Merge branch 'master' into TemporaryHackishBranch

Common/Vulkan/VulkanContext.cpp

@@ -1391,3 +1391,43 @@ void VulkanContext::GetImageMemoryRequirements(VkImage image, VkMemoryRequiremen
 		*dedicatedAllocation = false;
 	}
 }
+
+bool IsHashMaliDriverVersion(const VkPhysicalDeviceProperties &props) {
+	// ARM used to put a hash in place of the driver version.
+	// Now they only use major versions. We'll just make a bad heuristic.
+	uint32_t major = VK_VERSION_MAJOR(props.driverVersion);
+	uint32_t minor = VK_VERSION_MINOR(props.driverVersion);
+	uint32_t branch = VK_VERSION_PATCH(props.driverVersion);
+	if (branch > 0)
+		return true;
+	if (branch > 100 || major > 100)
+		return true;
+	return false;
+}
+
+// From Sascha's code
+std::string FormatDriverVersion(const VkPhysicalDeviceProperties &props) {
+	if (props.vendorID == VULKAN_VENDOR_NVIDIA) {
+		// For whatever reason, NVIDIA has their own scheme.
+		// 10 bits = major version (up to r1023)
+		// 8 bits = minor version (up to 255)
+		// 8 bits = secondary branch version/build version (up to 255)
+		// 6 bits = tertiary branch/build version (up to 63)
+		uint32_t major = (props.driverVersion >> 22) & 0x3ff;
+		uint32_t minor = (props.driverVersion >> 14) & 0x0ff;
+		uint32_t secondaryBranch = (props.driverVersion >> 6) & 0x0ff;
+		uint32_t tertiaryBranch = (props.driverVersion) & 0x003f;
+		return StringFromFormat("%d.%d.%d.%d", major, minor, secondaryBranch, tertiaryBranch);
+	} else if (props.vendorID == VULKAN_VENDOR_ARM) {
+		// ARM used to just put a hash here. No point in splitting it up.
+		if (IsHashMaliDriverVersion(props)) {
+			return StringFromFormat("(hash) %08x", props.driverVersion);
+		}
+	}
+	// Qualcomm has an inscrutable versioning scheme. Let's just display it as normal.
+	// Standard scheme, use the standard macros.
+	uint32_t major = VK_VERSION_MAJOR(props.driverVersion);
+	uint32_t minor = VK_VERSION_MINOR(props.driverVersion);
+	uint32_t branch = VK_VERSION_PATCH(props.driverVersion);
+	return StringFromFormat("%d.%d.%d (%08x)", major, minor, branch, props.driverVersion);
+}

Common/Vulkan/VulkanContext.h

@@ -359,3 +359,7 @@ void finalize_glslang();
 bool GLSLtoSPV(const VkShaderStageFlagBits shader_type, const char *pshader, std::vector<uint32_t> &spirv, std::string *errorMessage = nullptr);
 
 const char *VulkanResultToString(VkResult res);
+std::string FormatDriverVersion(const VkPhysicalDeviceProperties &props);
+
+// Simple heuristic.
+bool IsHashMaliDriverVersion(const VkPhysicalDeviceProperties &props);

GPU/Common/ShaderUniforms.cpp

@@ -150,7 +150,6 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool flipView
 		if (g_Config.iRenderingMode == FB_NON_BUFFERED_MODE && g_display_rotation != DisplayRotation::ROTATE_0) {
 			flippedMatrix = flippedMatrix * g_display_rot_matrix;
 		}
-
 		CopyMatrix4x4(ub->proj, flippedMatrix.getReadPtr());
 	}
 

GPU/GPUState.h

@@ -495,6 +495,7 @@ enum {
 	GPU_SUPPORTS_ARB_FRAMEBUFFER_BLIT = FLAG_BIT(26),
 	GPU_SUPPORTS_NV_FRAMEBUFFER_BLIT = FLAG_BIT(27),
 	GPU_SUPPORTS_OES_TEXTURE_NPOT = FLAG_BIT(28),
+	GPU_NEEDS_Z_EQUAL_W_HACK = FLAG_BIT(29),
 	GPU_PREFER_CPU_DOWNLOAD = FLAG_BIT(30),
 	GPU_PREFER_REVERSE_COLOR_ORDER = FLAG_BIT(31),
 };

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -970,7 +970,7 @@ void DrawEngineVulkan::DoFlush() {
 			int scissorY2 = gstate.getScissorY2() + 1;
 			framebufferManager_->SetSafeSize(scissorX2, scissorY2);
 
