Merge pull request #445 from zrax/avx2

Support AVX2

Sources/Plasma/CoreLib/hsCpuID.cpp

@@ -59,16 +59,31 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
 #include "hsCpuID.h"
 
 hsCpuId::hsCpuId() {
-    const unsigned int sse1_flag = 1<<25;
+    enum : unsigned int {
-    const unsigned int sse2_flag = 1<<26;
+        // EAX=1; EDX=:
-    const unsigned int sse3_flag = 1<<0;
+        sse1_flag  = 1U<<25,
-    const unsigned int ssse3_flag = 1<<9;
+        sse2_flag  = 1U<<26,
-    const unsigned int sse41_flag = 1<<19;
-    const unsigned int sse42_flag = 1<<20;
-    const unsigned int avx_flag = 1<<28;
 
-    unsigned int ax = 0, bx = 0, cx = 0, dx = 0;
+        // EAX=1; ECX=:
+        sse3_flag  = 1U<<0,
+        ssse3_flag = 1U<<9,
+        sse41_flag = 1U<<19,
+        sse42_flag = 1U<<20,
+        avx_flag   = 1U<<28,
 
+        // EAX=7; ECX=0; EBX=:
+        avx2_flag  = 1U<<5
+    };
+
+    union RegSet {
+        struct {
+            unsigned int eax, ebx, ecx, edx;
+        };
+        int array[4];
+    };
+
+    RegSet CPUInfo_Features = { 0, 0, 0, 0 };
+    RegSet CPUInfo_Ext = { 0, 0, 0, 0 };
 
     /**
      * Portable implementation of CPUID, successfully tested with:
@@ -80,33 +95,33 @@ hsCpuId::hsCpuId() {
      *
      * Ref: http://primesieve.googlecode.com/svn-history/r388/trunk/soe/cpuid.h
      */
-    #if defined(MSC_COMPATIBLE)
+#if defined(MSC_COMPATIBLE)
-      int CPUInfo[4] = {ax, bx, cx, dx};
+    __cpuid(CPUInfo_Features.array, 0);
-      __cpuid(CPUInfo, 0);
 
     // check if the CPU supports the cpuid instruction.
-      if (CPUInfo[0] != 0) {
+    if (CPUInfo_Features.eax != 0) {
-        __cpuid(CPUInfo, 1);
+        __cpuid(CPUInfo_Features.array, 1);
-        ax = CPUInfo[0];
+        __cpuid(CPUInfo_Ext.array, 7);
-        bx = CPUInfo[1];
-        cx = CPUInfo[2];
-        dx = CPUInfo[3];
     }
-    #elif defined(GCC_COMPATIBLE)
+#elif defined(GCC_COMPATIBLE)
-      __get_cpuid(1, &ax, &bx, &cx, &dx);
+    __get_cpuid(1, &CPUInfo_Features.eax, &CPUInfo_Features.ebx,
-    #endif
+                   &CPUInfo_Features.ecx, &CPUInfo_Features.edx);
+    __get_cpuid(7, &CPUInfo_Ext.eax, &CPUInfo_Ext.ebx,
+                   &CPUInfo_Ext.ecx, &CPUInfo_Ext.edx);
+#endif
 
 
-    has_sse1    = (dx & sse1_flag)  || false;
+    has_sse1    = (CPUInfo_Features.edx & sse1_flag)  || false;
-    has_sse2    = (dx & sse2_flag)  || false;
+    has_sse2    = (CPUInfo_Features.edx & sse2_flag)  || false;
-    has_sse3    = (cx & sse3_flag)  || false;
+    has_sse3    = (CPUInfo_Features.ecx & sse3_flag)  || false;
-    has_ssse3   = (cx & ssse3_flag) || false;
+    has_ssse3   = (CPUInfo_Features.ecx & ssse3_flag) || false;
-    has_sse41   = (cx & sse41_flag) || false;
+    has_sse41   = (CPUInfo_Features.ecx & sse41_flag) || false;
-    has_sse42   = (cx & sse42_flag) || false;
+    has_sse42   = (CPUInfo_Features.ecx & sse42_flag) || false;
-    has_avx     = (cx & avx_flag)   || false;
+    has_avx     = (CPUInfo_Features.ecx & avx_flag)   || false;
+    has_avx2    = (CPUInfo_Ext.ebx      & avx2_flag)  || false;
 }
 
-const hsCpuId& hsCpuId::instance()
+const hsCpuId& hsCpuId::Instance()
 {
     static hsCpuId self;
     return self;

