moar changes

mac implementation had _mm_div (x86 intrinsic)

Engine/source/math/impl/float4_impl.inl

@@ -1,5 +1,6 @@
 #pragma once
 #include <cmath>   // for sqrtf, etc.
+#include "../mConstants.h"
 
 namespace math_backend::float4
 {
@@ -46,7 +47,7 @@ namespace math_backend::float4
    {
       f32x4 va = v_load(a);
       f32x4 vb = v_load(b);
-      f32x4 vr = _mm_div_ps(va, vb);
+      f32x4 vr = v_div(va, vb);
       v_store(r, vr);
    }
 
@@ -84,7 +85,7 @@ namespace math_backend::float4
    inline void float4_normalize_impl(float* a)
    {
       float len = float4_length_impl(a);
-      if (len > 1e-6f) // safe threshold
+      if (len > POINT_EPSILON) // safe threshold
       {
          float4_mul_scalar_impl(a, 1.0f / len, a);
       }
@@ -94,7 +95,7 @@ namespace math_backend::float4
    inline void float4_normalize_mag_impl(float* a, float r)
    {
       float len = float4_length_impl(a);
-      if (len > 1e-6f)
+      if (len > POINT_EPSILON)
       {
          float4_mul_scalar_impl(a, r / len, a);
       }

Engine/source/math/isa/avx/float4.cpp

@@ -0,0 +1,67 @@
+
+#include "float4_dispatch.h"
+#include <immintrin.h> // AVX/AVX2 intrinsics
+
+namespace
+{
+   typedef __m128 f32x4;
+
+   // Load 4 floats from memory into a SIMD register
+   inline f32x4 v_load(const float* p) { return _mm_loadu_ps(p); }
+
+   // Store 4 floats from SIMD register back to memory
+   inline void v_store(float* dst, f32x4 v) { _mm_storeu_ps(dst, v); }
+
+   // Broadcast a single float across all 4 lanes
+   inline f32x4 v_set1(float s) { return _mm_set1_ps(s); }
+
+   // Element-wise multiply
+   inline f32x4 v_mul(f32x4 a, f32x4 b) { return _mm_mul_ps(a, b); }
+
+   // Element-wise divide
+   inline f32x4 v_div(f32x4 a, f32x4 b) { return _mm_div_ps(a, b); }
+
+   // Element-wise add
+   inline f32x4 v_add(f32x4 a, f32x4 b) { return _mm_add_ps(a, b); }
+
+   // Element-wise subtract
+   inline f32x4 v_sub(f32x4 a, f32x4 b) { return _mm_sub_ps(a, b); }
+
+   // Horizontal sum of all 4 elements (for dot product, length, etc.)
+   inline float v_hadd4(f32x4 a)
+   {
+      __m128 t1 = _mm_hadd_ps(a, a);  // sums pairs: [a0+a1, a2+a3, ...]
+      __m128 t2 = _mm_hadd_ps(t1, t1); // sums again: first element = a0+a1+a2+a3
+      return _mm_cvtss_f32(t2);       // extract first element
+   }
+
+   // specialized dot product for AVX
+   float float4_dot_avx(const float* a, const float* b)
+   {
+      f32x4 va = _mm_loadu_ps(a);
+      f32x4 vb = _mm_loadu_ps(b);
+      __m128 dp = _mm_dp_ps(va, vb, 0xF1); // multiply all 4, sum all 4, lowest lane
+      return _mm_cvtss_f32(dp);
+   }
+}
+
+#include "float4_impl.inl"
+
+namespace math_backend::float4::dispatch
+{
+   // Install AVX backend
+   void install_avx()
+   {
+      gFloat4.add = float4_add_impl;
+      gFloat4.sub = float4_sub_impl;
+      gFloat4.mul = float4_mul_impl;
+      gFloat4.mul_scalar = float4_mul_scalar_impl;
+      gFloat4.div = float4_div_impl;
+      gFloat4.div_scalar = float4_div_scalar_impl;
+      gFloat4.dot = float4_dot_avx;
+      gFloat4.length = float4_length_impl;
+      gFloat4.lengthSquared = float4_length_squared_impl;
+      gFloat4.normalize = float4_normalize_impl;
+      gFloat4.lerp = float4_lerp_impl;
+   }
+}

