diff --git a/DirectXTex/DirectXTex.h b/DirectXTex/DirectXTex.h
index c3a997f..5b09eb6 100644
--- a/DirectXTex/DirectXTex.h
+++ b/DirectXTex/DirectXTex.h
@@ -67,6 +67,9 @@ namespace DirectX
         CP_FLAGS_NONE               = 0x0,      // Normal operation
         CP_FLAGS_LEGACY_DWORD       = 0x1,      // Assume pitch is DWORD aligned instead of BYTE aligned
         CP_FLAGS_PARAGRAPH          = 0x2,      // Assume pitch is 16-byte aligned instead of BYTE aligned
+        CP_FLAGS_YMM                = 0x4,      // Assume pitch is 32-byte aligned instead of BYTE aligned
+        CP_FLAGS_ZMM                = 0x8,      // Assume pitch is 64-byte aligned instead of BYTE aligned
+        CP_FLAGS_PAGE4K             = 0x200,    // Assume pitch is 4096-byte aligned instead of BYTE aligned
         CP_FLAGS_24BPP              = 0x10000,  // Override with a legacy 24 bits-per-pixel format size
         CP_FLAGS_16BPP              = 0x20000,  // Override with a legacy 16 bits-per-pixel format size
         CP_FLAGS_8BPP               = 0x40000,  // Override with a legacy 8 bits-per-pixel format size
diff --git a/DirectXTex/DirectXTexCompress.cpp b/DirectXTex/DirectXTexCompress.cpp
index a66bbdf..0e338ed 100644
--- a/DirectXTex/DirectXTexCompress.cpp
+++ b/DirectXTex/DirectXTexCompress.cpp
@@ -110,7 +110,7 @@ static HRESULT _CompressBC( _In_ const Image& image, _In_ const Image& result, _
         uint8_t* dptr = pDest;
         size_t ph = std::min<size_t>( 4, image.height - h );
         size_t w = 0;
-        for( size_t count = 0; count < rowPitch; count += sbpp*4, w += 4 )
+        for( size_t count = 0; count < result.rowPitch; count += blocksize, w += 4 )
         {
             size_t pw = std::min<size_t>( 4, image.width - w );
             assert( pw > 0 && ph > 0 );
diff --git a/DirectXTex/DirectXTexUtil.cpp b/DirectXTex/DirectXTexUtil.cpp
index 5a343fd..cd8a88f 100644
--- a/DirectXTex/DirectXTexUtil.cpp
+++ b/DirectXTex/DirectXTexUtil.cpp
@@ -654,20 +654,39 @@ void ComputePitch( DXGI_FORMAT fmt, size_t width, size_t height,
         else
             bpp = BitsPerPixel( fmt );
 
-        if ( flags & CP_FLAGS_LEGACY_DWORD )
+        if ( flags & ( CP_FLAGS_LEGACY_DWORD | CP_FLAGS_PARAGRAPH | CP_FLAGS_YMM | CP_FLAGS_ZMM | CP_FLAGS_PAGE4K ) )
         {
-            // Special computation for some incorrectly created DDS files based on
-            // legacy DirectDraw assumptions about pitch alignment
-            rowPitch = ( ( width * bpp + 31 ) / 32 ) * sizeof(uint32_t);
-            slicePitch = rowPitch * height;
-        }
-        else if ( flags & CP_FLAGS_PARAGRAPH )
-        {
-            rowPitch = ( ( width * bpp + 127 ) / 128 ) * 16;
-            slicePitch = rowPitch * height;
+            if ( flags & CP_FLAGS_PAGE4K )
+            {
+                rowPitch = ( ( width * bpp + 32767 ) / 32768 ) * 4096;
+                slicePitch = rowPitch * height;
+            }
+            else if ( flags & CP_FLAGS_ZMM )
+            {
+                rowPitch = ( ( width * bpp + 511 ) / 512 ) * 64;
+                slicePitch = rowPitch * height;
+            }
+            else if ( flags & CP_FLAGS_YMM )
+            {
+                rowPitch = ( ( width * bpp + 255 ) / 256) * 32;
+                slicePitch = rowPitch * height;
+            }
+            else if ( flags & CP_FLAGS_PARAGRAPH )
+            {
+                rowPitch = ( ( width * bpp + 127 ) / 128 ) * 16;
+                slicePitch = rowPitch * height;
+            }
+            else // DWORD alignment
+            {
+                // Special computation for some incorrectly created DDS files based on
+                // legacy DirectDraw assumptions about pitch alignment
+                rowPitch = ( ( width * bpp + 31 ) / 32 ) * sizeof(uint32_t);
+                slicePitch = rowPitch * height;
+            }
         }
         else
         {
+            // Default byte alignment
             rowPitch = ( width * bpp + 7 ) / 8;
             slicePitch = rowPitch * height;
         }