GS/HW: Fix various issues with texture replacements

- Replacement textures now show in HC usage to give a clearer picture
   of VRAM usage.
 - Fixed crashes when loading compressed and mipmapped DDS textures.
 - Fixed compressed mipmapped textures in Direct3D 12.
 - Fixed GPU crashes in D3D11/D3D12/Vulkan when compressed textures went
   down the last 1x1 mipmap level.

common/Vulkan/Texture.cpp

@@ -369,14 +369,14 @@ namespace Vulkan
 	}
 
 	void Texture::UpdateFromBuffer(VkCommandBuffer cmdbuf, u32 level, u32 layer, u32 x, u32 y, u32 width, u32 height,
-		u32 row_length, VkBuffer buffer, u32 buffer_offset)
+		u32 buffer_height, u32 row_length, VkBuffer buffer, u32 buffer_offset)
 	{
 		const VkImageLayout old_layout = m_layout;
 		if (old_layout != VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL)
 			TransitionSubresourcesToLayout(
 				cmdbuf, level, 1, layer, 1, old_layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
 
-		const VkBufferImageCopy bic = {static_cast<VkDeviceSize>(buffer_offset), row_length, height,
+		const VkBufferImageCopy bic = {static_cast<VkDeviceSize>(buffer_offset), row_length, buffer_height,
 			{VK_IMAGE_ASPECT_COLOR_BIT, level, layer, 1u}, {static_cast<int32_t>(x), static_cast<int32_t>(y), 0},
 			{width, height, 1u}};
 

common/Vulkan/Texture.h

@@ -72,7 +72,7 @@ namespace Vulkan
 		VkFramebuffer CreateFramebuffer(VkRenderPass render_pass);
 
 		void UpdateFromBuffer(VkCommandBuffer cmdbuf, u32 level, u32 layer, u32 x, u32 y, u32 width, u32 height,
-			u32 row_length, VkBuffer buffer, u32 buffer_offset);
+			u32 buffer_height, u32 row_length, VkBuffer buffer, u32 buffer_offset);
 
 	private:
 		u32 m_width = 0;

pcsx2/Frontend/VulkanHostDisplay.cpp

@@ -222,7 +222,7 @@ static bool UploadBufferToTexture(
 	StringUtil::StrideMemCpy(buf.GetCurrentHostPointer(), upload_stride, data, data_stride, row_size, height);
 	buf.CommitMemory(upload_size);
 
-	texture->UpdateFromBuffer(cmdbuf, 0, 0, 0, 0, width, height, upload_stride / texel_size, buf.GetBuffer(), buf_offset);
+	texture->UpdateFromBuffer(cmdbuf, 0, 0, 0, 0, width, height, height, upload_stride / texel_size, buf.GetBuffer(), buf_offset);
 	return true;
 }
 

pcsx2/Frontend/imgui_impl_vulkan.cpp

@@ -428,7 +428,7 @@ bool ImGui_ImplVulkan_CreateFontsTexture()
     std::memcpy(ai.pMappedData, pixels, upload_size);
     vmaFlushAllocation(g_vulkan_context->GetAllocator(), allocation, 0, upload_size);
     bd->FontTexture.TransitionToLayout(g_vulkan_context->GetCurrentInitCommandBuffer(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
-    bd->FontTexture.UpdateFromBuffer(g_vulkan_context->GetCurrentInitCommandBuffer(), 0, 0, 0, 0, width, height, width, buffer, 0);
+    bd->FontTexture.UpdateFromBuffer(g_vulkan_context->GetCurrentInitCommandBuffer(), 0, 0, 0, 0, width, height, height, width, buffer, 0);
     bd->FontTexture.TransitionToLayout(g_vulkan_context->GetCurrentInitCommandBuffer(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 
     // Immediately queue it for freeing after the command buffer finishes, since it's only needed for the copy.

pcsx2/GS/Renderers/DX11/GSTexture11.cpp

@@ -18,6 +18,7 @@
 #include "GSTexture11.h"
 #include "GS/GSPng.h"
 #include "GS/GSPerfMon.h"
+#include "common/Align.h"
 
