diff --git a/Common/Vulkan/VulkanImage.cpp b/Common/Vulkan/VulkanImage.cpp
index 8ad99b774d..614cd0e3fb 100644
--- a/Common/Vulkan/VulkanImage.cpp
+++ b/Common/Vulkan/VulkanImage.cpp
@@ -121,9 +121,10 @@ bool VulkanTexture::CreateDirect(VkCommandBuffer cmd, VulkanDeviceAllocator *all
 	if (initialLayout != VK_IMAGE_LAYOUT_UNDEFINED && initialLayout != VK_IMAGE_LAYOUT_PREINITIALIZED) {
 		switch (initialLayout) {
 		case VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL:
+		case VK_IMAGE_LAYOUT_GENERAL:
 			TransitionImageLayout2(cmd, image_, 0, numMips, VK_IMAGE_ASPECT_COLOR_BIT,
-				VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
-				VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+				VK_IMAGE_LAYOUT_UNDEFINED, initialLayout,
+				VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
 				0, VK_ACCESS_TRANSFER_WRITE_BIT);
 			break;
 		default:
@@ -208,10 +209,10 @@ void VulkanTexture::GenerateMip(VkCommandBuffer cmd, int mip) {
 		VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_TRANSFER_WRITE_BIT);
 }
 
-void VulkanTexture::EndCreate(VkCommandBuffer cmd, bool vertexTexture) {
+void VulkanTexture::EndCreate(VkCommandBuffer cmd, bool vertexTexture, VkImageLayout layout) {
 	TransitionImageLayout2(cmd, image_, 0, numMips_,
 		VK_IMAGE_ASPECT_COLOR_BIT,
-		VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+		layout, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
 		VK_PIPELINE_STAGE_TRANSFER_BIT, vertexTexture ? VK_PIPELINE_STAGE_VERTEX_SHADER_BIT : VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
 		VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT);
 }
@@ -222,6 +223,26 @@ void VulkanTexture::Touch() {
 	}
 }
 
+VkImageView VulkanTexture::CreateViewForMip(int mip) {
+	VkImageViewCreateInfo view_info = { VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO };
+	view_info.image = image_;
+	view_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
+	view_info.format = format_;
+	view_info.components.r = VK_COMPONENT_SWIZZLE_R;
+	view_info.components.g = VK_COMPONENT_SWIZZLE_G;
+	view_info.components.b = VK_COMPONENT_SWIZZLE_B;
+	view_info.components.a = VK_COMPONENT_SWIZZLE_A;
+	view_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+	view_info.subresourceRange.baseMipLevel = mip;
+	view_info.subresourceRange.levelCount = 1;
+	view_info.subresourceRange.baseArrayLayer = 0;
+	view_info.subresourceRange.layerCount = 1;
+	VkImageView view;
+	VkResult res = vkCreateImageView(vulkan_->GetDevice(), &view_info, NULL, &view);
+	assert(res == VK_SUCCESS);
+	return view;
+}
+
 void VulkanTexture::Destroy() {
 	if (view_ != VK_NULL_HANDLE) {
 		vulkan_->Delete().QueueDeleteImageView(view_);
diff --git a/Common/Vulkan/VulkanImage.h b/Common/Vulkan/VulkanImage.h
index 64f84a8997..4fd06e21c0 100644
--- a/Common/Vulkan/VulkanImage.h
+++ b/Common/Vulkan/VulkanImage.h
@@ -21,7 +21,11 @@ public:
 	bool CreateDirect(VkCommandBuffer cmd, VulkanDeviceAllocator *allocator, int w, int h, int numMips, VkFormat format, VkImageLayout initialLayout, VkImageUsageFlags usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, const VkComponentMapping *mapping = nullptr);
 	void UploadMip(VkCommandBuffer cmd, int mip, int mipWidth, int mipHeight, VkBuffer buffer, uint32_t offset, size_t rowLength);  // rowLength is in pixels
 	void GenerateMip(VkCommandBuffer cmd, int mip);
-	void EndCreate(VkCommandBuffer cmd, bool vertexTexture = false);
+	void EndCreate(VkCommandBuffer cmd, bool vertexTexture = false, VkImageLayout layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+
+	// When loading mips from compute shaders, you need to pass VK_IMAGE_LAYOUT_GENERAL to the above function.
+	// In addition, ignore UploadMip and GenerateMip, and instead use GetViewForMip. Make sure to delete the returned views when used.
+	VkImageView CreateViewForMip(int mip);
 
 	void Destroy();
 
diff --git a/Common/Vulkan/VulkanMemory.cpp b/Common/Vulkan/VulkanMemory.cpp
index 8edae36aca..3673d3ea6f 100644
--- a/Common/Vulkan/VulkanMemory.cpp
+++ b/Common/Vulkan/VulkanMemory.cpp
@@ -23,8 +23,8 @@
 #include "base/timeutil.h"
 #include "math/math_util.h"
 
-VulkanPushBuffer::VulkanPushBuffer(VulkanContext *vulkan, size_t size, VkBufferUsageFlags usage)
-		: vulkan_(vulkan), size_(size), usage_(usage) {
+VulkanPushBuffer::VulkanPushBuffer(VulkanContext *vulkan, size_t size, VkBufferUsageFlags usage, VkMemoryPropertyFlags memoryPropertyMask)
+		: vulkan_(vulkan), memoryPropertyMask_(memoryPropertyMask), size_(size), usage_(usage) {
 	bool res = AddBuffer();
 	assert(res);
 }
@@ -58,7 +58,7 @@ bool VulkanPushBuffer::AddBuffer() {
 	// Okay, that's the buffer. Now let's allocate some memory for it.
 	VkMemoryAllocateInfo alloc{ VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
 	alloc.allocationSize = reqs.size;
-	vulkan_->MemoryTypeFromProperties(reqs.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, &alloc.memoryTypeIndex);
+	vulkan_->MemoryTypeFromProperties(reqs.memoryTypeBits, memoryPropertyMask_, &alloc.memoryTypeIndex);
 
 	res = vkAllocateMemory(device, &alloc, nullptr, &info.deviceMemory);
 	if (VK_SUCCESS != res) {
@@ -89,7 +89,8 @@ void VulkanPushBuffer::Destroy(VulkanContext *vulkan) {
 
 void VulkanPushBuffer::NextBuffer(size_t minSize) {
 	// First, unmap the current memory.
-	Unmap();
+	if (memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+		Unmap();
 
 	buf_++;
 	if (buf_ >= buffers_.size() || minSize > size_) {
@@ -108,7 +109,8 @@ void VulkanPushBuffer::NextBuffer(size_t minSize) {
 
 	// Now, move to the next buffer and map it.
 	offset_ = 0;
-	Map();
+	if (memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+		Map();
 }
 
 void VulkanPushBuffer::Defragment(VulkanContext *vulkan) {
@@ -142,14 +144,15 @@ void VulkanPushBuffer::Map() {
 
 void VulkanPushBuffer::Unmap() {
 	_dbg_assert_(G3D, writePtr_ != 0);
-	/*
-	// Should not need this since we use coherent memory.
-	VkMappedMemoryRange range{ VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE };
-	range.offset = 0;
-	range.size = offset_;
-	range.memory = buffers_[buf_].deviceMemory;
-	vkFlushMappedMemoryRanges(device_, 1, &range);
-	*/
+
+	if ((memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) == 0) {
+		VkMappedMemoryRange range{ VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE };
+		range.offset = 0;
+		range.size = offset_;
+		range.memory = buffers_[buf_].deviceMemory;
+		vkFlushMappedMemoryRanges(vulkan_->GetDevice(), 1, &range);
+	}
+
 	vkUnmapMemory(vulkan_->GetDevice(), buffers_[buf_].deviceMemory);
 	writePtr_ = nullptr;
 }
diff --git a/Common/Vulkan/VulkanMemory.h b/Common/Vulkan/VulkanMemory.h
index 50e7017eb9..5c4cec8b1e 100644
--- a/Common/Vulkan/VulkanMemory.h
+++ b/Common/Vulkan/VulkanMemory.h
@@ -22,7 +22,10 @@ class VulkanPushBuffer {
 	};
 