Sources/Plasma/CoreLib/hsCpuID.h

@@ -65,7 +65,8 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
 //  typedef float(*func_ptr)();
 //  static hsFunctionDispatcher<func_ptr> my_func;
 //
-//  hsFunctionDispatcher<float::func_ptr> float::my_func(float::my_func_fpu, 0, 0, 0, 0, 0, 0, float::my_func_avx);
+//  hsFunctionDispatcher<func_ptr> my_func(my_func_fpu, 0, 0, 0, 0, 0, 0, my_func_avx);
+//  my_func();
 //
 //////////////////////////////////////////////////////////////////////
 
@@ -74,6 +75,12 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
 #ifndef hsCpuID_inc
 #define hsCpuID_inc
 
+#if defined __AVX2__ || _MSC_VER >= 1600
+#define HS_AVX2
+#ifndef HS_SIMD_INCLUDE
+# define HS_SIMD_INCLUDE "immintrin.h"
+#endif
+#endif
 #if defined __AVX__ || _MSC_VER >= 1600
 #define HS_AVX
 #ifndef HS_SIMD_INCLUDE
@@ -126,16 +133,31 @@ struct hsCpuId {
     bool has_sse41;
     bool has_sse42;
     bool has_avx;
+    bool has_avx2;
 
     hsCpuId();
-    static const hsCpuId& instance();
+    static const hsCpuId& Instance();
 };
 
 template <typename func_ptr>
-struct hsFunctionDispatcher {
+struct hsCpuFunctionDispatcher {
-    hsFunctionDispatcher(func_ptr fpu, func_ptr sse1=0, func_ptr sse2=0, func_ptr sse3=0, func_ptr ssse3=0, func_ptr sse41=0, func_ptr sse42=0, func_ptr avx=0) {
+    hsCpuFunctionDispatcher(func_ptr fpu,
+                            func_ptr sse1 = nullptr,
+                            func_ptr sse2 = nullptr,
+                            func_ptr sse3 = nullptr,
+                            func_ptr ssse3 = nullptr,
+                            func_ptr sse41 = nullptr,
+                            func_ptr sse42 = nullptr,
+                            func_ptr avx = nullptr,
+                            func_ptr avx2 = nullptr)
+    {
         hsAssert(fpu, "FPU fallback function required.");
-        const hsCpuId& cpu = hsCpuId::instance();
+        const hsCpuId& cpu = hsCpuId::Instance();
+#ifdef HS_AVX2
+        if (cpu.has_avx2 && avx2) {
+            call = avx2;
+        } else
+#endif
 #ifdef HS_AVX
         if (cpu.has_avx && avx) {
             call = avx;

Sources/Plasma/CoreLib/hsMatrix44.cpp

@@ -57,9 +57,6 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
 static hsMatrix44 myIdent = hsMatrix44().Reset();
 const hsMatrix44& hsMatrix44::IdentityMatrix() { return myIdent; }
 
-// CPU-optimized functions requiring dispatch
-hsFunctionDispatcher<hsMatrix44::mat_mult_ptr> hsMatrix44::mat_mult(hsMatrix44::mat_mult_fpu, 0, 0, hsMatrix44::mat_mult_sse3);
-
 /*
     For the rotation:
          <20>        2     2                                      <20>
@@ -102,7 +99,7 @@ void hsMatrix44::DecompRigid(hsScalarTriple &translate, hsQuat &rotate) const
     rotate.QuatFromMatrix44(*this);
 }
 
-hsMatrix44 hsMatrix44::mat_mult_fpu(const hsMatrix44 &a, const hsMatrix44 &b)
+static hsMatrix44 mat_mult_fpu(const hsMatrix44 &a, const hsMatrix44 &b)
 {
     hsMatrix44 c;
 
@@ -153,7 +150,7 @@ hsMatrix44 hsMatrix44::mat_mult_fpu(const hsMatrix44 &a, const hsMatrix44 &b)
         _mm_storeu_ps(c.fMap[i], xmm[1]);
 #endif  // HS_SSE3
 
-hsMatrix44 hsMatrix44::mat_mult_sse3(const hsMatrix44 &a, const hsMatrix44 &b)
+static hsMatrix44 mat_mult_sse3(const hsMatrix44 &a, const hsMatrix44 &b)
 {
     hsMatrix44 c;
 #ifdef HS_SSE3
@@ -201,6 +198,26 @@ hsMatrix44 hsMatrix44::mat_mult_sse3(const hsMatrix44 &a, const hsMatrix44 &b)
     return c;
 }
 
