Add float conversion flags

Author: elasota

DirectXTex/DirectXTex.h

@@ -450,6 +450,12 @@ namespace DirectX
         TEX_FILTER_FLOAT_X2BIAS     = 0x200,
             // Enable *2 - 1 conversion cases for unorm<->float and positive-only float formats
 
+        TEX_FILTER_FLOAT_SATURATE_TO_INF = 0x400,
+            // When converting to smaller float formats, saturate out-of-range values to +INF/-INF instead of largest in-range value
+
+        TEX_FILTER_FLOAT_KEEP_NANS = 0x800,
+            // When converting to smaller float formats, preserve NaNs instead of converting to zero
+
         TEX_FILTER_RGB_COPY_RED     = 0x1000,
         TEX_FILTER_RGB_COPY_GREEN   = 0x2000,
         TEX_FILTER_RGB_COPY_BLUE    = 0x4000,

DirectXTex/DirectXTexConvert.cpp

@@ -192,6 +192,30 @@ namespace
     const XMVECTORF32 g_HalfMin   = { { { -65504.f, -65504.f, -65504.f, -65504.f } } };
     const XMVECTORF32 g_HalfMax   = { { { 65504.f, 65504.f, 65504.f, 65504.f } } };
     const XMVECTORF32 g_8BitBias  = { { { 0.5f / 255.f, 0.5f / 255.f, 0.5f / 255.f, 0.5f / 255.f } } };
+
+    inline void SanitizeFloat16(XMVECTOR& v, DWORD flags)
+    {
+        // Convert NaNs to 0.0f
+        if (!(flags & TEX_FILTER_FLOAT_KEEP_NANS))
+            v = XMVectorSelect(XMVectorZero(), v, XMVectorIsNaN(v));
+
+        if (!(flags & TEX_FILTER_FLOAT_SATURATE_TO_INF))
+            v = XMVectorClamp(v, g_HalfMin, g_HalfMax);
+    }
+
+    inline void SanitizeFloat16(float& v, DWORD flags)
+    {
+        if (std::isnan(v))
+        {
+            if (!(flags & TEX_FILTER_FLOAT_KEEP_NANS))
+                v = 0.0f;
+        }
+        else
+        {
+            if (!(flags & TEX_FILTER_FLOAT_SATURATE_TO_INF))
+                v = std::min(std::max(v, -65504.f), 65504.f);
+        }
+    }
 }
 
 //-------------------------------------------------------------------------------------
@@ -1609,7 +1633,8 @@ bool DirectX::_StoreScanline(
     DXGI_FORMAT format,
     const XMVECTOR* pSource,
     size_t count,
-    float threshold)
+    float threshold,
+    DWORD flags)
 {
     assert(pDestination && size > 0);
     assert(pSource && count > 0 && (((uintptr_t)pSource & 0xF) == 0));
@@ -1649,7 +1674,7 @@ bool DirectX::_StoreScanline(
             {
                 if (sPtr >= ePtr) break;
                 XMVECTOR v = *sPtr++;
-                v = XMVectorClamp(v, g_HalfMin, g_HalfMax);
+                SanitizeFloat16(v, flags);
                 XMStoreHalf4(dPtr++, v);
             }
             return true;
@@ -1743,7 +1768,7 @@ bool DirectX::_StoreScanline(
             {
                 if (sPtr >= ePtr) break;
                 XMVECTOR v = *sPtr++;
-                v = XMVectorClamp(v, g_HalfMin, g_HalfMax);
+                SanitizeFloat16(v, flags);
                 XMStoreHalf2(dPtr++, v);
             }
             return true;
@@ -1842,7 +1867,7 @@ bool DirectX::_StoreScanline(
             {
                 if (sPtr >= ePtr) break;
                 float v = XMVectorGetX(*sPtr++);
-                v = std::max<float>(std::min<float>(v, 65504.f), -65504.f);
+                SanitizeFloat16(v, flags);
                 *(dPtr++) = XMConvertFloatToHalf(v);
             }
             return true;
@@ -2699,7 +2724,7 @@ bool DirectX::_StoreScanlineLinear(
         }
     }
 
-    return _StoreScanline(pDestination, size, format, pSource, count, threshold);
+    return _StoreScanline(pDestination, size, format, pSource, count, threshold, flags);
 }
 
 
@@ -4542,7 +4567,7 @@ namespace
 
                     _ConvertScanline(scanline.get(), width, destImage.format, srcImage.format, filter);
 
-                    if (!_StoreScanline(pDest, destImage.rowPitch, destImage.format, scanline.get(), width, threshold))
+                    if (!_StoreScanline(pDest, destImage.rowPitch, destImage.format, scanline.get(), width, threshold, filter))
                         return E_FAIL;
 
                     pSrc += srcImage.rowPitch;