 public:
-	VulkanPushBuffer(VulkanContext *vulkan, size_t size, VkBufferUsageFlags usage);
+	// NOTE: If you create a push buffer with only VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+	// then you can't use any of the push functions as pointers will not be reachable from the CPU.
+	// You must in this case use Allocate() only, and pass the returned offset and the VkBuffer to Vulkan APIs.
+	VulkanPushBuffer(VulkanContext *vulkan, size_t size, VkBufferUsageFlags usage, VkMemoryPropertyFlags memoryPropertyMask = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
 	~VulkanPushBuffer();
 
 	void Destroy(VulkanContext *vulkan);
@@ -35,15 +38,18 @@ public:
 		offset_ = 0;
 		// Note: we must defrag because some buffers may be smaller than size_.
 		Defragment(vulkan);
-		Map();
+		if (memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+			Map();
 	}
 
 	void BeginNoReset() {
-		Map();
+		if (memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+			Map();
 	}
 
 	void End() {
-		Unmap();
+		if (memoryPropertyMask_ & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+			Unmap();
 	}
 
 	void Map();
@@ -109,6 +115,8 @@ private:
 	void Defragment(VulkanContext *vulkan);
 
 	VulkanContext *vulkan_;
+	VkMemoryPropertyFlags memoryPropertyMask_;
+
 	std::vector<BufInfo> buffers_;
 	size_t buf_ = 0;
 	size_t offset_ = 0;
diff --git a/Core/Config.cpp b/Core/Config.cpp
index f7ce02c292..5590630ec3 100644
--- a/Core/Config.cpp
+++ b/Core/Config.cpp
@@ -751,6 +751,7 @@ static ConfigSetting graphicsSettings[] = {
 	ReportedConfigSetting("TexScalingLevel", &g_Config.iTexScalingLevel, 1, true, true),
 	ReportedConfigSetting("TexScalingType", &g_Config.iTexScalingType, 0, true, true),
 	ReportedConfigSetting("TexDeposterize", &g_Config.bTexDeposterize, false, true, true),
+	ReportedConfigSetting("TexHardwareScaling", &g_Config.bTexHardwareScaling, false, true, true),
 	ConfigSetting("VSyncInterval", &g_Config.bVSync, false, true, true),
 	ReportedConfigSetting("BloomHack", &g_Config.iBloomHack, 0, true, true),
 
diff --git a/Core/Config.h b/Core/Config.h
index f06f9acbdc..917bda6faa 100644
--- a/Core/Config.h
+++ b/Core/Config.h
@@ -170,6 +170,7 @@ public:
 	int iTexScalingLevel; // 0 = auto, 1 = off, 2 = 2x, ..., 5 = 5x
 	int iTexScalingType; // 0 = xBRZ, 1 = Hybrid
 	bool bTexDeposterize;
+	bool bTexHardwareScaling;
 	int iFpsLimit1;
 	int iFpsLimit2;
 	int iMaxRecent;
diff --git a/GPU/Vulkan/DrawEngineVulkan.cpp b/GPU/Vulkan/DrawEngineVulkan.cpp
index af4ee397a5..b5ca090dc4 100644
--- a/GPU/Vulkan/DrawEngineVulkan.cpp
+++ b/GPU/Vulkan/DrawEngineVulkan.cpp
@@ -153,6 +153,8 @@ void DrawEngineVulkan::InitDeviceObjects() {
 		frame_[i].pushUBO = new VulkanPushBuffer(vulkan_, 8 * 1024 * 1024, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT);
 		frame_[i].pushVertex = new VulkanPushBuffer(vulkan_, 2 * 1024 * 1024, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
 		frame_[i].pushIndex = new VulkanPushBuffer(vulkan_, 1 * 1024 * 1024, VK_BUFFER_USAGE_INDEX_BUFFER_BIT);
+
+		frame_[i].pushLocal = new VulkanPushBuffer(vulkan_, 1 * 1024 * 1024, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
 	}
 
 	VkPipelineLayoutCreateInfo pl{ VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO };
@@ -209,6 +211,11 @@ void DrawEngineVulkan::FrameData::Destroy(VulkanContext *vulkan) {
 		delete pushIndex;
 		pushIndex = nullptr;
 	}
+	if (pushLocal) {
+		pushLocal->Destroy(vulkan);
+		delete pushLocal;
+		pushLocal = nullptr;
+	}
 }
 
 void DrawEngineVulkan::DestroyDeviceObjects() {
@@ -264,10 +271,12 @@ void DrawEngineVulkan::BeginFrame() {
 	frame->pushUBO->Reset();
 	frame->pushVertex->Reset();
 	frame->pushIndex->Reset();
+	frame->pushLocal->Reset();
 
 	frame->pushUBO->Begin(vulkan_);
 	frame->pushVertex->Begin(vulkan_);
 	frame->pushIndex->Begin(vulkan_);
+	frame->pushLocal->Begin(vulkan_);
 
 	// TODO: How can we make this nicer...
 	tessDataTransferVulkan->SetPushBuffer(frame->pushUBO);
@@ -324,6 +333,7 @@ void DrawEngineVulkan::EndFrame() {
 	frame->pushUBO->End();
 	frame->pushVertex->End();
 	frame->pushIndex->End();
+	frame->pushLocal->End();
 	vertexCache_->End();
 }
 
diff --git a/GPU/Vulkan/DrawEngineVulkan.h b/GPU/Vulkan/DrawEngineVulkan.h
index cece2c08c6..cbbaeebb5b 100644
--- a/GPU/Vulkan/DrawEngineVulkan.h
+++ b/GPU/Vulkan/DrawEngineVulkan.h
@@ -188,6 +188,11 @@ public:
 		return frame_[vulkan_->GetCurFrame()].pushUBO;
 	}
 
+	// Only use Allocate on this one.
+	VulkanPushBuffer *GetPushBufferLocal() {
+		return frame_[vulkan_->GetCurFrame()].pushLocal;
+	}
+
 	const DrawEngineVulkanStats &GetStats() const {
 		return stats_;
 	}
@@ -257,6 +262,10 @@ private:
 		VulkanPushBuffer *pushUBO = nullptr;
 		VulkanPushBuffer *pushVertex = nullptr;
 		VulkanPushBuffer *pushIndex = nullptr;
+
+		// Special push buffer in GPU local memory, for texture data conversion and similar tasks.
+		VulkanPushBuffer *pushLocal;
+
 		// We do rolling allocation and reset instead of caching across frames. That we might do later.
 		DenseHashMap<DescriptorSetKey, VkDescriptorSet, (VkDescriptorSet)VK_NULL_HANDLE> descSets;
 
diff --git a/GPU/Vulkan/TextureCacheVulkan.cpp b/GPU/Vulkan/TextureCacheVulkan.cpp
index d092893d40..1de9ba2c9e 100644
--- a/GPU/Vulkan/TextureCacheVulkan.cpp
+++ b/GPU/Vulkan/TextureCacheVulkan.cpp
@@ -67,6 +67,450 @@ static const VkComponentMapping VULKAN_1555_SWIZZLE = { VK_COMPONENT_SWIZZLE_B,
 static const VkComponentMapping VULKAN_565_SWIZZLE = { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_G, VK_COMPONENT_SWIZZLE_B, VK_COMPONENT_SWIZZLE_A };
 static const VkComponentMapping VULKAN_8888_SWIZZLE = { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_G, VK_COMPONENT_SWIZZLE_B, VK_COMPONENT_SWIZZLE_A };
 