-			if ((gstate_c.featureFlags & GPU_USE_CLEAR_RAM_HACK) && gstate.isClearModeColorMask() && (gstate.isClearModeAlphaMask() || gstate.FrameBufFormat() == GE_FORMAT_565)) {
+			if (gstate_c.Supports(GPU_USE_CLEAR_RAM_HACK) && gstate.isClearModeColorMask() && (gstate.isClearModeAlphaMask() || gstate.FrameBufFormat() == GE_FORMAT_565)) {
 				framebufferManager_->ApplyClearToMemory(scissorX1, scissorY1, scissorX2, scissorY2, result.color);
 			}
 		}

GPU/Vulkan/GPU_Vulkan.cpp

@@ -17,6 +17,7 @@
 // https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
 
 #include <thread>
+
 #include "base/logging.h"
 #include "base/timeutil.h"
 #include "profiler/profiler.h"
@@ -197,15 +198,23 @@ void GPU_Vulkan::CheckGPUFeatures() {
 		features |= GPU_SUPPORTS_ACCURATE_DEPTH;
 		break;
 	case VULKAN_VENDOR_ARM:
-		// Also required on older ARM Mali drivers, like the one on many Galaxy S7.
-		if (!PSP_CoreParameter().compat.flags().DisableAccurateDepth ||
-			  vulkan_->GetPhysicalDeviceProperties().properties.driverVersion <= VK_MAKE_VERSION(428, 811, 2674)) {
+	{
+		// This check is probably not exactly accurate. But old drivers had problems with reverse-Z, just like AMD and Qualcomm.
+		bool driverTooOld = IsHashMaliDriverVersion(vulkan_->GetPhysicalDeviceProperties().properties)
+			|| VK_VERSION_MAJOR(vulkan_->GetPhysicalDeviceProperties().properties.driverVersion) < 14;
+		if (!PSP_CoreParameter().compat.flags().DisableAccurateDepth || driverTooOld) {
 			features |= GPU_SUPPORTS_ACCURATE_DEPTH;
 		}
+		// These GPUs (up to some certain hardware version?) has a bug where draws where gl_Position.w == .z
+		// corrupt the depth buffer. This is easily worked around by simply scaling Z down a tiny bit when this case
+		// is detected. See: https://github.com/hrydgard/ppsspp/issues/11937
+		features |= GPU_NEEDS_Z_EQUAL_W_HACK;
 		break;
+	}
 	default:
-		if (!PSP_CoreParameter().compat.flags().DisableAccurateDepth)
+		if (!PSP_CoreParameter().compat.flags().DisableAccurateDepth) {
 			features |= GPU_SUPPORTS_ACCURATE_DEPTH;
+		}
 		break;
 	}
 

GPU/Vulkan/VertexShaderGeneratorVulkan.cpp

@@ -176,8 +176,7 @@ bool GenerateVulkanGLSLVertexShader(const VShaderID &id, char *buffer) {
 		if (!useHWTransform && doTextureTransform && !isModeThrough) {
 			WRITE(p, "layout (location = %d) in vec3 texcoord;\n", (int)PspAttributeLocation::TEXCOORD);
 			texcoordInVec3 = true;
-		}
-		else
+		} else
 			WRITE(p, "layout (location = %d) in vec2 texcoord;\n", (int)PspAttributeLocation::TEXCOORD);
 	}
 	if (hasColor) {
@@ -640,6 +639,10 @@ bool GenerateVulkanGLSLVertexShader(const VShaderID &id, char *buffer) {
 	}
 	WRITE(p, "  gl_Position = outPos;\n");
 
+	if (gstate_c.Supports(GPU_NEEDS_Z_EQUAL_W_HACK)) {
+		// See comment in GPU_Vulkan.cpp.
+		WRITE(p, "  if (gl_Position.z == gl_Position.w) gl_Position.z *= 0.999999;\n");
+	}
 	WRITE(p, "}\n");
 	return true;
 }

ext/native/base/display.cpp

@@ -1,3 +1,4 @@
+#include <algorithm>
 #include "base/display.h"
 
 int dp_xres;
@@ -17,3 +18,36 @@ float display_hz = 60.0f;
 
 DisplayRotation g_display_rotation;
 Matrix4x4 g_display_rot_matrix;
+
+void RotateRectToDisplay(FRect &rect, float curRTWidth, float curRTHeight) {
+	switch (g_display_rotation) {
+	case DisplayRotation::ROTATE_180:
+		rect.x = curRTWidth - rect.w - rect.x;
+		rect.y = curRTHeight - rect.h - rect.y;
+		break;
+	case DisplayRotation::ROTATE_90: {
+		// Note that curRTWidth_ and curRTHeight_ are "swapped"!
+		float origX = rect.x;
+		float origY = rect.y;
+		float rtw = curRTHeight;
+		float rth = curRTWidth;
+		rect.x = rth - rect.h - origY;
+		rect.y = origX;
+		std::swap(rect.w, rect.h);
+		break;
+	}
+	case DisplayRotation::ROTATE_270: {
+		float origX = rect.x;
+		float origY = rect.y;
+		float rtw = curRTHeight;
+		float rth = curRTWidth;
+		rect.x = origY;
+		rect.y = rtw - rect.w - origX;
+		std::swap(rect.w, rect.h);
+		break;
+	}
+	case DisplayRotation::ROTATE_0:
+	default:
+		break;
+	}
+}

