GregMiscellaneous: zzogl-pg: Switch to _aligned_malloc in GetMemoryTarget, take 2. (If at first you don't succeed...)

git-svn-id: http://pcsx2.googlecode.com/svn/branches/GregMiscellaneous@3809 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/zzogl-pg/opengl/targets.cpp

@@ -2202,13 +2202,13 @@ ZeroGS::CMemoryTarget* ZeroGS::CMemoryTargetMngr::GetMemoryTarget(const tex0Info
 
 	memcpy_amd(targ->ptex->memptr, g_pbyGSMemory + 4 * GPU_TEXWIDTH * targ->realy, 4 * GPU_TEXWIDTH * targ->height);
 
-	vector<u8> texdata;
+	__aligned16 u8* ptexdata = NULL;
-	u8* ptexdata = NULL;
+	bool has_data = false;
 
 	if (PSMT_ISCLUT(tex0.psm))
 	{
-		texdata.resize(((tex0.cpsm <= 1) ? 4 : 2) * texW * texH);
+		ptexdata = (u8*)_aligned_malloc(((tex0.cpsm <= 1) ? 4 : 2) * texW * texH, 16);
-		ptexdata = &texdata[0];
+		has_data = true;
 
 		u8* psrc = (u8*)(g_pbyGSMemory + 4 * GPU_TEXWIDTH * targ->realy);
 
@@ -2231,43 +2231,16 @@ ZeroGS::CMemoryTarget* ZeroGS::CMemoryTargetMngr::GetMemoryTarget(const tex0Info
 	{
 		if (tex0.psm == PSMT16Z || tex0.psm == PSMT16SZ)
 		{
-			texdata.resize(4 * texW * texH
+			ptexdata = (u8*)_aligned_malloc(4 * texW * texH, 16);
-#if defined(ZEROGS_SSE2)
+			has_data = true;
-			+ 15 						// reserve additional elements for alignment if SSE2 used.
-										// better do it now, so less resizing would be needed
-#endif
-							);
-
-			ptexdata = &texdata[0];
 
 			// needs to be 8 bit, use xmm for unpacking
 			u16* dst = (u16*)ptexdata;
 			u16* src = (u16*)(g_pbyGSMemory + 4 * GPU_TEXWIDTH * targ->realy);
 
 #if defined(ZEROGS_SSE2)
-			if (((u32)(uptr)dst) % 16 != 0)
+			assert(((u32)(uptr)dst) % 16 == 0);
-			{
+			SSE2_UnswizzleZ16Target(dst, src, targ->height * GPU_TEXWIDTH / 16);
-				// This is not unusual situation, when vector<u8> does not 16bit alignment, that is destructive for SSE2
-				// instruction movdqa [%eax], xmm0
-				// The idea would be resize vector to 15 elements, that set ptxedata to aligned position.
-				// Later we would move eax by 16, so only we should verify is first element align
-				
-				// FIXME. As I see, texdata used only once here, it does not have any impact on other code.
-				// Probably, usage of _aligned_maloc() would be preferable.
-				
-				// Update: Apparently not, as Zeydlitz says it leads to a heavy slowdown. --arcum42
-
-				// Note: this often happens when changing AA.
-				int disalignment = 16 - ((u32)(uptr)dst) % 16;		// This is value of shift. It could be 0 < disalignment <= 15
-				ptexdata = &texdata[disalignment];			// Set pointer to aligned element
-				dst = (u16*)ptexdata;
-				ZZLog::GS_Log("Made alignment for texdata, 0x%x", dst);
-				assert(((u32)(uptr)dst) % 16 == 0);			// Assert, because at future could be vectors with uncontigious spaces
-			}
-
-			int iters = targ->height * GPU_TEXWIDTH / 16;
-
-			SSE2_UnswizzleZ16Target(dst, src, iters) ;
 #else // ZEROGS_SSE2
 
 			for (int i = 0; i < targ->height; ++i)
@@ -2288,6 +2261,8 @@ ZeroGS::CMemoryTarget* ZeroGS::CMemoryTargetMngr::GetMemoryTarget(const tex0Info
 		else
 		{
 			ptexdata = targ->ptex->memptr;
+			// We really don't want to deallocate memptr. As a reminder...
+			has_data = false;
 		}
 	}
 
@@ -2320,6 +2295,7 @@ ZeroGS::CMemoryTarget* ZeroGS::CMemoryTargetMngr::GetMemoryTarget(const tex0Info
 			{
 				ZZLog::Error_Log("Failed to create %dx%x texture.", GPU_TEXWIDTH*channels*widthmult, (realheight + widthmult - 1) / widthmult);
 				channels = 1;
+				if (has_data) _aligned_free(ptexdata);
 				return NULL;
 			}
 
@@ -2330,6 +2306,7 @@ ZeroGS::CMemoryTarget* ZeroGS::CMemoryTargetMngr::GetMemoryTarget(const tex0Info
 	}
 
 	setRectWrap(GL_CLAMP);
+	if (has_data) _aligned_free(ptexdata);
 
 	assert(tex0.psm != 0xd);