+// 4xBRZ shader - Copyright (C) 2014-2016 DeSmuME team (GPL2+)
+// Hyllian's xBR-vertex code and texel mapping
+// Copyright (C) 2011/2016 Hyllian - sergiogdb@gmail.com
+// TODO: Handles alpha badly for PSP.
+const char *shader4xbrz = R"(
+vec4 premultiply_alpha(vec4 c) { float a = clamp(c.a, 0.0, 1.0); return vec4(c.rgb * a, a); }
+vec4 postdivide_alpha(vec4 c) { return c.a < 0.001? vec4(0.0,0.0,0.0,0.0) : vec4(c.rgb / c.a, c.a); }
+
+#define BLEND_ALPHA 1
+#define BLEND_NONE 0
+#define BLEND_NORMAL 1
+#define BLEND_DOMINANT 2
+#define LUMINANCE_WEIGHT 1.0
+#define EQUAL_COLOR_TOLERANCE 30.0/255.0
+#define STEEP_DIRECTION_THRESHOLD 2.2
+#define DOMINANT_DIRECTION_THRESHOLD 3.6
+
+float reduce(vec4 color) {
+	return dot(color.rgb, vec3(65536.0, 256.0, 1.0));
+}
+
+float DistYCbCr(vec4 pixA, vec4 pixB) {
+	const vec3 w = vec3(0.2627, 0.6780, 0.0593);
+	const float scaleB = 0.5 / (1.0 - w.b);
+	const float scaleR = 0.5 / (1.0 - w.r);
+	vec4 diff = pixA - pixB;
+	float Y = dot(diff.rgb, w);
+	float Cb = scaleB * (diff.b - Y);
+	float Cr = scaleR * (diff.r - Y);
+
+	return sqrt( ((LUMINANCE_WEIGHT * Y) * (LUMINANCE_WEIGHT * Y)) + (Cb * Cb) + (Cr * Cr) + (diff.a * diff.a));
+}
+
+bool IsPixEqual(const vec4 pixA, const vec4 pixB) {
+	return (DistYCbCr(pixA, pixB) < EQUAL_COLOR_TOLERANCE);
+}
+
+bool IsBlendingNeeded(const ivec4 blend) {
+	ivec4 diff = blend - ivec4(BLEND_NONE);
+	return diff.x != 0 || diff.y != 0 || diff.z != 0 || diff.w != 0;
+}
+
+vec4 applyScalingf(uvec2 origxy, uvec2 xy) {
+	float dx = 1.0 / params.width;
+	float dy = 1.0 / params.height;
+
+	//    A1 B1 C1
+	// A0 A  B  C C4
+	// D0 D  E  F F4
+	// G0 G  H  I I4
+	//    G5 H5 I5
+
+	uvec4 t1 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 2); // A1 B1 C1
+	uvec4 t2 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 1); // A B C
+	uvec4 t3 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 0); // D E F
+	uvec4 t4 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 1); // G H I
+	uvec4 t5 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 2); // G5 H5 I5
+	uvec4 t6 = uvec4(origxy.x - 2, origxy.y - 1, origxy.y, origxy.y + 1); // A0 D0 G0
+	uvec4 t7 = uvec4(origxy.x + 2, origxy.y - 1, origxy.y, origxy.y + 1); // C4 F4 I4
+
+	vec2 f = fract(vec2(float(xy.x) / float(params.scale), float(xy.y) / float(params.scale)));
+
+	//---------------------------------------
+	// Input Pixel Mapping:    |21|22|23|
+	//                       19|06|07|08|09
+	//                       18|05|00|01|10
+	//                       17|04|03|02|11
+	//                         |15|14|13|
+
+	vec4 src[25];
+
+	src[21] = premultiply_alpha(readColorf(t1.xw));
+	src[22] = premultiply_alpha(readColorf(t1.yw));
+	src[23] = premultiply_alpha(readColorf(t1.zw));
+	src[ 6] = premultiply_alpha(readColorf(t2.xw));
+	src[ 7] = premultiply_alpha(readColorf(t2.yw));
+	src[ 8] = premultiply_alpha(readColorf(t2.zw));
+	src[ 5] = premultiply_alpha(readColorf(t3.xw));
+	src[ 0] = premultiply_alpha(readColorf(t3.yw));
+	src[ 1] = premultiply_alpha(readColorf(t3.zw));
+	src[ 4] = premultiply_alpha(readColorf(t4.xw));
+	src[ 3] = premultiply_alpha(readColorf(t4.yw));
+	src[ 2] = premultiply_alpha(readColorf(t4.zw));
+	src[15] = premultiply_alpha(readColorf(t5.xw));
+	src[14] = premultiply_alpha(readColorf(t5.yw));
+	src[13] = premultiply_alpha(readColorf(t5.zw));
+	src[19] = premultiply_alpha(readColorf(t6.xy));
+	src[18] = premultiply_alpha(readColorf(t6.xz));
+	src[17] = premultiply_alpha(readColorf(t6.xw));
+	src[ 9] = premultiply_alpha(readColorf(t7.xy));
+	src[10] = premultiply_alpha(readColorf(t7.xz));
+	src[11] = premultiply_alpha(readColorf(t7.xw));
+
+	float v[9];
+	v[0] = reduce(src[0]);
+	v[1] = reduce(src[1]);
+	v[2] = reduce(src[2]);
+	v[3] = reduce(src[3]);
+	v[4] = reduce(src[4]);
+	v[5] = reduce(src[5]);
+	v[6] = reduce(src[6]);
+	v[7] = reduce(src[7]);
+	v[8] = reduce(src[8]);
+
+	ivec4 blendResult = ivec4(BLEND_NONE);
+
+	// Preprocess corners
+	// Pixel Tap Mapping: --|--|--|--|--
+	//                    --|--|07|08|--
+	//                    --|05|00|01|10
+	//                    --|04|03|02|11
+	//                    --|--|14|13|--
+	// Corner (1, 1)
+	if ( ((v[0] == v[1] && v[3] == v[2]) || (v[0] == v[3] && v[1] == v[2])) == false) {
+		float dist_03_01 = DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + DistYCbCr(src[14], src[ 2]) + DistYCbCr(src[ 2], src[10]) + (4.0 * DistYCbCr(src[ 3], src[ 1]));
+		float dist_00_02 = DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[ 3], src[13]) + DistYCbCr(src[ 7], src[ 1]) + DistYCbCr(src[ 1], src[11]) + (4.0 * DistYCbCr(src[ 0], src[ 2]));
+		bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_03_01) < dist_00_02;
+		blendResult[2] = ((dist_03_01 < dist_00_02) && (v[0] != v[1]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
+	}
+
+	// Pixel Tap Mapping: --|--|--|--|--
+	//                    --|06|07|--|--
+	//                    18|05|00|01|--
+	//                    17|04|03|02|--
+	//                    --|15|14|--|--
+	// Corner (0, 1)
+	if ( ((v[5] == v[0] && v[4] == v[3]) || (v[5] == v[4] && v[0] == v[3])) == false) {
+		float dist_04_00 = DistYCbCr(src[17], src[ 5]) + DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[15], src[ 3]) + DistYCbCr(src[ 3], src[ 1]) + (4.0 * DistYCbCr(src[ 4], src[ 0]));
+		float dist_05_03 = DistYCbCr(src[18], src[ 4]) + DistYCbCr(src[ 4], src[14]) + DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + (4.0 * DistYCbCr(src[ 5], src[ 3]));
+		bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_03) < dist_04_00;
+		blendResult[3] = ((dist_04_00 > dist_05_03) && (v[0] != v[5]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
+	}
+
+	// Pixel Tap Mapping: --|--|22|23|--
+	//                    --|06|07|08|09
+	//                    --|05|00|01|10
+	//                    --|--|03|02|--
+	//                    --|--|--|--|--
+	// Corner (1, 0)
+	if ( ((v[7] == v[8] && v[0] == v[1]) || (v[7] == v[0] && v[8] == v[1])) == false) {
+		float dist_00_08 = DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[ 7], src[23]) + DistYCbCr(src[ 3], src[ 1]) + DistYCbCr(src[ 1], src[ 9]) + (4.0 * DistYCbCr(src[ 0], src[ 8]));
+		float dist_07_01 = DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + DistYCbCr(src[22], src[ 8]) + DistYCbCr(src[ 8], src[10]) + (4.0 * DistYCbCr(src[ 7], src[ 1]));
+		bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_07_01) < dist_00_08;
+		blendResult[1] = ((dist_00_08 > dist_07_01) && (v[0] != v[7]) && (v[0] != v[1])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
+	}
+
+	// Pixel Tap Mapping: --|21|22|--|--
+	//                    19|06|07|08|--
+	//                    18|05|00|01|--
+	//                    --|04|03|--|--
+	//                    --|--|--|--|--
+	// Corner (0, 0)
+	if ( ((v[6] == v[7] && v[5] == v[0]) || (v[6] == v[5] && v[7] == v[0])) == false) {
+		float dist_05_07 = DistYCbCr(src[18], src[ 6]) + DistYCbCr(src[ 6], src[22]) + DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + (4.0 * DistYCbCr(src[ 5], src[ 7]));
+		float dist_06_00 = DistYCbCr(src[19], src[ 5]) + DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[21], src[ 7]) + DistYCbCr(src[ 7], src[ 1]) + (4.0 * DistYCbCr(src[ 6], src[ 0]));
+		bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_07) < dist_06_00;
+		blendResult[0] = ((dist_05_07 < dist_06_00) && (v[0] != v[5]) && (v[0] != v[7])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
+	}
+
+	vec4 dst[16];
+	dst[ 0] = src[0];
+	dst[ 1] = src[0];
+	dst[ 2] = src[0];
+	dst[ 3] = src[0];
+	dst[ 4] = src[0];
+	dst[ 5] = src[0];
+	dst[ 6] = src[0];
+	dst[ 7] = src[0];
+	dst[ 8] = src[0];
+	dst[ 9] = src[0];
+	dst[10] = src[0];
+	dst[11] = src[0];
+	dst[12] = src[0];
+	dst[13] = src[0];
+	dst[14] = src[0];
+	dst[15] = src[0];
+
+	// Scale pixel
+	if (IsBlendingNeeded(blendResult) == true) {
+		float dist_01_04 = DistYCbCr(src[1], src[4]);
+		float dist_03_08 = DistYCbCr(src[3], src[8]);
+		bool haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[4]) && (v[5] != v[4]);
+		bool haveSteepLine   = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[8]) && (v[7] != v[8]);
+		bool needBlend = (blendResult[2] != BLEND_NONE);
+		bool doLineBlend = (  blendResult[2] >= BLEND_DOMINANT ||
+			((blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
+			(blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
+			(IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[0], src[2]) == false) ) == false );
+
+		vec4 blendPix = ( DistYCbCr(src[0], src[1]) <= DistYCbCr(src[0], src[3]) ) ? src[1] : src[3];
+		dst[ 2] = mix(dst[ 2], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
+		dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
+		dst[10] = mix(dst[10], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
+		dst[11] = mix(dst[11], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[12] = mix(dst[12], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
+		dst[13] = mix(dst[13], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[14] = mix(dst[14], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
+		dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
+
+		dist_01_04 = DistYCbCr(src[7], src[2]);
+		dist_03_08 = DistYCbCr(src[1], src[6]);
+		haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[2]) && (v[3] != v[2]);
+		haveSteepLine   = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[6]) && (v[5] != v[6]);
+		needBlend = (blendResult[1] != BLEND_NONE);
+		doLineBlend = (  blendResult[1] >= BLEND_DOMINANT ||
+			!((blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
+			(blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
