From ba30da0945c266fa8e765ba812d8d90b9d29796a Mon Sep 17 00:00:00 2001
From: Chuck Walbourn <chuckw@windows.microsoft.com>
Date: Tue, 3 Jul 2018 19:33:07 -0700
Subject: [PATCH] BC7 CPU codec optimization to skip mode 7 for opaque blocks

---
 DirectXTex/BC6HBC7.cpp | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/DirectXTex/BC6HBC7.cpp b/DirectXTex/BC6HBC7.cpp
index ef752e9..594d22a 100644
--- a/DirectXTex/BC6HBC7.cpp
+++ b/DirectXTex/BC6HBC7.cpp
@@ -2733,6 +2733,7 @@ void D3DX_BC7::Encode(DWORD flags, const HDRColorA* const pIn)
     D3DX_BC7 final = *this;
     EncodeParams EP(pIn);
     float fMSEBest = FLT_MAX;
+    uint32_t alphaMask = 0xFF;
 
     for (size_t i = 0; i < NUM_PIXELS_PER_BLOCK; ++i)
     {
@@ -2740,8 +2741,11 @@ void D3DX_BC7::Encode(DWORD flags, const HDRColorA* const pIn)
         EP.aLDRPixels[i].g = uint8_t(std::max<float>(0.0f, std::min<float>(255.0f, pIn[i].g * 255.0f + 0.01f)));
         EP.aLDRPixels[i].b = uint8_t(std::max<float>(0.0f, std::min<float>(255.0f, pIn[i].b * 255.0f + 0.01f)));
         EP.aLDRPixels[i].a = uint8_t(std::max<float>(0.0f, std::min<float>(255.0f, pIn[i].a * 255.0f + 0.01f)));
+        alphaMask &= EP.aLDRPixels[i].a;
     }
 
+    const bool bHasAlpha = (alphaMask != 0xFF);
+
     for (EP.uMode = 0; EP.uMode < 8 && fMSEBest > 0; ++EP.uMode)
     {
         if (!(flags & BC_FLAGS_USE_3SUBSETS) && (EP.uMode == 0 || EP.uMode == 2))
@@ -2756,6 +2760,12 @@ void D3DX_BC7::Encode(DWORD flags, const HDRColorA* const pIn)
             continue;
         }
 
+        if ((!bHasAlpha) && (EP.uMode == 7))
+        {
+            // There is no value in using mode 7 for completely opaque blocks (the other 2 subset modes handle this case for opaque blocks), so skip it for a small perf win.
+            continue;
+        }
+
         const size_t uShapes = size_t(1) << ms_aInfo[EP.uMode].uPartitionBits;
         assert(uShapes <= BC7_MAX_SHAPES);
         _Analysis_assume_(uShapes <= BC7_MAX_SHAPES);