 GSTexture11::GSTexture11(wil::com_ptr_nothrow<ID3D11Texture2D> texture, const D3D11_TEXTURE2D_DESC& desc,
 	GSTexture::Type type, GSTexture::Format format)
@@ -44,7 +45,9 @@ bool GSTexture11::Update(const GSVector4i& r, const void* data, int pitch, int l
 
 	g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
-	const D3D11_BOX box = {(UINT)r.left, (UINT)r.top, 0U, (UINT)r.right, (UINT)r.bottom, 1U};
+	const u32 bs = GetCompressedBlockSize();
+	const D3D11_BOX box = {Common::AlignDownPow2((u32)r.left, bs), Common::AlignDownPow2((u32)r.top, bs),
+		Common::AlignUpPow2((u32)r.right, bs), Common::AlignUpPow2((u32)r.bottom, bs), 1U};
 	const UINT subresource = layer; // MipSlice + (ArraySlice * MipLevels).
 
 	GSDevice11::GetInstance()->GetD3DContext()->UpdateSubresource(m_texture.get(), subresource, &box, data, pitch, 0);

pcsx2/GS/Renderers/DX12/GSTexture12.cpp

@@ -48,13 +48,11 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, u32 width, u32 heigh
 	{
 		case Type::Texture:
 		{
-			// this is a little annoying. basically, to do mipmap generation, we need to be a render target.
+			// This is a little annoying. basically, to do mipmap generation, we need to be a render target.
-			const D3D12_RESOURCE_FLAGS flags = (levels > 1) ? D3D12_RESOURCE_FLAG_ALLOW_RENDER_TARGET : D3D12_RESOURCE_FLAG_NONE;
+			// If it's a compressed texture, we won't be generating mips anyway, so this should be fine.
-			if (levels > 1 && d3d_format != DXGI_FORMAT_R8G8B8A8_UNORM)
+			const D3D12_RESOURCE_FLAGS flags = (levels > 1 && !IsCompressedFormat(format)) ?
-			{
+												   D3D12_RESOURCE_FLAG_ALLOW_RENDER_TARGET :
-				Console.Warning("DX12: Refusing to create a %ux%u format %u mipmapped texture", width, height, format);
+                                                   D3D12_RESOURCE_FLAG_NONE;
-				return nullptr;
-			}
 
 			D3D12::Texture texture;
 			if (!texture.Create(width, height, levels, d3d_format, srv_format,
@@ -184,10 +182,12 @@ bool GSTexture12::Update(const GSVector4i& r, const void* data, int pitch, int l
 
 	g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
-	const u32 width = r.width();
+	// Footprint and box must be block aligned for compressed textures.
-	const u32 height = r.height();
+	const u32 block_size = GetCompressedBlockSize();
+	const u32 width = Common::AlignUpPow2(r.width(), block_size);
+	const u32 height = Common::AlignUpPow2(r.height(), block_size);
 	const u32 upload_pitch = Common::AlignUpPow2<u32>(pitch, D3D12_TEXTURE_DATA_PITCH_ALIGNMENT);
-	const u32 required_size = CalcUploadSize(height, upload_pitch);
+	const u32 required_size = CalcUploadSize(r.height(), upload_pitch);
 
 	D3D12_TEXTURE_COPY_LOCATION srcloc;
 	srcloc.Type = D3D12_TEXTURE_COPY_TYPE_PLACED_FOOTPRINT;
@@ -250,8 +250,9 @@ bool GSTexture12::Update(const GSVector4i& r, const void* data, int pitch, int l
 	dstloc.Type = D3D12_TEXTURE_COPY_TYPE_SUBRESOURCE_INDEX;
 	dstloc.SubresourceIndex = layer;
 
-	const D3D12_BOX srcbox{0u, 0u, 0u, static_cast<UINT>(r.width()), static_cast<UINT>(r.height()), 1u};
+	const D3D12_BOX srcbox{0u, 0u, 0u, width, height, 1u};
-	cmdlist->CopyTextureRegion(&dstloc, r.x, r.y, 0, &srcloc, &srcbox);
+	cmdlist->CopyTextureRegion(&dstloc, Common::AlignDownPow2((u32)r.x, block_size),
+		Common::AlignDownPow2((u32)r.y, block_size), 0, &srcloc, &srcbox);
 