+			(IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && !IsPixEqual(src[0], src[8])) ) );
+
+		blendPix = ( DistYCbCr(src[0], src[7]) <= DistYCbCr(src[0], src[1]) ) ? src[7] : src[1];
+		dst[ 1] = mix(dst[ 1], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
+		dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
+		dst[ 7] = mix(dst[ 7], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
+		dst[ 8] = mix(dst[ 8], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[ 9] = mix(dst[ 9], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
+		dst[10] = mix(dst[10], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[11] = mix(dst[11], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
+		dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
+
+		dist_01_04 = DistYCbCr(src[5], src[8]);
+		dist_03_08 = DistYCbCr(src[7], src[4]);
+		haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[8]) && (v[1] != v[8]);
+		haveSteepLine   = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[4]) && (v[3] != v[4]);
+		needBlend = (blendResult[0] != BLEND_NONE);
+		doLineBlend = (  blendResult[0] >= BLEND_DOMINANT ||
+			!((blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
+			(blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
+			(IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && !IsPixEqual(src[0], src[6])) ) );
+
+		blendPix = ( DistYCbCr(src[0], src[5]) <= DistYCbCr(src[0], src[7]) ) ? src[5] : src[7];
+		dst[ 0] = mix(dst[ 0], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
+		dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
+		dst[ 4] = mix(dst[ 4], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
+		dst[ 5] = mix(dst[ 5], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[ 6] = mix(dst[ 6], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
+		dst[ 7] = mix(dst[ 7], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[ 8] = mix(dst[ 8], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
+		dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
+
+		dist_01_04 = DistYCbCr(src[3], src[6]);
+		dist_03_08 = DistYCbCr(src[5], src[2]);
+		haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[6]) && (v[7] != v[6]);
+		haveSteepLine   = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[2]) && (v[1] != v[2]);
+		needBlend = (blendResult[3] != BLEND_NONE);
+		doLineBlend = (  blendResult[3] >= BLEND_DOMINANT ||
+			!((blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
+			(blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
+			(IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && !IsPixEqual(src[0], src[4])) ) );
+
+		blendPix = ( DistYCbCr(src[0], src[3]) <= DistYCbCr(src[0], src[5]) ) ? src[3] : src[5];
+		dst[ 3] = mix(dst[ 3], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
+		dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
+		dst[13] = mix(dst[13], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
+		dst[14] = mix(dst[14], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[15] = mix(dst[15], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
+		dst[ 4] = mix(dst[ 4], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
+		dst[ 5] = mix(dst[ 5], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
+		dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
+	}
+
+	// select output pixel
+	vec4 res = mix(mix(mix(mix(dst[ 6], dst[ 7], step(0.25, f.x)),
+					mix(dst[ 8], dst[ 9], step(0.75, f.x)),
+					step(0.50, f.x)),
+				mix(mix(dst[ 5], dst[ 0], step(0.25, f.x)),
+					mix(dst[ 1], dst[10], step(0.75, f.x)),
+					step(0.50, f.x)),
+				step(0.25, f.y)),
+		mix(mix(mix(dst[ 4], dst[ 3], step(0.25, f.x)),
+					mix(dst[ 2], dst[11], step(0.75, f.x)),
+					step(0.50, f.x)),
+				mix(mix(dst[15], dst[14], step(0.25, f.x)),
+					mix(dst[13], dst[12], step(0.75, f.x)),
+					step(0.50, f.x)),
+				step(0.75, f.y)),
+		step(0.50, f.y));
+
+	return postdivide_alpha(res);
+}
+
+uint applyScalingu(uvec2 origxy, uvec2 xy) {
+	return packUnorm4x8(applyScalingf(origxy, xy));
+}
+)";
+
+const char *copyShader = R"(
+#version 450
+#extension GL_ARB_separate_shader_objects : enable
+
+// No idea what's optimal here...
+#define WORKGROUP_SIZE 16
+layout (local_size_x = WORKGROUP_SIZE, local_size_y = WORKGROUP_SIZE, local_size_z = 1) in;
+
+layout(std430, binding = 1) buffer Buf1 {
+	uint data[];
+} buf1;
+
+layout(std430, binding = 2) buffer Buf2 {
+	uint data[];
+} buf2;
+
+layout(push_constant) uniform Params {
+	int width;
+	int height;
+	int scale;
+	int fmt;
+} params;
+
+uint readColoru(uvec2 p) {
+	// Note that if the pixels are packed, we can do multiple stores
+	// and only launch this compute shader for every N pixels,
+	// by slicing the width in half and multiplying x by 2, for example.
+	if (params.fmt == 0) {
+		return buf1.data[p.y * params.width + p.x];
+	} else {
+		uint offset = p.y * params.width + p.x;
+		uint data = buf1.data[offset / 2];
+		if ((offset & 1) != 0) {
+			data = data >> 16;
+		}
+		if (params.fmt == 6) {
+			uint r = ((data << 3) & 0xF8) | ((data >> 2) & 0x07);
+			uint g = ((data >> 3) & 0xFC) | ((data >> 9) & 0x03);
+			uint b = ((data >> 8) & 0xF8) | ((data >> 13) & 0x07);
+			return 0xFF000000 | (b << 16) | (g << 8) | r;
+		} else if (params.fmt == 5) {
+			uint r = ((data << 3) & 0xF8) | ((data >> 2) & 0x07);
+			uint g = ((data >> 2) & 0xF8) | ((data >> 7) & 0x07);
+			uint b = ((data >> 7) & 0xF8) | ((data >> 12) & 0x07);
+			uint a = ((data >> 15) & 0x01) == 0 ? 0x00 : 0xFF;
+			return (a << 24) | (b << 16) | (g << 8) | r;
+		} else if (params.fmt == 4) {
+			uint r = (data & 0x0F) | ((data << 4) & 0x0F);
+			uint g = (data & 0xF0) | ((data >> 4) & 0x0F);
+			uint b = ((data >> 8) & 0x0F) | ((data >> 4) & 0xF0);
+			uint a = ((data >> 12) & 0x0F) | ((data >> 8) & 0xF0);
+			return (a << 24) | (b << 16) | (g << 8) | r;
+		}
+	}
+}
+
+vec4 readColorf(uvec2 p) {
+	return unpackUnorm4x8(readColoru(p));
+}
+
+%s
+
+void main() {
+	uvec2 xy = gl_GlobalInvocationID.xy;
+	// Kill off any out-of-image threads to avoid stray writes.
+	// Should only happen on the tiniest mipmaps as PSP textures are power-of-2,
+	// and we use a 16x16 workgroup size.
+	if (xy.x >= params.width || xy.y >= params.height)
+		return;
+
+	uvec2 origxy = xy / params.scale;
+	if (params.scale == 1) {
+		buf2.data[xy.y * params.width + xy.x] = readColoru(origxy);
+	} else {
+		buf2.data[xy.y * params.width + xy.x] = applyScalingu(origxy, xy);
+	}
+}
+)";
+
+const char *uploadShader = R"(
+#version 450
+#extension GL_ARB_separate_shader_objects : enable
+
+// No idea what's optimal here...
+#define WORKGROUP_SIZE 16
+layout (local_size_x = WORKGROUP_SIZE, local_size_y = WORKGROUP_SIZE, local_size_z = 1) in;
+
+uniform layout(binding = 0, rgba8) writeonly image2D img;
+
+layout(std430, binding = 1) buffer Buf {
+	uint data[];
+} buf;
+
+layout(push_constant) uniform Params {
+	int width;
+	int height;
+	int scale;
+	int fmt;
+} params;
+
+uint readColoru(uvec2 p) {
+	// Note that if the pixels are packed, we can do multiple stores
+	// and only launch this compute shader for every N pixels,
+	// by slicing the width in half and multiplying x by 2, for example.
+	if (params.fmt == 0) {
+		return buf.data[p.y * params.width + p.x];
+	} else {
+		uint offset = p.y * params.width + p.x;
+		uint data = buf.data[offset / 2];
+		if ((offset & 1) != 0) {
+			data = data >> 16;
+		}
+		if (params.fmt == 6) {
+			uint r = ((data << 3) & 0xF8) | ((data >> 2) & 0x07);
+			uint g = ((data >> 3) & 0xFC) | ((data >> 9) & 0x03);
+			uint b = ((data >> 8) & 0xF8) | ((data >> 13) & 0x07);
+			return 0xFF000000 | (b << 16) | (g << 8) | r;
+		} else if (params.fmt == 5) {
+			uint r = ((data << 3) & 0xF8) | ((data >> 2) & 0x07);
+			uint g = ((data >> 2) & 0xF8) | ((data >> 7) & 0x07);
+			uint b = ((data >> 7) & 0xF8) | ((data >> 12) & 0x07);
+			uint a = ((data >> 15) & 0x01) == 0 ? 0x00 : 0xFF;
+			return (a << 24) | (b << 16) | (g << 8) | r;
+		} else if (params.fmt == 4) {
+			uint r = (data & 0x0F) | ((data << 4) & 0x0F);
+			uint g = (data & 0xF0) | ((data >> 4) & 0x0F);
+			uint b = ((data >> 8) & 0x0F) | ((data >> 4) & 0xF0);
+			uint a = ((data >> 12) & 0x0F) | ((data >> 8) & 0xF0);
+			return (a << 24) | (b << 16) | (g << 8) | r;
+		}
+	}
+}
+
+vec4 readColorf(uvec2 p) {
+	// Unpack the color (we could look it up in a CLUT here if we wanted...)
+	// It's a bit silly that we need to unpack to float and then have imageStore repack,
+	// but the alternative is to store to a buffer, and then launch a vkCmdCopyBufferToImage instead.
+	return unpackUnorm4x8(readColoru(p));
+}
+
+%s
+
+void main() {
+	uvec2 xy = gl_GlobalInvocationID.xy;
+	// Kill off any out-of-image threads to avoid stray writes.
+	// Should only happen on the tiniest mipmaps as PSP textures are power-of-2,
+	// and we use a 16x16 workgroup size.
+	if (xy.x >= params.width || xy.y >= params.height)
+		return;
+
+	uvec2 origxy = xy / params.scale;
+	if (params.scale == 1) {
+		imageStore(img, ivec2(xy.x, xy.y), readColorf(origxy));
+	} else {
+		imageStore(img, ivec2(xy.x, xy.y), applyScalingf(origxy, xy));
+	}
+}
+)";
+
 SamplerCache::~SamplerCache() {
 	DeviceLost();
 }
@@ -141,7 +585,8 @@ std::vector<std::string> SamplerCache::DebugGetSamplerIDs() const {
 TextureCacheVulkan::TextureCacheVulkan(Draw::DrawContext *draw, VulkanContext *vulkan)
 	: TextureCacheCommon(draw),
 		vulkan_(vulkan),
-		samplerCache_(vulkan) {
+		samplerCache_(vulkan),
+		computeShaderManager_(vulkan) {
 	timesInvalidatedAllThisFrame_ = 0;
 	DeviceRestore(vulkan, draw);
 	SetupTextureDecoder();
@@ -180,6 +625,13 @@ void TextureCacheVulkan::DeviceLost() {
 	if (samplerNearest_)
 		vulkan_->Delete().QueueDeleteSampler(samplerNearest_);
 
