreen::preRender() {
 	// We do, however, start the frame in other ways.
 
 	bool useBufferedRendering = g_Config.iRenderingMode != FB_NON_BUFFERED_MODE;
-	if (!useBufferedRendering) {
+	if (!useBufferedRendering && !g_Config.bSoftwareRendering) {
 		// We need to clear here already so that drawing during the frame is done on a clean slate.
 		DrawContext *draw = screenManager()->getDrawContext();
 		draw->BindFramebufferAsRenderTarget(nullptr, { RPAction::CLEAR, RPAction::CLEAR, 0xFF000000 });

UI/GameSettingsScreen.cpp

@@ -187,7 +187,7 @@ void GameSettingsScreen::CreateViews() {
 		renderingBackendChoice->HideChoice(2);  // D3D11
 	}
 #endif
-#if !defined(_WIN32)
+#if !defined(_WIN32) && !PPSSPP_PLATFORM(ANDROID)
 	// TODO: Add dynamic runtime check for Vulkan support on Android
 	renderingBackendChoice->HideChoice(3);
 #endif

UI/MiscScreens.cpp

@@ -510,7 +510,7 @@ void LogoScreen::render() {
 		dc.DrawTextShadow(boot_filename.c_str(), bounds.centerX(), bounds.centerY() + 180, textColor, ALIGN_CENTER);
 	}
 
-#if defined(_WIN32) && !PPSSPP_PLATFORM(UWP)
+#if (defined(_WIN32) && !PPSSPP_PLATFORM(UWP)) || PPSSPP_PLATFORM(ANDROID)
 	// Draw the graphics API, except on UWP where it's always D3D11
 	dc.DrawText(screenManager()->getDrawContext()->GetInfoString(InfoField::APINAME).c_str(), bounds.centerX(), bounds.y2() - 100, textColor, ALIGN_CENTER);
 #endif

ext/native/thin3d/thin3d_gl.cpp

@@ -1557,7 +1557,7 @@ void OpenGLContext::BindFramebufferAsRenderTarget(Framebuffer *fbo, const Render
 		glstate.colorMask.force(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
 	}
 	if (rp.depth == RPAction::CLEAR) {
-		glClearDepth(rp.clearDepth);
+		glClearDepthf(rp.clearDepth);
 		glClearStencil(rp.clearStencil);
 		clearFlags |= GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT;
 		glstate.depthWrite.force(GL_TRUE);