 	if (m_texture.GetState() != D3D12_RESOURCE_STATE_COPY_DEST)
 		m_texture.TransitionSubresourceToState(cmdlist, layer, D3D12_RESOURCE_STATE_COPY_DEST, m_texture.GetState());
@@ -293,7 +294,8 @@ bool GSTexture12::Map(GSMap& m, const GSVector4i* r, int layer)
 
 void GSTexture12::Unmap()
 {
-	pxAssert(m_map_level < m_texture.GetLevels());
+	// this can't handle blocks/compressed formats at the moment.
+	pxAssert(m_map_level < m_texture.GetLevels() && !IsCompressedFormat());
 	g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
 	// TODO: non-tightly-packed formats
@@ -351,6 +353,8 @@ void GSTexture12::Unmap()
 
 void GSTexture12::GenerateMipmap()
 {
+	pxAssert(!IsCompressedFormat(m_format));
+
 	for (int dst_level = 1; dst_level < m_mipmap_levels; dst_level++)
 	{
 		const int src_level = dst_level - 1;

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -1644,9 +1644,7 @@ void GSTextureCache::IncAge()
 		HashCacheEntry& e = it->second;
 		if (e.refcount == 0 && ++e.age > max_hash_cache_age)
 		{
-			if (!e.is_replacement)
+			m_hash_cache_memory_usage -= e.texture->GetMemUsage();
-				m_hash_cache_memory_usage -= e.texture->GetMemUsage();
-
 			g_gs_device->Recycle(e.texture);
 			m_hash_cache.erase(it++);
 		}
@@ -2129,6 +2127,7 @@ GSTextureCache::HashCacheEntry* GSTextureCache::LookupHashCache(const GIFRegTEX0
 			// found a replacement texture! insert it into the hash cache, and clear paltex (since it's not indexed)
 			paltex = false;
 			const HashCacheEntry entry{replacement_tex, 1u, 0u, true};
+			m_hash_cache_memory_usage += entry.texture->GetMemUsage();
 			return &m_hash_cache.emplace(key, entry).first->second;
 		}
 		else if (
@@ -3163,6 +3162,9 @@ void GSTextureCache::InvalidateTemporarySource()
 
 void GSTextureCache::InjectHashCacheTexture(const HashCacheKey& key, GSTexture* tex)
 {
+	// When we insert we update memory usage. Old texture gets removed below.
+	m_hash_cache_memory_usage += tex->GetMemUsage();
+
 	auto it = m_hash_cache.find(key);
 	if (it == m_hash_cache.end())
 	{

pcsx2/GS/Renderers/HW/GSTextureReplacementLoaders.cpp

@@ -68,7 +68,7 @@ static u32 GetBlockCount(u32 extent, u32 block_size)
 static void CalcBlockMipmapSize(u32 block_size, u32 bytes_per_block, u32 base_width, u32 base_height, u32 mip, u32& width, u32& height, u32& pitch, u32& size)
 {
 	width = std::max<u32>(base_width >> mip, 1u);
-	height = std::max<u32>(base_width >> mip, 1u);
+	height = std::max<u32>(base_height >> mip, 1u);
 
 	const u32 blocks_wide = GetBlockCount(width, block_size);
 	const u32 blocks_high = GetBlockCount(height, block_size);
@@ -635,9 +635,9 @@ bool DDSLoader(const std::string& filename, GSTextureReplacements::ReplacementTe
 		for (u32 level = 1; level <= info.mip_count; level++)
 		{
 			GSTextureReplacements::ReplacementTexture::MipData md;
-			u32 mip_width, mip_height, mip_size;
+			u32 mip_size;
-			CalcBlockMipmapSize(info.block_size, info.bytes_per_block, info.width, info.height, level, mip_width, mip_height, md.pitch, mip_size);
+			CalcBlockMipmapSize(info.block_size, info.bytes_per_block, info.width, info.height, level, md.width, md.height, md.pitch, mip_size);
-			if (!ReadDDSMipLevel(fp.get(), filename, level, info, mip_width, mip_height, md.data, md.pitch, mip_size))
+			if (!ReadDDSMipLevel(fp.get(), filename, level, info, md.width, md.height, md.data, md.pitch, mip_size))
 				break;
 
 			tex->mips.push_back(std::move(md));

pcsx2/GS/Renderers/HW/GSTextureReplacements.cpp

@@ -23,6 +23,8 @@
 #include "common/ScopedGuard.h"
 
