SIMD implementation for XMFloatLoad/Store3SE (#308)

Author: walbourn

.github/workflows/arm64.yml

@@ -62,3 +62,39 @@ jobs:
       - name: 'Build'
         working-directory: ${{ github.workspace }}
         run: cmake --build out/build/${{ matrix.build_type }}
+
+  buildarm32:
+    runs-on: windows-11-arm
+
+    strategy:
+      fail-fast: false
+
+      matrix:
+        build_type: [arm-Debug, arm-Release]
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Clone test repository
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          repository: walbourn/directxmathtest
+          path: Tests
+          ref: main
+
+      - name: 'Install Ninja'
+        run: choco install ninja
+
+      # ARM32 is deprecated, so more recent Windows SDKs no longer support it
+      - uses: ilammy/msvc-dev-cmd@0b201ec74fa43914dc39ae48a89fd1d8cb592756 # v1.13.0
+        with:
+          arch: arm64_arm
+          sdk: 10.0.22621.0
+
+      - name: 'Configure CMake'
+        working-directory: ${{ github.workspace }}
+        run: cmake --preset=${{ matrix.build_type }}
+
+      - name: 'Build'
+        working-directory: ${{ github.workspace }}
+        run: cmake --build out/build/${{ matrix.build_type }}

CMakePresets.json

@@ -36,6 +36,17 @@
       },
       "hidden": true
     },
+    {
+      "name": "ARM32",
+      "architecture": {
+        "value": "arm",
+        "strategy": "external"
+      },
+      "cacheVariables": {
+        "DXMATH_ARCHITECTURE": "arm"
+      },
+      "hidden": true
+    },
     {
       "name": "ARM64",
       "architecture": {
@@ -205,6 +216,8 @@
     { "name": "arm64-Release"  , "description": "MSVC for ARM64 (Release) - ARM-NEON", "inherits": [ "base", "ARM64", "Release", "MSVC" ] },
     { "name": "arm64ec-Debug"  , "description": "MSVC for ARM64EC (Debug) - ARM-NEON", "inherits": [ "base", "ARM64EC", "Debug", "MSVC" ] },
     { "name": "arm64ec-Release", "description": "MSVC for ARM64EC (Release) - ARM-NEON", "inherits": [ "base", "ARM64EC", "Release", "MSVC" ] },
+    { "name": "arm-Debug"      , "description": "MSVC for ARM32 [Deprecated] (Debug) - ARM-NEON", "inherits": [ "base", "ARM32", "Debug", "MSVC" ] },
+    { "name": "arm-Release"    , "description": "MSVC for ARM32 [Deprecated] (Release) - ARM-NEON", "inherits": [ "base", "ARM32", "Release", "MSVC" ] },
 
     { "name": "x64-Debug-Clang"    , "description": "Clang/LLVM for x64 (Debug) - SSE/SSE2", "inherits": [ "base", "x64", "Debug", "Clang" ] },
     { "name": "x64-Release-Clang"  , "description": "Clang/LLVM for x64 (Release) - SSE/SSE2", "inherits": [ "base", "x64", "Release", "Clang" ] },

Inc/DirectXPackedVector.inl

@@ -1279,6 +1279,8 @@ inline XMVECTOR XM_CALLCONV XMLoadFloat3SE(const XMFLOAT3SE* pSource) noexcept
 {
     assert(pSource);
 