+	if (uploadCS_ != VK_NULL_HANDLE)
+		vulkan_->Delete().QueueDeleteShaderModule(uploadCS_);
+	if (copyCS_ != VK_NULL_HANDLE)
+		vulkan_->Delete().QueueDeleteShaderModule(copyCS_);
+
+	computeShaderManager_.DeviceLost();
+
 	nextTexture_ = nullptr;
 }
 
@@ -200,6 +652,19 @@ void TextureCacheVulkan::DeviceRestore(VulkanContext *vulkan, Draw::DrawContext
 	samp.minFilter = VK_FILTER_NEAREST;
 	samp.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
 	vkCreateSampler(vulkan_->GetDevice(), &samp, nullptr, &samplerNearest_);
+
+	std::string error;
+	std::string fullUploadShader = StringFromFormat(uploadShader, shader4xbrz);
+	std::string fullCopyShader = StringFromFormat(copyShader, shader4xbrz);
+
+	if (g_Config.bTexHardwareScaling) {
+		uploadCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullUploadShader.c_str(), &error);
+		_dbg_assert_msg_(G3D, uploadCS_ != VK_NULL_HANDLE, "failed to compile upload shader");
+		copyCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullCopyShader.c_str(), &error);
+		_dbg_assert_msg_(G3D, copyCS_!= VK_NULL_HANDLE, "failed to compile copy shader");
+	}
+
+	computeShaderManager_.DeviceRestore(vulkan);
 }
 
 void TextureCacheVulkan::ReleaseTexture(TexCacheEntry *entry, bool delete_them) {
@@ -272,10 +737,12 @@ void TextureCacheVulkan::StartFrame() {
 	}
 
 	allocator_->Begin();
+	computeShaderManager_.BeginFrame();
 }
 
 void TextureCacheVulkan::EndFrame() {
 	allocator_->End();
+	computeShaderManager_.EndFrame();
 
 	if (texelsScaledThisFrame_) {
 		// INFO_LOG(G3D, "Scaled %i texels", texelsScaledThisFrame_);
@@ -570,14 +1037,14 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 	// Don't scale the PPGe texture.
 	if (entry->addr > 0x05000000 && entry->addr < PSP_GetKernelMemoryEnd())
 		scaleFactor = 1;
-	if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1) {
+	if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1 && !g_Config.bTexHardwareScaling) {
 		// 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) {
+		if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED && !g_Config.bTexHardwareScaling) {
 			entry->status |= TexCacheEntry::STATUS_TO_SCALE;
 			scaleFactor = 1;
 		} else {
@@ -597,6 +1064,9 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		actualFmt = ToVulkanFormat(replaced.Format(0));
 	}
 