 #include "Config.h"
+#include "Host.h"
+#include "IconsFontAwesome5.h"
 #include "GS/GSLocalMemory.h"
 #include "GS/Renderers/HW/GSTextureReplacements.h"
 
@@ -325,7 +327,7 @@ void GSTextureReplacements::ReloadReplacementMap()
 		if (!name.has_value())
 			continue;
 
-		DevCon.WriteLn("Found %ux%u replacement '%.*s'", name->Width(), name->Height(), static_cast<int>(filename.size()), filename.data());
+		DbgCon.WriteLn("Found %ux%u replacement '%.*s'", name->Width(), name->Height(), static_cast<int>(filename.size()), filename.data());
 		s_replacement_texture_filenames.emplace(name.value(), std::move(fd.FileName));
 
 		// zero out the CLUT hash, because we need this for checking if there's any replacements with this hash when using paltex
@@ -529,7 +531,10 @@ GSTexture* GSTextureReplacements::CreateReplacementTexture(const ReplacementText
 		static bool log_once = false;
 		if (!log_once)
 		{
-			Console.Warning("Disabling mipmaps on one or more compressed replacement textures.");
+			static const char* message =
+				"Disabling autogenerated mipmaps on one or more compressed replacement textures. Please generate mipmaps when compressing your textures.";
+			Console.Warning(message);
+			Host::AddIconOSDMessage("DisablingReplacementAutoGeneratedMipmap", ICON_FA_EXCLAMATION_CIRCLE, message, Host::OSD_WARNING_DURATION);
 			log_once = true;
 		}
 
@@ -548,10 +553,8 @@ GSTexture* GSTextureReplacements::CreateReplacementTexture(const ReplacementText
 	{
 		for (u32 i = 0; i < static_cast<u32>(rtex.mips.size()); i++)
 		{
-			const u32 mip = i + 1;
+			const ReplacementTexture::MipData& mip = rtex.mips[i];
-			const u32 mipw = std::max<u32>(rtex.width >> mip, 1u);
+			tex->Update(GSVector4i(0, 0, static_cast<int>(mip.width), static_cast<int>(mip.height)), mip.data.data(), mip.pitch, i + 1);
-			const u32 miph = std::max<u32>(rtex.height >> mip, 1u);
-			tex->Update(GSVector4i(0, 0, mipw, miph), rtex.mips[i].data.data(), rtex.mips[i].pitch, mip);
 		}
 	}
 

pcsx2/GS/Renderers/HW/GSTextureReplacements.h

@@ -30,6 +30,8 @@ namespace GSTextureReplacements
 
 		struct MipData
 		{
+			u32 width;
+			u32 height;
 			u32 pitch;
 			std::vector<u8> data;
 		};

pcsx2/GS/Renderers/Vulkan/GSTextureVK.cpp

@@ -262,6 +262,7 @@ bool GSTextureVK::Update(const GSVector4i& r, const void* data, int pitch, int l
 	}
 
 	m_texture.UpdateFromBuffer(cmdbuf, layer, 0, r.x, r.y, width, height,
+		Common::AlignUpPow2(height, GetCompressedBlockSize()),
 		CalcUploadRowLengthFromPitch(upload_pitch), buffer, buffer_offset);
 	m_texture.TransitionToLayout(cmdbuf, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 
@@ -303,7 +304,8 @@ bool GSTextureVK::Map(GSMap& m, const GSVector4i* r, int layer)
 
 void GSTextureVK::Unmap()
 {
-	pxAssert(m_map_level < m_texture.GetLevels());
+	// this can't handle blocks/compressed formats at the moment.
+	pxAssert(m_map_level < m_texture.GetLevels() && !IsCompressedFormat());
 	g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
 	// TODO: non-tightly-packed formats
@@ -334,6 +336,7 @@ void GSTextureVK::Unmap()
 	}
 
 	m_texture.UpdateFromBuffer(cmdbuf, m_map_level, 0, m_map_area.x, m_map_area.y, width, height,
+		Common::AlignUpPow2(height, GetCompressedBlockSize()),
 		CalcUploadRowLengthFromPitch(pitch), buffer.GetBuffer(), buffer_offset);
 	m_texture.TransitionToLayout(cmdbuf, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);