Engine/source/math/isa/avx2/float4.cpp

@@ -34,6 +34,15 @@ namespace
       __m128 t2 = _mm_hadd_ps(t1, t1); // sums again: first element = a0+a1+a2+a3
       return _mm_cvtss_f32(t2);       // extract first element
    }
+
+   // specialized dot product for AVX
+   float float4_dot_avx(const float* a, const float* b)
+   {
+      f32x4 va = _mm_loadu_ps(a);
+      f32x4 vb = _mm_loadu_ps(b);
+      __m128 dp = _mm_dp_ps(va, vb, 0xF1); // multiply all 4, sum all 4, lowest lane
+      return _mm_cvtss_f32(dp);
+   }
 }
 
 #include "float4_impl.inl"
@@ -49,7 +58,7 @@ namespace math_backend::float4::dispatch
       gFloat4.mul_scalar   = float4_mul_scalar_impl;
       gFloat4.div          = float4_div_impl;
       gFloat4.div_scalar   = float4_div_scalar_impl;
-      gFloat4.dot          = float4_dot_impl;
+      gFloat4.dot          = float4_dot_avx;
       gFloat4.length       = float4_length_impl;
       gFloat4.lengthSquared = float4_length_squared_impl;
       gFloat4.normalize    = float4_normalize_impl;

Engine/source/math/isa/sse2/float4.cpp

@@ -40,19 +40,19 @@ namespace
 
 namespace math_backend::float4::dispatch
 {
-   // Install AVX2 backend
+   // Install SSE2 backend
    void install_sse2()
    {
-      gFloat4.add = float4_add_impl;
+      gFloat4.add          = float4_add_impl;
-      gFloat4.sub = float4_sub_impl;
+      gFloat4.sub          = float4_sub_impl;
-      gFloat4.mul = float4_mul_impl;
+      gFloat4.mul          = float4_mul_impl;
-      gFloat4.mul_scalar = float4_mul_scalar_impl;
+      gFloat4.mul_scalar   = float4_mul_scalar_impl;
-      gFloat4.div = float4_div_impl;
+      gFloat4.div          = float4_div_impl;
-      gFloat4.div_scalar = float4_div_scalar_impl;
+      gFloat4.div_scalar   = float4_div_scalar_impl;
-      gFloat4.dot = float4_dot_impl;
+      gFloat4.dot          = float4_dot_impl;
-      gFloat4.length = float4_length_impl;
+      gFloat4.length       = float4_length_impl;
       gFloat4.lengthSquared = float4_length_squared_impl;
-      gFloat4.normalize = float4_normalize_impl;
+      gFloat4.normalize    = float4_normalize_impl;
-      gFloat4.lerp = float4_lerp_impl;
+      gFloat4.lerp         = float4_lerp_impl;
    }
 }

Engine/source/math/isa/sse41/float4.cpp

@@ -0,0 +1,67 @@
+
+#include "float4_dispatch.h"
+#include <smmintrin.h> // SSE41 intrinsics
+
+namespace
+{
+   typedef __m128 f32x4;
+
+   // Load 4 floats from memory into a SIMD register
+   inline f32x4 v_load(const float* p) { return _mm_loadu_ps(p); }
+
+   // Store 4 floats from SIMD register back to memory
+   inline void v_store(float* dst, f32x4 v) { _mm_storeu_ps(dst, v); }
+
+   // Broadcast a single float across all 4 lanes
+   inline f32x4 v_set1(float s) { return _mm_set1_ps(s); }
+
+   // Element-wise multiply
+   inline f32x4 v_mul(f32x4 a, f32x4 b) { return _mm_mul_ps(a, b); }
+
+   // Element-wise divide
+   inline f32x4 v_div(f32x4 a, f32x4 b) { return _mm_div_ps(a, b); }
+
+   // Element-wise add
+   inline f32x4 v_add(f32x4 a, f32x4 b) { return _mm_add_ps(a, b); }
+
+   // Element-wise subtract
+   inline f32x4 v_sub(f32x4 a, f32x4 b) { return _mm_sub_ps(a, b); }
+
+   // Horizontal sum of all 4 elements (for dot product, length, etc.)
+   inline float v_hadd4(f32x4 a)
+   {
+      __m128 t1 = _mm_hadd_ps(a, a);  // sums pairs: [a0+a1, a2+a3, ...]
+      __m128 t2 = _mm_hadd_ps(t1, t1); // sums again: first element = a0+a1+a2+a3
+      return _mm_cvtss_f32(t2);       // extract first element
+   }
+
+   // specialized dot product for SSE4.1
+   float float4_dot_sse41(const float* a, const float* b)
+   {
+      f32x4 va = _mm_loadu_ps(a);
+      f32x4 vb = _mm_loadu_ps(b);
+      __m128 dp = _mm_dp_ps(va, vb, 0xF1); // multiply all 4, sum all 4, lowest lane
+      return _mm_cvtss_f32(dp);
+   }
+}
+
+#include "float4_impl.inl"
+
+namespace math_backend::float4::dispatch
+{
+   // Install SSE41 backend
+   void install_sse41()
+   {
+      gFloat4.add          = float4_add_impl;
+      gFloat4.sub          = float4_sub_impl;
+      gFloat4.mul          = float4_mul_impl;
+      gFloat4.mul_scalar   = float4_mul_scalar_impl;
+      gFloat4.div          = float4_div_impl;
+      gFloat4.div_scalar   = float4_div_scalar_impl;
+      gFloat4.dot          = float4_dot_sse41;
+      gFloat4.length       = float4_length_impl;
+      gFloat4.lengthSquared = float4_length_squared_impl;
+      gFloat4.normalize    = float4_normalize_impl;
+      gFloat4.lerp         = float4_lerp_impl;
+   }
+}