+	bool computeUpload = false;
+	bool computeCopy = false;
+
 	{
 		delete entry->vkTex;
 		entry->vkTex = new VulkanTexture(vulkan_);
@@ -621,11 +1091,29 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 			break;
 		}
 
+		VkImageLayout imageLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+		VkImageUsageFlags usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
+		// If we want to use the GE debugger, we should add VK_IMAGE_USAGE_TRANSFER_SRC_BIT too...
+
+		// Compute experiment
+		if (actualFmt == VULKAN_8888_FORMAT && scaleFactor > 1 && g_Config.bTexHardwareScaling) {
+			// Enable the experiment you want.
+			if (uploadCS_ != VK_NULL_HANDLE)
+				computeUpload = true;
+			else if (copyCS_ != VK_NULL_HANDLE)
+				computeCopy = true;
+		}
+
+		if (computeUpload) {
+			usage |= VK_IMAGE_USAGE_STORAGE_BIT;
+			imageLayout = VK_IMAGE_LAYOUT_GENERAL;
+		}
+
 		char texName[128]{};
 		snprintf(texName, sizeof(texName), "Texture%08x", entry->addr);
 		image->SetTag(texName);
 
-		bool allocSuccess = image->CreateDirect(cmdInit, allocator_, w * scaleFactor, h * scaleFactor, maxLevel + 1, actualFmt, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, mapping);
+		bool allocSuccess = image->CreateDirect(cmdInit, allocator_, w * scaleFactor, h * scaleFactor, maxLevel + 1, actualFmt, imageLayout, usage, mapping);
 		if (!allocSuccess && !lowMemoryMode_) {
 			WARN_LOG_REPORT(G3D, "Texture cache ran out of GPU memory; switching to low memory mode");
 			lowMemoryMode_ = true;
@@ -677,6 +1165,9 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 			if (replaced.Valid()) {
 				replaced.GetSize(i, mipWidth, mipHeight);
 			}
+			int srcBpp = dstFmt == VULKAN_8888_FORMAT ? 4 : 2;
+			int srcStride = mipWidth * srcBpp;
+			int srcSize = srcStride * mipHeight;
 			int bpp = actualFmt == VULKAN_8888_FORMAT ? 4 : 2;
 			int stride = (mipWidth * bpp + 15) & ~15;
 			int size = stride * mipHeight;
@@ -684,22 +1175,85 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 			VkBuffer texBuf;
 			// nvidia returns 1 but that can't be healthy... let's align by 16 as a minimum.
 			int pushAlignment = std::max(16, (int)vulkan_->GetPhysicalDeviceProperties().properties.limits.optimalBufferCopyOffsetAlignment);
-			void *data = drawEngine_->GetPushBufferForTextureData()->PushAligned(size, &bufferOffset, &texBuf, pushAlignment);
+			void *data;
+			bool dataScaled = true;
 			if (replaced.Valid()) {
+				data = drawEngine_->GetPushBufferForTextureData()->PushAligned(size, &bufferOffset, &texBuf, pushAlignment);
 				replaced.Load(i, data, stride);
+				entry->vkTex->UploadMip(cmdInit, i, mipWidth, mipHeight, texBuf, bufferOffset, stride / bpp);
 			} else {
+				auto dispatchCompute = [&](VkDescriptorSet descSet) {
+					struct Params { int x; int y; int s; int fmt; } params{ mipWidth, mipHeight, scaleFactor, 0 };
+					if (dstFmt == VULKAN_4444_FORMAT) {
+						params.fmt = 4;
+					} else if (dstFmt == VULKAN_1555_FORMAT) {
+						params.fmt = 5;
+					} else if (dstFmt == VULKAN_565_FORMAT) {
+						params.fmt = 6;
+					}
+					vkCmdBindDescriptorSets(cmdInit, VK_PIPELINE_BIND_POINT_COMPUTE, computeShaderManager_.GetPipelineLayout(), 0, 1, &descSet, 0, nullptr);
+					vkCmdPushConstants(cmdInit, computeShaderManager_.GetPipelineLayout(), VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(params), &params);
+					vkCmdDispatch(cmdInit, (mipWidth + 15) / 16, (mipHeight + 15) / 16, 1);
+				};
+
 				if (fakeMipmap) {
+					data = drawEngine_->GetPushBufferForTextureData()->PushAligned(size, &bufferOffset, &texBuf, pushAlignment);
 					LoadTextureLevel(*entry, (uint8_t *)data, stride, level, scaleFactor, dstFmt);
 					entry->vkTex->UploadMip(cmdInit, 0, mipWidth, mipHeight, texBuf, bufferOffset, stride / bpp);
 					break;
 				} else {
-					LoadTextureLevel(*entry, (uint8_t *)data, stride, i, scaleFactor, dstFmt);
+					if (computeUpload) {
+						data = drawEngine_->GetPushBufferForTextureData()->PushAligned(srcSize, &bufferOffset, &texBuf, pushAlignment);
+						dataScaled = false;
+						LoadTextureLevel(*entry, (uint8_t *)data, srcStride, i, 1, dstFmt);
+						// This format can be used with storage images.
+						VkImageView view = entry->vkTex->CreateViewForMip(i);
+						VkDescriptorSet descSet = computeShaderManager_.GetDescriptorSet(view, texBuf, bufferOffset, srcSize);
+						vkCmdBindPipeline(cmdInit, VK_PIPELINE_BIND_POINT_COMPUTE, computeShaderManager_.GetPipeline(uploadCS_));
+						dispatchCompute(descSet);
+						vulkan_->Delete().QueueDeleteImageView(view);
+					} else if (computeCopy) {
+						data = drawEngine_->GetPushBufferForTextureData()->PushAligned(srcSize, &bufferOffset, &texBuf, pushAlignment);
+						dataScaled = false;
+						LoadTextureLevel(*entry, (uint8_t *)data, srcStride, i, 1, dstFmt);
+						// Simple test of using a "copy shader" before the upload. This one could unswizzle or whatever
+						// and will work for any texture format including 16-bit as long as the shader is written to pack it into int32 size bits
+						// which is the smallest possible write.
+						VkBuffer localBuf;
+						uint32_t localOffset;
+						uint32_t localSize = size;
+						localOffset = (uint32_t)drawEngine_->GetPushBufferLocal()->Allocate(localSize, &localBuf);
+
+						VkDescriptorSet descSet = computeShaderManager_.GetDescriptorSet(VK_NULL_HANDLE, texBuf, bufferOffset, srcSize, localBuf, localOffset, localSize);
+						vkCmdBindPipeline(cmdInit, VK_PIPELINE_BIND_POINT_COMPUTE, computeShaderManager_.GetPipeline(copyCS_));
+						dispatchCompute(descSet);
+
+						// After the compute, before the copy, we need a memory barrier.
+						VkBufferMemoryBarrier barrier{ VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER };
+						barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
+						barrier.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
+						barrier.buffer = localBuf;
+						barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+						barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+						barrier.offset = localOffset;
+						barrier.size = localSize;
+						vkCmdPipelineBarrier(cmdInit, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+							0, 0, nullptr, 1, &barrier, 0, nullptr);
+
+						entry->vkTex->UploadMip(cmdInit, i, mipWidth, mipHeight, localBuf, localOffset, stride / bpp);
+					} else {
+						data = drawEngine_->GetPushBufferForTextureData()->PushAligned(size, &bufferOffset, &texBuf, pushAlignment);
+						LoadTextureLevel(*entry, (uint8_t *)data, stride, i, scaleFactor, dstFmt);
+						entry->vkTex->UploadMip(cmdInit, i, mipWidth, mipHeight, texBuf, bufferOffset, stride / bpp);
+					}
 				}
 				if (replacer_.Enabled()) {
-					replacer_.NotifyTextureDecoded(replacedInfo, data, stride, i, mipWidth, mipHeight);
+					// When hardware texture scaling is enabled, this saves the original.
+					int w = dataScaled ? mipWidth : mipWidth / scaleFactor;
+					int h = dataScaled ? mipHeight : mipHeight / scaleFactor;
+					replacer_.NotifyTextureDecoded(replacedInfo, data, stride, i, w, h);
 				}
 			}
-			entry->vkTex->UploadMip(cmdInit, i, mipWidth, mipHeight, texBuf, bufferOffset, stride / bpp);
 		}
 
