diff --git a/Core/Core.cpp b/Core/Core.cpp
index c1caba9f5f..5153b2e4b7 100644
--- a/Core/Core.cpp
+++ b/Core/Core.cpp
@@ -180,7 +180,6 @@ void GPU_SwapBuffers() {
 		D3D9_SwapBuffers();
 		break;
 	}
-	PROFILE_END_FRAME();
 }
 
 #endif
diff --git a/GPU/GLES/TextureCache.cpp b/GPU/GLES/TextureCache.cpp
index 797f74a48a..1f87fa3564 100644
--- a/GPU/GLES/TextureCache.cpp
+++ b/GPU/GLES/TextureCache.cpp
@@ -1893,7 +1893,14 @@ TextureCache::TexCacheEntry::Status TextureCache::CheckAlpha(const u32 *pixelDat
 }
 
 void TextureCache::LoadTextureLevel(TexCacheEntry &entry, int level, bool replaceImages, int scaleFactor, GLenum dstFmt) {
-	PROFILE_THIS_SCOPE("loadtex");
+	int w = gstate.getTextureWidth(level);
+	int h = gstate.getTextureHeight(level);
+	bool useUnpack = false;
+	bool useBGRA;
+	u32 *pixelData;
+	{
+
+	PROFILE_THIS_SCOPE("decodetex");
 
 	// TODO: only do this once
 	u32 texByteAlign = 1;
@@ -1905,13 +1912,9 @@ void TextureCache::LoadTextureLevel(TexCacheEntry &entry, int level, bool replac
 		return;
 	}
 
-	int w = gstate.getTextureWidth(level);
-	int h = gstate.getTextureHeight(level);
-
 	gpuStats.numTexturesDecoded++;
 
 	// Can restore these and remove the fixup at the end of DecodeTextureLevel on desktop GL and GLES 3.
-	bool useUnpack = false;
 	if ((g_Config.iTexScalingLevel == 1 && gl_extensions.EXT_unpack_subimage) && w != bufw) {
 		glPixelStorei(GL_UNPACK_ROW_LENGTH, bufw);
 		useUnpack = true;
@@ -1919,9 +1922,9 @@ void TextureCache::LoadTextureLevel(TexCacheEntry &entry, int level, bool replac
 
 	glPixelStorei(GL_UNPACK_ALIGNMENT, texByteAlign);
 
-	bool useBGRA = UseBGRA8888() && dstFmt == GL_UNSIGNED_BYTE;
+	useBGRA = UseBGRA8888() && dstFmt == GL_UNSIGNED_BYTE;
 
-	u32 *pixelData = (u32 *)finalBuf;
+	pixelData = (u32 *)finalBuf;
 	if (scaleFactor > 1 && (entry.status & TexCacheEntry::STATUS_CHANGE_FREQUENT) == 0)
 		scaler.Scale(pixelData, dstFmt, w, h, scaleFactor);
 
@@ -1931,6 +1934,7 @@ void TextureCache::LoadTextureLevel(TexCacheEntry &entry, int level, bool replac
 	} else {
 		entry.SetAlphaStatus(TexCacheEntry::STATUS_ALPHA_UNKNOWN);
 	}
+	}
 
 	GLuint components = dstFmt == GL_UNSIGNED_SHORT_5_6_5 ? GL_RGB : GL_RGBA;
 
@@ -1940,8 +1944,10 @@ void TextureCache::LoadTextureLevel(TexCacheEntry &entry, int level, bool replac
 	}
 
 	if (replaceImages) {
+		PROFILE_THIS_SCOPE("repltex");
 		glTexSubImage2D(GL_TEXTURE_2D, level, 0, 0, w, h, components2, dstFmt, pixelData);
 	} else {
+		PROFILE_THIS_SCOPE("loadtex");
 		glTexImage2D(GL_TEXTURE_2D, level, components, w, h, 0, components2, dstFmt, pixelData);
 		if (!lowMemoryMode_) {
 			GLenum err = glGetError();
diff --git a/UI/NativeApp.cpp b/UI/NativeApp.cpp
index 264017e052..4182232968 100644
--- a/UI/NativeApp.cpp
+++ b/UI/NativeApp.cpp
@@ -48,7 +48,6 @@
 #include "file/zip_read.h"
 #include "thread/thread.h"
 #include "net/http_client.h"
-#include "gfx_es2/gl_state.h"  // should've been only for screenshot - but actually not, cleanup?
 #include "gfx_es2/draw_text.h"
 #include "gfx/gl_lost_manager.h"
 #include "gfx/texture.h"
@@ -57,6 +56,7 @@
 #include "math/fast/fast_math.h"
 #include "math/math_util.h"
 #include "math/lin/matrix4x4.h"
+#include "profiler/profiler.h"
 #include "thin3d/thin3d.h"
 #include "ui/ui.h"
 #include "ui/screen.h"
@@ -762,6 +762,8 @@ void HandleGlobalMessage(const std::string &msg, const std::string &value) {
 }
 
 void NativeUpdate(InputState &input) {
+	PROFILE_END_FRAME();
+
 	{
 		lock_guard lock(pendingMutex);
 		for (size_t i = 0; i < pendingMessages.size(); i++) {