Vulkan: Be more restrictive about texture upscaling on "slow" devices.

Applies the same rules as for software upscaling in this case. Should fix the stutters seen in #15109
hrydgard · Dec 16, 2021 · 80ae4b0 · 80ae4b0
1 parent d294b6a
commit 80ae4b0
Show file tree

Hide file tree

Showing 4 changed files with 55 additions and 5 deletions.
diff --git a/Common/GPU/Vulkan/VulkanContext.cpp b/Common/GPU/Vulkan/VulkanContext.cpp
@@ -5,6 +5,7 @@
 #include <cstring>
 #include <iostream>
 
+#include "Common/System/System.h"
 #include "Common/System/Display.h"
 #include "Common/Log.h"
 #include "Common/GPU/Vulkan/VulkanContext.h"
@@ -680,6 +681,36 @@ VkResult VulkanContext::CreateDevice() {
 	allocatorInfo.instance = instance_;
 	vmaCreateAllocator(&allocatorInfo, &allocator_);
 
+	// Examine the physical device to figure out super rough performance grade.
+	// Basically all we want to do is to identify low performance mobile devices
+	// so we can make decisions on things like texture scaling strategy.
+	auto &props = physicalDeviceProperties_[physical_device_].properties;
+	switch (props.vendorID) {
+	case VULKAN_VENDOR_AMD:
+	case VULKAN_VENDOR_NVIDIA:
+	case VULKAN_VENDOR_INTEL:
+		devicePerfClass_ = PerfClass::FAST;
+		break;
+
+	case VULKAN_VENDOR_ARM:
+	case VULKAN_VENDOR_QUALCOMM:
+		devicePerfClass_ = PerfClass::SLOW;
+		// I haven't seen the bad compute shader slowdowns on any Android 11 device, so let's do
+		// a really silly heuristic here, and improve as needed.
+#if PPSSPP_PLATFORM(ANDROID)
+		if (System_GetPropertyInt(SYSPROP_SYSTEMVERSION) >= 30) {
+			devicePerfClass_ = PerfClass::FAST;
+		}
+#endif
+		break;
+
+	case VULKAN_VENDOR_IMGTEC:
+	default:
+		devicePerfClass_ = PerfClass::SLOW;
+		break;
+	}
+
+
 	for (int i = 0; i < ARRAY_SIZE(frame_); i++) {
 		frame_[i].profiler.Init(this);
 	}

diff --git a/Common/GPU/Vulkan/VulkanContext.h b/Common/GPU/Vulkan/VulkanContext.h
@@ -16,7 +16,7 @@
 // other things as well. We also have a nice integrated render pass profiler in the queue
 // runner, but this one is more convenient for transient events.
 
-// #define VULKAN_PROFILER_ENABLED
+#define VULKAN_PROFILER_ENABLED
 
 #if defined(VULKAN_PROFILER_ENABLED)
 #define VK_PROFILE_BEGIN(vulkan, cmd, stage, message) vulkan->GetProfiler()->Begin(cmd, stage, message);
@@ -80,6 +80,12 @@ struct VulkanPhysicalDeviceInfo {
 
 class VulkanProfiler;
 
+// Extremely rough split of capabilities.
+enum class PerfClass {
+	SLOW,
+	FAST,
+};
+
 // This is a bit repetitive...
 class VulkanDeleteList {
 	struct BufferWithAlloc {
@@ -323,6 +329,10 @@ class VulkanContext {
 
 	const VulkanExtensions &Extensions() { return extensionsLookup_; }
 
+	PerfClass DevicePerfClass() const {
+		return devicePerfClass_;
+	}
+
 	void GetImageMemoryRequirements(VkImage image, VkMemoryRequirements *mem_reqs, bool *dedicatedAllocation);
 
 	VmaAllocator Allocator() const {
@@ -385,6 +395,7 @@ class VulkanContext {
 	VkExtent2D swapChainExtent_{};
 
 	int flags_ = 0;
+	PerfClass devicePerfClass_ = PerfClass::SLOW;
 
 	int inflightFrames_ = MAX_INFLIGHT_FRAMES;
 

diff --git a/GPU/Common/TextureCacheCommon.cpp b/GPU/Common/TextureCacheCommon.cpp
@@ -473,6 +473,7 @@ TexCacheEntry *TextureCacheCommon::SetTexture() {
 				reason = "scaling";
 			}
 		}
+
 		if (match && (entry->status & TexCacheEntry::STATUS_TO_REPLACE) && replacementTimeThisFrame_ < replacementFrameBudget_) {
 			int w0 = gstate.getTextureWidth(0);
 			int h0 = gstate.getTextureHeight(0);

diff --git a/GPU/Vulkan/TextureCacheVulkan.cpp b/GPU/Vulkan/TextureCacheVulkan.cpp
@@ -589,6 +589,8 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		return;
 	}
 
+	VulkanContext *vulkan = (VulkanContext *)draw_->GetNativeObject(Draw::NativeObject::CONTEXT);
+
 	// Adjust maxLevel to actually present levels..
 	bool badMipSizes = false;
 
@@ -659,14 +661,17 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 	if (entry->addr > 0x05000000 && entry->addr < PSP_GetKernelMemoryEnd()) {
 		scaleFactor = 1;
 	}
-	if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1 && !hardwareScaling) {
+
+	bool slowScaler = !hardwareScaling || vulkan->DevicePerfClass() == PerfClass::SLOW;  // Or the GPU is slow - TODO add check!
+
+	if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1 && slowScaler) {
 		// Remember for later that we /wanted/ to scale this texture.
 		entry->status |= TexCacheEntry::STATUS_TO_SCALE;
 		scaleFactor = 1;
 	}
 
 	if (scaleFactor != 1) {
-		if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED && !hardwareScaling) {
+		if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED && slowScaler) {
 			entry->status |= TexCacheEntry::STATUS_TO_SCALE;
 			scaleFactor = 1;
 		} else {
@@ -715,7 +720,6 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 
 	bool computeUpload = false;
 	VkCommandBuffer cmdInit = (VkCommandBuffer)draw_->GetNativeObject(Draw::NativeObject::INIT_COMMANDBUFFER);
-	VulkanContext *vulkan = (VulkanContext *)draw_->GetNativeObject(Draw::NativeObject::CONTEXT);
 
 	{
 		delete entry->vkTex;
@@ -745,8 +749,11 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		VkImageUsageFlags usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
 
 		if (actualFmt == VULKAN_8888_FORMAT && scaleFactor > 1 && hardwareScaling) {
-			if (uploadCS_ != VK_NULL_HANDLE)
+			if (uploadCS_ != VK_NULL_HANDLE) {
 				computeUpload = true;
+			} else {
+				WARN_LOG(G3D, "Falling back to software scaling, hardware shader didn't compile");
+			}
 		}
 
 		if (computeUpload) {