 		if (maxLevel == 0) {
@@ -710,7 +1264,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		if (replaced.Valid()) {
 			entry->SetAlphaStatus(TexCacheEntry::TexStatus(replaced.AlphaStatus()));
 		}
-		entry->vkTex->EndCreate(cmdInit);
+		entry->vkTex->EndCreate(cmdInit, false, computeUpload ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
 	}
 
 	gstate_c.SetTextureFullAlpha(entry->GetAlphaStatus() == TexCacheEntry::STATUS_ALPHA_FULL);
diff --git a/GPU/Vulkan/TextureCacheVulkan.h b/GPU/Vulkan/TextureCacheVulkan.h
index 77f55aaea8..3635d5299c 100644
--- a/GPU/Vulkan/TextureCacheVulkan.h
+++ b/GPU/Vulkan/TextureCacheVulkan.h
@@ -25,6 +25,7 @@
 #include "Common/Vulkan/VulkanContext.h"
 #include "GPU/Vulkan/TextureScalerVulkan.h"
 #include "GPU/Common/TextureCacheCommon.h"
+#include "GPU/Vulkan/VulkanUtil.h"
 
 struct VirtualFramebuffer;
 class FramebufferManagerVulkan;
@@ -126,6 +127,8 @@ private:
 	VulkanDeviceAllocator *allocator_ = nullptr;
 	VulkanPushBuffer *push_ = nullptr;
 
+	VulkanComputeShaderManager computeShaderManager_;
+
 	SamplerCache samplerCache_;
 
 	TextureScalerVulkan scaler;
@@ -142,6 +145,9 @@ private:
 	DrawEngineVulkan *drawEngine_;
 	Vulkan2D *vulkan2D_;
 
+	VkShaderModule uploadCS_ = VK_NULL_HANDLE;
+	VkShaderModule copyCS_ = VK_NULL_HANDLE;
+
 	// Bound state to emulate an API similar to the others
 	VkImageView imageView_ = VK_NULL_HANDLE;
 	VkSampler curSampler_ = VK_NULL_HANDLE;
diff --git a/GPU/Vulkan/VulkanUtil.cpp b/GPU/Vulkan/VulkanUtil.cpp
index f598086235..ad91ec05c2 100644
--- a/GPU/Vulkan/VulkanUtil.cpp
+++ b/GPU/Vulkan/VulkanUtil.cpp
@@ -16,6 +16,7 @@
 // https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
 
 #include "base/basictypes.h"
+#include "base/stringutil.h"
 #include "Common/Log.h"
 #include "Common/Vulkan/VulkanContext.h"
 #include "GPU/Vulkan/VulkanUtil.h"
@@ -51,17 +52,14 @@ void Vulkan2D::DestroyDeviceObjects() {
 	VkDevice device = vulkan_->GetDevice();
 	if (descriptorSetLayout_ != VK_NULL_HANDLE) {
 		vulkan_->Delete().QueueDeleteDescriptorSetLayout(descriptorSetLayout_);
-		descriptorSetLayout_ = VK_NULL_HANDLE;
 	}
 	if (pipelineLayout_ != VK_NULL_HANDLE) {
 		vulkan_->Delete().QueueDeletePipelineLayout(pipelineLayout_);
-		pipelineLayout_ = VK_NULL_HANDLE;
 	}
 
 	// pipelineBasicTex_ and pipelineBasicTex_ come from vulkan2D_.
 	if (pipelineCache_ != VK_NULL_HANDLE) {
 		vulkan_->Delete().QueueDeletePipelineCache(pipelineCache_);
-		pipelineCache_ = VK_NULL_HANDLE;
 	}
 }
 