+// CPU-optimized functions requiring dispatch
+hsCpuFunctionDispatcher<hsMatrix44::mat_mult_ptr> hsMatrix44::mat_mult {
+    &mat_mult_fpu,
+    nullptr,            // SSE1
+    nullptr,            // SSE2
+    &mat_mult_sse3
+};
+
+hsPoint3 hsMatrix44::operator*(const hsPoint3& p) const
+{
+    if (fFlags & hsMatrix44::kIsIdent)
+        return p;
+
+    hsPoint3 rVal;
+    rVal.fX = (p.fX * fMap[0][0]) + (p.fY * fMap[0][1]) + (p.fZ * fMap[0][2]) + fMap[0][3];
+    rVal.fY = (p.fX * fMap[1][0]) + (p.fY * fMap[1][1]) + (p.fZ * fMap[1][2]) + fMap[1][3];
+    rVal.fZ = (p.fX * fMap[2][0]) + (p.fY * fMap[2][1]) + (p.fZ * fMap[2][2]) + fMap[2][3];
+    return rVal;
+}
+
 hsVector3 hsMatrix44::operator*(const hsVector3& p) const
 {
     if( fFlags & hsMatrix44::kIsIdent )

Sources/Plasma/CoreLib/hsMatrix44.h

@@ -133,17 +133,7 @@ struct hsMatrix44 {
     void MakeZRotation(float radians);
 
 
-    hsPoint3        operator*(const hsPoint3& p) const
+    hsPoint3 operator*(const hsPoint3& p) const;
-                    {   
-                        if( fFlags & hsMatrix44::kIsIdent )
-                            return p;
-
-                        hsPoint3 rVal;
-                        rVal.fX = (p.fX * fMap[0][0]) + (p.fY * fMap[0][1]) + (p.fZ * fMap[0][2]) + fMap[0][3];
-                        rVal.fY = (p.fX * fMap[1][0]) + (p.fY * fMap[1][1]) + (p.fZ * fMap[1][2])  + fMap[1][3];
-                        rVal.fZ = (p.fX * fMap[2][0]) + (p.fY * fMap[2][1]) + (p.fZ * fMap[2][2])  + fMap[2][3];
-                        return rVal;
-                    }
     hsVector3 operator*(const hsVector3& p) const;
     hsMatrix44 operator *(const hsMatrix44& other) const { return mat_mult.call(*this, other); }
     
@@ -160,9 +150,7 @@ struct hsMatrix44 {
 
     //  CPU-optimized functions
     typedef hsMatrix44(*mat_mult_ptr)(const hsMatrix44&, const hsMatrix44&);
-    static hsMatrix44 mat_mult_fpu(const hsMatrix44&, const hsMatrix44&);
+    static hsCpuFunctionDispatcher<mat_mult_ptr> mat_mult;
-    static hsMatrix44 mat_mult_sse3(const hsMatrix44&, const hsMatrix44&);
-    static hsFunctionDispatcher<mat_mult_ptr> mat_mult;
 };
 
 ////////////////////////////////////////////////////////////////////////////

Sources/Plasma/PubUtilLib/plPipeline/DX/plDXPipeline.cpp

@@ -10714,9 +10714,14 @@ static void IBlendVertBuffer(plSpan* span, hsMatrix44* matrixPalette, int numMat
 }
 
 // CPU-optimized functions requiring dispatch
-hsFunctionDispatcher<plDXPipeline::blend_vert_buffer_ptr> plDXPipeline::blend_vert_buffer(
+hsCpuFunctionDispatcher<plDXPipeline::blend_vert_buffer_ptr> plDXPipeline::blend_vert_buffer {
-    IBlendVertBuffer<ISkinVertexFPU>, 0, 0, IBlendVertBuffer<ISkinVertexSSE3>, 0,
+    &IBlendVertBuffer<ISkinVertexFPU>,
-    IBlendVertBuffer<ISkinVertexSSE41>);
+    nullptr,                                // SSE1
+    nullptr,                                // SSE2
+    &IBlendVertBuffer<ISkinVertexSSE3>,
+    nullptr,                                // SSSE3
+    &IBlendVertBuffer<ISkinVertexSSE41>
+};
 
 // ISetPipeConsts //////////////////////////////////////////////////////////////////
 // A shader can request that the pipeline fill in certain constants that are indeterminate

Sources/Plasma/PubUtilLib/plPipeline/DX/plDXPipeline.h

@@ -798,8 +798,10 @@ public:
 
     //  CPU-optimized functions
 protected:
-    typedef void(*blend_vert_buffer_ptr)(plSpan*, hsMatrix44*, int, const uint8_t *, uint8_t , uint32_t, uint8_t *, uint32_t, uint32_t, uint16_t);
+    typedef void(*blend_vert_buffer_ptr)(plSpan*, hsMatrix44*, int, const uint8_t *,
-    static hsFunctionDispatcher<blend_vert_buffer_ptr> blend_vert_buffer;
+                                         uint8_t , uint32_t, uint8_t *, uint32_t,
+                                         uint32_t, uint16_t);
+    static hsCpuFunctionDispatcher<blend_vert_buffer_ptr> blend_vert_buffer;
 };