Engine/source/math/public/float4_dispatch.h

@@ -1,4 +1,8 @@
 #pragma once
+#ifndef _FLOAT4_DISPATCH_H_
+#define _FLOAT4_DISPATCH_H_
+
+
 #include <cstdint>
 
 namespace math_backend::float4::dispatch
@@ -32,3 +36,5 @@ namespace math_backend::float4::dispatch
    // Centralized installer (engine calls this once)
    void install_preferred();
 }
+
+#endif // !_FLOAT4_DISPATCH_H_

Engine/source/math/public/math_backend.cpp

@@ -0,0 +1,53 @@
+#pragma once
+#include "math/public/math_backend.h"
+
+math_backend::backend math_backend::choose_backend(U32 cpu_flags)
+{
+#if defined(__x86_64__) || defined(_M_X64) || defined(_M_IX86)
+
+   if (cpu_flags & CPU_PROP_AVX2)  return backend::avx2;
+   if (cpu_flags & CPU_PROP_AVX)   return backend::avx;
+   if (cpu_flags & CPU_PROP_SSE4_1) return backend::sse41;
+   if (cpu_flags & CPU_PROP_SSE2)  return backend::sse2;
+
+#elif defined(__aarch64__) || defined(__ARM_NEON)
+
+   if (cpu_flags & CPU_NEON) return backend::neon;
+
+#endif
+   return backend::scalar;
+}
+
+void math_backend::install_from_cpu_flags(uint32_t cpu_flags)
+{
+   {
+      g_backend = choose_backend(cpu_flags);
+
+      switch (g_backend)
+      {
+      case backend::avx2:
+         float4::dispatch::install_avx2();
+         break;
+
+      case backend::avx:
+         //float4::dispatch::install_avx();
+         break;
+
+      case backend::sse41:
+         float4::dispatch::install_sse41();
+         break;
+
+      case backend::sse2:
+         float4::dispatch::install_sse2();
+         break;
+
+      case backend::neon:
+         float4::dispatch::install_neon();
+         break;
+
+      default:
+         float4::dispatch::install_scalar();
+         break;
+      }
+   }
+}

Engine/source/math/public/math_backend.h

@@ -0,0 +1,27 @@
+#pragma once
+#ifndef _MCONSTANTS_H_
+#include "math/mConstants.h"
+#endif
+#ifndef _PLATFORMASSERT_H_
+#include "platform/platformAssert.h"
+#endif
+#ifndef _FLOAT4_DISPATCH_H_
+#include "math/public/float4_dispatch.h"
+#endif
+
+namespace math_backend
+{
+   enum class backend
+   {
+      scalar,
+      sse2,
+      sse41,
+      avx,
+      avx2,
+      neon
+   };
+
+   static backend g_backend = backend::scalar;
+   backend choose_backend(U32 cpu_flags);
+   void install_from_cpu_flags(uint32_t cpu_flags);
+}

Tools/CMake/torque_macros.cmake

@@ -154,7 +154,9 @@ function(add_math_backend name compile_defs)
 
     # ISA flags
     if(MSVC)
-        if(name STREQUAL "sse2" OR name STREQUAL "sse41")
+        if(name STREQUAL "sse2")
+            target_compile_options(math_${name} PRIVATE /arch:SSE2)
+        elseif(name STREQUAL "sse41")
             target_compile_options(math_${name} PRIVATE /arch:SSE2)
         elseif(name STREQUAL "avx")
             target_compile_options(math_${name} PRIVATE /arch:AVX)