@@ -388,7 +386,7 @@ VkShaderModule CompileShaderModule(VulkanContext *vulkan, VkShaderStageFlagBits
 			ERROR_LOG(G3D, "Error in shader compilation!");
 		}
 		ERROR_LOG(G3D, "Messages: %s", error->c_str());
-		ERROR_LOG(G3D, "Shader source:\n%s", code);
+		ERROR_LOG(G3D, "Shader source:\n%s", LineNumberString(code).c_str());
 		OutputDebugStringUTF8("Messages:\n");
 		OutputDebugStringUTF8(error->c_str());
 		return VK_NULL_HANDLE;
@@ -401,3 +399,169 @@ VkShaderModule CompileShaderModule(VulkanContext *vulkan, VkShaderStageFlagBits
 		}
 	}
 }
+
+VulkanComputeShaderManager::VulkanComputeShaderManager(VulkanContext *vulkan) : vulkan_(vulkan), pipelines_(8) {
+}
+VulkanComputeShaderManager::~VulkanComputeShaderManager() {}
+
+void VulkanComputeShaderManager::InitDeviceObjects() {
+	VkPipelineCacheCreateInfo pc{ VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO };
+	VkResult res = vkCreatePipelineCache(vulkan_->GetDevice(), &pc, nullptr, &pipelineCache_);
+	assert(VK_SUCCESS == res);
+
+	VkDescriptorSetLayoutBinding bindings[3] = {};
+	bindings[0].descriptorCount = 1;
+	bindings[0].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
+	bindings[0].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
+	bindings[0].binding = 0;
+	bindings[1].descriptorCount = 1;
+	bindings[1].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+	bindings[1].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
+	bindings[1].binding = 1;
+	bindings[2].descriptorCount = 1;
+	bindings[2].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+	bindings[2].stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
+	bindings[2].binding = 2;
+
+	VkDevice device = vulkan_->GetDevice();
+
+	VkDescriptorSetLayoutCreateInfo dsl = { VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO };
+	dsl.bindingCount = ARRAY_SIZE(bindings);
+	dsl.pBindings = bindings;
+	res = vkCreateDescriptorSetLayout(device, &dsl, nullptr, &descriptorSetLayout_);
+	assert(VK_SUCCESS == res);
+
+	VkDescriptorPoolSize dpTypes[2];
+	dpTypes[0].descriptorCount = 8192;
+	dpTypes[0].type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+	dpTypes[1].descriptorCount = 4096;
+	dpTypes[1].type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
+
+	VkDescriptorPoolCreateInfo dp = { VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO };
+	dp.flags = 0;   // Don't want to mess around with individually freeing these, let's go fixed each frame and zap the whole array. Might try the dynamic approach later.
+	dp.maxSets = 4096;  // GTA can end up creating more than 1000 textures in the first frame!
+	dp.pPoolSizes = dpTypes;
+	dp.poolSizeCount = ARRAY_SIZE(dpTypes);
+	for (int i = 0; i < ARRAY_SIZE(frameData_); i++) {
+		VkResult res = vkCreateDescriptorPool(vulkan_->GetDevice(), &dp, nullptr, &frameData_[i].descPool);
+		assert(VK_SUCCESS == res);
+	}
+
+	VkPushConstantRange push = {};
+	push.offset = 0;
+	push.size = 16;
+	push.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
+
+	VkPipelineLayoutCreateInfo pl = { VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO };
+	pl.pPushConstantRanges = &push;
+	pl.pushConstantRangeCount = 1;
+	pl.setLayoutCount = 1;
+	pl.pSetLayouts = &descriptorSetLayout_;
+	pl.flags = 0;
+	res = vkCreatePipelineLayout(device, &pl, nullptr, &pipelineLayout_);
+	assert(VK_SUCCESS == res);
+}
+
+void VulkanComputeShaderManager::DestroyDeviceObjects() {
+	for (int i = 0; i < ARRAY_SIZE(frameData_); i++) {
+		vulkan_->Delete().QueueDeleteDescriptorPool(frameData_[i].descPool);
+	}
+	if (descriptorSetLayout_) {
+		vulkan_->Delete().QueueDeleteDescriptorSetLayout(descriptorSetLayout_);
+	}
+	pipelines_.Iterate([&](const PipelineKey &key, VkPipeline pipeline) {
+		vulkan_->Delete().QueueDeletePipeline(pipeline);
+	});
+	pipelines_.Clear();
+
+	if (pipelineLayout_) {
+		vulkan_->Delete().QueueDeletePipelineLayout(pipelineLayout_);
+	}
+	if (pipelineCache_ != VK_NULL_HANDLE) {
+		vulkan_->Delete().QueueDeletePipelineCache(pipelineCache_);
+	}
+}
+
+VkDescriptorSet VulkanComputeShaderManager::GetDescriptorSet(VkImageView image, VkBuffer buffer, VkDeviceSize offset, VkDeviceSize range, VkBuffer buffer2, VkDeviceSize offset2, VkDeviceSize range2) {
+	int curFrame = vulkan_->GetCurFrame();
+	FrameData &frameData = frameData_[curFrame];
+	frameData_[curFrame].numDescriptors++;
+	VkDescriptorSet desc;
+	VkDescriptorSetAllocateInfo descAlloc = { VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO };
+	descAlloc.pSetLayouts = &descriptorSetLayout_;
+	descAlloc.descriptorPool = frameData.descPool;
+	descAlloc.descriptorSetCount = 1;
+	VkResult result = vkAllocateDescriptorSets(vulkan_->GetDevice(), &descAlloc, &desc);
+	assert(result == VK_SUCCESS);
+
+	VkWriteDescriptorSet writes[2]{};
+	int n = 0;
+	VkDescriptorImageInfo imageInfo = {};
+	VkDescriptorBufferInfo bufferInfo[2] = {};
+	if (image) {
+		imageInfo.imageLayout = VK_IMAGE_LAYOUT_GENERAL;
+		imageInfo.imageView = image;
+		imageInfo.sampler = VK_NULL_HANDLE;
+		writes[n].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+		writes[n].dstBinding = 0;
+		writes[n].pImageInfo = &imageInfo;
+		writes[n].descriptorCount = 1;
+		writes[n].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
+		writes[n].dstSet = desc;
+		n++;
+	}
+	bufferInfo[0].buffer = buffer;
+	bufferInfo[0].offset = offset;
+	bufferInfo[0].range = range;
+	writes[n].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+	writes[n].dstBinding = 1;
+	writes[n].pBufferInfo = &bufferInfo[0];
+	writes[n].descriptorCount = 1;
+	writes[n].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+	writes[n].dstSet = desc;
+	n++;
+	if (buffer2) {
+		bufferInfo[1].buffer = buffer2;
+		bufferInfo[1].offset = offset2;
+		bufferInfo[1].range = range2;
+		writes[n].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+		writes[n].dstBinding = 2;
+		writes[n].pBufferInfo = &bufferInfo[1];
+		writes[n].descriptorCount = 1;
+		writes[n].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+		writes[n].dstSet = desc;
+		n++;
+	}
+	vkUpdateDescriptorSets(vulkan_->GetDevice(), n, writes, 0, nullptr);
+	return desc;
+}
+
+VkPipeline VulkanComputeShaderManager::GetPipeline(VkShaderModule cs) {
+	PipelineKey key{ cs };
+	VkPipeline pipeline = pipelines_.Get(key);
+	if (pipeline)
+		return pipeline;
+
+	VkComputePipelineCreateInfo pci{ VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO };
+	pci.stage.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
+	pci.stage.module = cs;
+	pci.stage.pName = "main";
+	pci.stage.stage = VK_SHADER_STAGE_COMPUTE_BIT;
+	pci.layout = pipelineLayout_;
+	pci.flags = 0;
+
+	vkCreateComputePipelines(vulkan_->GetDevice(), pipelineCache_, 1, &pci, nullptr, &pipeline);
+
+	pipelines_.Insert(key, pipeline);
+	return pipeline;
+}
+
+void VulkanComputeShaderManager::BeginFrame() {
+	int curFrame = vulkan_->GetCurFrame();
+	FrameData &frame = frameData_[curFrame];
+	frameData_[curFrame].numDescriptors = 0;
+	vkResetDescriptorPool(vulkan_->GetDevice(), frame.descPool, 0);
+}
+
+void VulkanComputeShaderManager::EndFrame() {
+}
diff --git a/GPU/Vulkan/VulkanUtil.h b/GPU/Vulkan/VulkanUtil.h
index 19b1ef4a37..53fd2ff858 100644
--- a/GPU/Vulkan/VulkanUtil.h
+++ b/GPU/Vulkan/VulkanUtil.h
@@ -20,6 +20,7 @@
 #include <tuple>
 #include <map>
 
+#include "Common/Hashmaps.h"
 #include "Common/Vulkan/VulkanContext.h"
 #include "Common/Vulkan/VulkanLoader.h"
 #include "Common/Vulkan/VulkanImage.h"
@@ -124,5 +125,52 @@ private:
 	std::vector<VkPipeline> keptPipelines_;
 };
 
+// Manager for compute shaders that upload things (and those have two bindings: a storage buffer to read from and an image to write to).
+class VulkanComputeShaderManager {
+public:
+	VulkanComputeShaderManager(VulkanContext *vulkan);
+	~VulkanComputeShaderManager();
+
+	void DeviceLost() {
+		DestroyDeviceObjects();
+	}
+	void DeviceRestore(VulkanContext *vulkan) {
+		vulkan_ = vulkan;
+		InitDeviceObjects();
+	}
+
+	// Note: This doesn't cache. The descriptor is for immediate use only.
+	VkDescriptorSet GetDescriptorSet(VkImageView image, VkBuffer buffer, VkDeviceSize offset, VkDeviceSize range, VkBuffer buffer2 = VK_NULL_HANDLE, VkDeviceSize offset2 = 0, VkDeviceSize range2 = 0);
+
+	// This of course caches though.
+	VkPipeline GetPipeline(VkShaderModule cs);
+	VkPipelineLayout GetPipelineLayout() const { return pipelineLayout_; }
+
+	void BeginFrame();
+	void EndFrame();
+
+private:
+	void InitDeviceObjects();
+	void DestroyDeviceObjects();
+
+	VulkanContext *vulkan_ = nullptr;
+	VkPipelineCache cache_ = VK_NULL_HANDLE;
+	VkDescriptorSetLayout descriptorSetLayout_ = VK_NULL_HANDLE;
+	VkPipelineLayout pipelineLayout_ = VK_NULL_HANDLE;
+	VkPipelineCache pipelineCache_ = VK_NULL_HANDLE;
+
+	struct FrameData {
+		VkDescriptorPool descPool;
+		int numDescriptors;
+	};
+	FrameData frameData_[VulkanContext::MAX_INFLIGHT_FRAMES];
+
+	struct PipelineKey {
+		VkShaderModule module;
+	};
+	
+	DenseHashMap<PipelineKey, VkPipeline, (VkPipeline)VK_NULL_HANDLE> pipelines_;
+};
+
 
 VkShaderModule CompileShaderModule(VulkanContext *vulkan, VkShaderStageFlagBits stage, const char *code, std::string *error);