ext/native/base/display.h

@@ -30,3 +30,8 @@ enum class DisplayRotation {
 
 extern DisplayRotation g_display_rotation;
 extern Matrix4x4 g_display_rot_matrix;
+
+struct FRect {
+	float x, y, w, h;
+};
+void RotateRectToDisplay(FRect &rect, float curRTWidth, float curRTHeight);

ext/native/math/math_util.h

@@ -220,5 +220,3 @@ void EnableFZ();
 // very slow VFP units).
 // http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0274h/Babffifj.html
 void FPU_SetFastMode();
-
-

ext/native/thin3d/VulkanRenderManager.h

@@ -10,6 +10,7 @@
 #include <mutex>
 #include <thread>
 
+#include "base/display.h"
 #include "Common/Vulkan/VulkanContext.h"
 #include "math/dataconv.h"
 #include "math/math_util.h"
@@ -119,15 +120,21 @@ public:
 	void SetViewport(const VkViewport &vp) {
 		_dbg_assert_(G3D, curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER);
 		_dbg_assert_(G3D, (int)vp.width >= 0);
+		FRect rc{ vp.x, vp.y, vp.width, vp.height };
+		if (curRenderStep_->render.framebuffer == nullptr) { // Only the backbuffer is actually rotated wrong!
+			RotateRectToDisplay(rc, (float)vulkan_->GetBackbufferWidth(), (float)vulkan_->GetBackbufferHeight());
+		}
 		VkRenderData data{ VKRRenderCommand::VIEWPORT };
-		data.viewport.vp.x = vp.x;
-		data.viewport.vp.y = vp.y;
-		data.viewport.vp.width = vp.width;
-		data.viewport.vp.height = vp.height;
+		data.viewport.vp.x = rc.x;
+		data.viewport.vp.y = rc.y;
+		data.viewport.vp.width = rc.w;
+		data.viewport.vp.height = rc.h;
 		// We can't allow values outside this range unless we use VK_EXT_depth_range_unrestricted.
 		// Sometimes state mapping produces 65536/65535 which is slightly outside.
+		// TODO: This should be fixed at the source.
 		data.viewport.vp.maxDepth = clamp_value(vp.maxDepth, 0.0f, 1.0f);
 		data.viewport.vp.minDepth = clamp_value(vp.minDepth, 0.0f, 1.0f);
+
 		curRenderStep_->commands.push_back(data);
 	}
 
@@ -136,7 +143,15 @@ public:
 		_dbg_assert_(G3D, (int)rc.extent.width >= 0);
 		_dbg_assert_(G3D, (int)rc.extent.height >= 0);
 		VkRenderData data{ VKRRenderCommand::SCISSOR };
-		data.scissor.scissor = rc;
+		if (curRenderStep_->render.framebuffer == nullptr) {
+			FRect frc{ (float)rc.offset.x, (float)rc.offset.y, (float)rc.extent.width, (float)rc.extent.height };
+			if (curRenderStep_->render.framebuffer == nullptr) { // Only the backbuffer is actually rotated wrong!
+				RotateRectToDisplay(frc, (float)vulkan_->GetBackbufferWidth(), (float)vulkan_->GetBackbufferHeight());
+			}
+			data.scissor.scissor = VkRect2D{ { (int32_t)frc.x, (int32_t)frc.y }, { (uint32_t)frc.w, (uint32_t)frc.h} };
+		} else {
+			data.scissor.scissor = rc;
+		}
 		curRenderStep_->commands.push_back(data);
 	}
 

ext/native/thin3d/thin3d.cpp

@@ -353,38 +353,6 @@ DrawContext::~DrawContext() {
 	DestroyPresets();
 }
 
-void DrawContext::RotateRectToDisplay(FRect &rect, float curRTWidth, float curRTHeight) {
-	if (g_display_rotation == DisplayRotation::ROTATE_0)
-		return;
-	switch (g_display_rotation) {
-	case DisplayRotation::ROTATE_180:
-		rect.x = curRTWidth - rect.w - rect.x;
-		rect.y = curRTHeight - rect.h - rect.y;
-		break;
-	case DisplayRotation::ROTATE_90: {
-		// Note that curRTWidth_ and curRTHeight_ are "swapped"!
-		float origX = rect.x;
-		float origY = rect.y;
-		float rtw = curRTHeight;
-		float rth = curRTWidth;
-		rect.x = rth - rect.h - origY;
-		rect.y = origX;
-		std::swap(rect.w, rect.h);
-		break;
-	}
-	case DisplayRotation::ROTATE_270: {
-		float origX = rect.x;
-		float origY = rect.y;
-		float rtw = curRTHeight;
-		float rth = curRTWidth;
-		rect.x = origY;
-		rect.y = rtw - rect.w - origX;
-		std::swap(rect.w, rect.h);
-		break;
-	}
-	}
-}
-
 // TODO: SSE/NEON
 // Could also make C fake-simd for 64-bit, two 8888 pixels fit in a register :)
 void ConvertFromRGBA8888(uint8_t *dst, const uint8_t *src, uint32_t dstStride, uint32_t srcStride, uint32_t width, uint32_t height, DataFormat format) {