+#if defined(_XM_NO_INTRINSICS_)
+
     union { float f; int32_t i; } fi;
     fi.i = 0x33800000 + (pSource->e << 23);
     float Scale = fi.f;
@@ -1289,6 +1291,41 @@ inline XMVECTOR XM_CALLCONV XMLoadFloat3SE(const XMFLOAT3SE* pSource) noexcept
             Scale * float(pSource->zm),
             1.0f } } };
     return v;
+
+#elif defined(_XM_ARM_NEON_INTRINSICS_)
+
+    // Build scale factor from shared exponent
+    union { float f; int32_t i; } fi;
+    fi.i = 0x33800000 + (pSource->e << 23);
+
+    // Extract 9-bit mantissas into vector lanes
+    uint32x4_t mantissas = vdupq_n_u32(0);
+    mantissas = vsetq_lane_u32(pSource->xm, mantissas, 0);
+    mantissas = vsetq_lane_u32(pSource->ym, mantissas, 1);
+    mantissas = vsetq_lane_u32(pSource->zm, mantissas, 2);
+
+    // Convert to float, scale, and set w = 1.0f
+    float32x4_t result = vmulq_n_f32(vcvtq_f32_u32(mantissas), fi.f);
+    return vsetq_lane_f32(1.0f, result, 3);
+
+#elif defined(_XM_SSE_INTRINSICS_)
+
+    // Build scale factor from shared exponent
+    union { float f; int32_t i; } fi;
+    fi.i = 0x33800000 + (pSource->e << 23);
+
+    // Extract 9-bit mantissas, convert to float, and scale
+    __m128i mantissas = _mm_set_epi32(
+        0,
+        static_cast<int>(pSource->zm),
+        static_cast<int>(pSource->ym),
+        static_cast<int>(pSource->xm));
+    __m128 result = _mm_mul_ps(_mm_cvtepi32_ps(mantissas), _mm_set1_ps(fi.f));
+
+    // Set w = 1.0f (w lane is +0.0f so bitwise OR inserts 1.0f cleanly)
+    return _mm_or_ps(result, g_XMIdentityR3);
+
+#endif
 }
 
 //------------------------------------------------------------------------------
@@ -2639,6 +2676,8 @@ inline void XM_CALLCONV XMStoreFloat3SE
 {
     assert(pDestination);
 
+#if defined(_XM_NO_INTRINSICS_)
+
     XMFLOAT3A tmp;
     XMStoreFloat3A(&tmp, V);
 