DirectXTex/DirectXTexMipmaps.cpp

@@ -641,7 +641,7 @@ namespace
     }
 
     //--- 2D Point Filter ---
-    HRESULT Generate2DMipsPointFilter(size_t levels, const ScratchImage& mipChain, size_t item)
+    HRESULT Generate2DMipsPointFilter(size_t levels, const ScratchImage& mipChain, size_t item, DWORD flags)
     {
         if (!mipChain.GetImages())
             return E_INVALIDARG;
@@ -706,7 +706,7 @@ namespace
                     sx += xinc;
                 }
 
-                if (!_StoreScanline(pDest, dest->rowPitch, dest->format, target, nwidth))
+                if (!_StoreScanline(pDest, dest->rowPitch, dest->format, target, nwidth, 0.0f, flags))
                     return E_FAIL;
                 pDest += dest->rowPitch;
 
@@ -1386,7 +1386,7 @@ namespace
 
 
     //--- 3D Point Filter ---
-    HRESULT Generate3DMipsPointFilter(size_t depth, size_t levels, const ScratchImage& mipChain)
+    HRESULT Generate3DMipsPointFilter(size_t depth, size_t levels, const ScratchImage& mipChain, DWORD flags)
     {
         if (!depth || !mipChain.GetImages())
             return E_INVALIDARG;
@@ -1460,7 +1460,7 @@ namespace
                             sx += xinc;
                         }
 
-                        if (!_StoreScanline(pDest, dest->rowPitch, dest->format, target, nwidth))
+                        if (!_StoreScanline(pDest, dest->rowPitch, dest->format, target, nwidth, 0.0f, flags))
                             return E_FAIL;
                         pDest += dest->rowPitch;
 
@@ -2658,7 +2658,7 @@ HRESULT DirectX::GenerateMipMaps(
             if (FAILED(hr))
                 return hr;
 
-            hr = Generate2DMipsPointFilter(levels, mipChain, 0);
+            hr = Generate2DMipsPointFilter(levels, mipChain, 0, filter);
             if (FAILED(hr))
                 mipChain.Release();
             return hr;
@@ -2856,7 +2856,7 @@ HRESULT DirectX::GenerateMipMaps(
 
             for (size_t item = 0; item < metadata.arraySize; ++item)
             {
-                hr = Generate2DMipsPointFilter(levels, mipChain, item);
+                hr = Generate2DMipsPointFilter(levels, mipChain, item, filter);
                 if (FAILED(hr))
                     mipChain.Release();
             }
@@ -2978,7 +2978,7 @@ HRESULT DirectX::GenerateMipMaps3D(
         if (FAILED(hr))
             return hr;
 
-        hr = Generate3DMipsPointFilter(depth, levels, mipChain);
+        hr = Generate3DMipsPointFilter(depth, levels, mipChain, filter);
         if (FAILED(hr))
             mipChain.Release();
         return hr;
@@ -3094,7 +3094,7 @@ HRESULT DirectX::GenerateMipMaps3D(
         if (FAILED(hr))
             return hr;
 
-        hr = Generate3DMipsPointFilter(metadata.depth, levels, mipChain);
+        hr = Generate3DMipsPointFilter(metadata.depth, levels, mipChain, filter);
         if (FAILED(hr))
             mipChain.Release();
         return hr;

DirectXTex/DirectXTexMisc.cpp

@@ -364,7 +364,7 @@ HRESULT DirectX::CopyRectangle(
 
         _ConvertScanline(scanline.get(), srcRect.w, dstImage.format, srcImage.format, filter);
 
-        if (!_StoreScanline(pDest, copyD, dstImage.format, scanline.get(), srcRect.w))
+        if (!_StoreScanline(pDest, copyD, dstImage.format, scanline.get(), srcRect.w, 0.0f, filter))
             return E_FAIL;
 
         pSrc += srcImage.rowPitch;

DirectXTex/DirectXTexP.h

@@ -242,7 +242,7 @@ namespace DirectX
 
     _Success_(return != false) bool __cdecl _StoreScanline(
         _Out_writes_bytes_(size) void* pDestination, _In_ size_t size, _In_ DXGI_FORMAT format,
-        _In_reads_(count) const XMVECTOR* pSource, _In_ size_t count, _In_ float threshold = 0);
+        _In_reads_(count) const XMVECTOR* pSource, _In_ size_t count, _In_ float threshold = 0, _In_ DWORD flags = 0);
 
     _Success_(return != false) bool __cdecl _StoreScanlineLinear(
         _Out_writes_bytes_(size) void* pDestination, _In_ size_t size, _In_ DXGI_FORMAT format,