ext/native/thin3d/thin3d.h

@@ -661,11 +661,6 @@ public:
 	virtual void FlushState() {}
 
 protected:
-	struct FRect {
-		float x, y, w, h;
-	};
-	void RotateRectToDisplay(FRect &rect, float curRTWidth, float curRTHeight);
-
 	ShaderModule *vsPresets_[VS_MAX_PRESET];
 	ShaderModule *fsPresets_[FS_MAX_PRESET];
 

ext/native/thin3d/thin3d_vulkan.cpp

@@ -429,30 +429,6 @@ public:
 
 	void FlushState() override {}
 
-	// From Sascha's code
-	static std::string FormatDriverVersion(const VkPhysicalDeviceProperties &props) {
-		if (props.vendorID == VULKAN_VENDOR_NVIDIA) {
-			// 10 bits = major version (up to r1023)
-			// 8 bits = minor version (up to 255)
-			// 8 bits = secondary branch version/build version (up to 255)
-			// 6 bits = tertiary branch/build version (up to 63)
-			uint32_t major = (props.driverVersion >> 22) & 0x3ff;
-			uint32_t minor = (props.driverVersion >> 14) & 0x0ff;
-			uint32_t secondaryBranch = (props.driverVersion >> 6) & 0x0ff;
-			uint32_t tertiaryBranch = (props.driverVersion) & 0x003f;
-			return StringFromFormat("%d.%d.%d.%d (%08x)", major, minor, secondaryBranch, tertiaryBranch, props.driverVersion);
-		} else if (props.vendorID == VULKAN_VENDOR_ARM) {
-			// ARM just puts a hash here, let's just output it as is.
-			return StringFromFormat("%08x", props.driverVersion);
-		} else {
-			// Standard scheme, use the standard macros.
-			uint32_t major = VK_VERSION_MAJOR(props.driverVersion);
-			uint32_t minor = VK_VERSION_MINOR(props.driverVersion);
-			uint32_t branch = VK_VERSION_PATCH(props.driverVersion);
-			return StringFromFormat("%d.%d.%d (%08x)", major, minor, branch, props.driverVersion);
-		}
-	}
-
 	std::string GetInfoString(InfoField info) const override {
 		// TODO: Make these actually query the right information
 		switch (info) {
@@ -1059,13 +1035,7 @@ Pipeline *VKContext::CreateGraphicsPipeline(const PipelineDesc &desc) {
 }
 
 void VKContext::SetScissorRect(int left, int top, int width, int height) {
-	FRect rc{ (float)left, (float)top, (float)width, (float)height };
-	if (curFramebuffer_ == nullptr) { // Only the backbuffer is actually rotated wrong!
-		int curRTWidth, curRTHeight;
-		GetFramebufferDimensions((Framebuffer *)curFramebuffer_, &curRTWidth, &curRTHeight);
-		RotateRectToDisplay(rc, (float)curRTWidth, (float)curRTHeight);
-	}
-	VkRect2D scissor{ {(int32_t)rc.x, (int32_t)rc.y}, {(uint32_t)rc.w, (uint32_t)rc.h} };
+	VkRect2D scissor{ {(int32_t)left, (int32_t)top}, {(uint32_t)width, (uint32_t)height} };
 	renderManager_.SetScissor(scissor);
 }
 
@@ -1074,11 +1044,6 @@ void VKContext::SetViewports(int count, Viewport *viewports) {
 		// Ignore viewports more than the first.
 		VkViewport viewport;
 		FRect rc{ viewports[0].TopLeftX , viewports[0].TopLeftY, viewports[0].Width, viewports[0].Height };
-		if (curFramebuffer_ == nullptr) { // Only the backbuffer is actually rotated wrong!
-			int curRTWidth, curRTHeight;
-			GetFramebufferDimensions((Framebuffer *)curFramebuffer_, &curRTWidth, &curRTHeight);
-			RotateRectToDisplay(rc, (float)curRTWidth, (float)curRTHeight);
-		}
 		viewport.x = rc.x;
 		viewport.y = rc.y;
 		viewport.width = rc.w;