@@ -2667,6 +2706,92 @@ inline void XM_CALLCONV XMStoreFloat3SE
     pDestination->xm = static_cast<uint32_t>(MathInternal::round_to_nearest(x * ScaleR));
     pDestination->ym = static_cast<uint32_t>(MathInternal::round_to_nearest(y * ScaleR));
     pDestination->zm = static_cast<uint32_t>(MathInternal::round_to_nearest(z * ScaleR));
+
+#elif defined(_XM_ARM_NEON_INTRINSICS_)
+
+    static const XMVECTORF32 MaxFloat9 = { { { float(0x1FF << 7), float(0x1FF << 7), float(0x1FF << 7), float(0x1FF << 7) } } };
+    static constexpr float minf9 = float(1.f / (1 << 16));
+
+    // Clamp to [0, maxf9] then zero w lane
+    float32x4_t clamped = vminq_f32(vmaxq_f32(V, vdupq_n_f32(0)), MaxFloat9);
+    clamped = vsetq_lane_f32(0.0f, clamped, 3);
+
+    // Horizontal max of xyz for shared exponent
+#if defined(_M_ARM64) || defined(_M_HYBRID_X86_ARM64) || defined(_M_ARM64EC) || __aarch64__
+    float maxVal = vmaxvq_f32(clamped);
+#else
+    float32x2_t vlow = vget_low_f32(clamped);
+    float32x2_t vhigh = vget_high_f32(clamped);
+    float32x2_t maxPair = vpmax_f32(vlow, vhigh);
+    maxPair = vpmax_f32(maxPair, maxPair);
+    float maxVal = vget_lane_f32(maxPair, 0);
+#endif
+
+    if (maxVal < minf9) maxVal = minf9;
+
+    // Compute shared exponent (inherently scalar)
+    union { float f; int32_t i; } fi;
+    fi.f = maxVal;
+    fi.i += 0x00004000; // round up leaving 9 bits in fraction (including assumed 1)
+
+    auto exp = static_cast<uint32_t>(fi.i) >> 23;
+    fi.i = static_cast<int32_t>(0x83000000 - (exp << 23));
+
+    // Scale all channels and convert to integer
+    float32x4_t scaled = vmulq_n_f32(clamped, fi.f);
+#if defined(_M_ARM64) || defined(_M_HYBRID_X86_ARM64) || defined(_M_ARM64EC) || __aarch64__
+    uint32x4_t ints = vcvtnq_u32_f32(scaled);
+#else
+    scaled = vaddq_f32(scaled, vdupq_n_f32(0.5f));
+    uint32x4_t ints = vcvtq_u32_f32(scaled);
+#endif
+
+    // Extract and pack into bitfields
+    pDestination->xm = vgetq_lane_u32(ints, 0) & 0x1FF;
+    pDestination->ym = vgetq_lane_u32(ints, 1) & 0x1FF;
+    pDestination->zm = vgetq_lane_u32(ints, 2) & 0x1FF;
+    pDestination->e = exp - 0x6f;
+
+#elif defined(_XM_SSE_INTRINSICS_)
+
+    static const XMVECTORF32 MaxFloat9 = { { { float(0x1FF << 7), float(0x1FF << 7), float(0x1FF << 7), float(0x1FF << 7) } } };
+    static constexpr float minf9 = float(1.f / (1 << 16));
+
+    // Clamp to [0, maxf9] then mask w to zero
+    __m128 clamped = _mm_min_ps(_mm_max_ps(V, _mm_setzero_ps()), MaxFloat9);
+    clamped = _mm_and_ps(clamped, g_XMMask3);
+
+    // Horizontal max of xyz for shared exponent
+    __m128 maxV = clamped;
+    __m128 temp = XM_PERMUTE_PS(maxV, _MM_SHUFFLE(1, 1, 1, 1));
+    maxV = _mm_max_ps(maxV, temp);
+    temp = XM_PERMUTE_PS(clamped, _MM_SHUFFLE(2, 2, 2, 2));
+    maxV = _mm_max_ps(maxV, temp);
+
+    // Ensure minimum threshold
+    maxV = _mm_max_ss(maxV, _mm_set_ss(minf9));
+
+    // Compute shared exponent (inherently scalar)
+    union { float f; int32_t i; } fi;
+    _mm_store_ss(&fi.f, maxV);
+    fi.i += 0x00004000; // round up leaving 9 bits in fraction (including assumed 1)
+
+    auto exp = static_cast<uint32_t>(fi.i) >> 23;
+    fi.i = static_cast<int32_t>(0x83000000 - (exp << 23));
+
+    // Scale all channels and round to nearest integer
+    __m128 scaled = _mm_mul_ps(clamped, _mm_set1_ps(fi.f));
+    __m128i ints = _mm_cvtps_epi32(scaled);
+
+    // Extract and pack into bitfields
+    XM_ALIGNED_DATA(16) uint32_t ivals[4];
+    _mm_store_si128(reinterpret_cast<__m128i*>(ivals), ints);
+
+    pDestination->xm = ivals[0] & 0x1FF;
+    pDestination->ym = ivals[1] & 0x1FF;
+    pDestination->zm = ivals[2] & 0x1FF;
+    pDestination->e = exp - 0x6f;
+#endif
 }
 
 //------------------------------------------------------------------------------

README.md

@@ -93,6 +93,8 @@ For a full change history, see [CHANGELOG.md](https://github.com/microsoft/Direc
 
 * The clang/LLVM toolset currently does not respect the ``float_control`` pragma for SSE instrinsics. Therefore, the use of ``/fp:fast`` is not recommended on clang/LLVM until this issue is fixed. See [55713](https://github.com/llvm/llvm-project/issues/55713).
 
+* AArch32/ARM32 (ARMv7) support is deprecated in Windows 11. Compiler support for ARM32 is deprecated in Visual Studio 2026, and the system libraries are no longer present in the Windows SDK (26100) or later. Therefore, support for ARM32 is deprecated in DirectXMath and will be removed in a future release. Since most codepaths are shared for AArch32 and AArch64, the codepaths will be refactored to assume AArch64 (ARMv8).
+
 ## Support
 
 For questions, consider using [Stack Overflow](https://stackoverflow.com/questions/tagged/directxmath) with the *directxmath* tag, or the [DirectX Discord Server](https://discord.gg/directx) in the *dx12-developers* or *dx9-dx11-developers* channel.