Micro-optimize float4x4 operations when SIMD is not supported

include/bx/float4x4_t.h

@@ -22,9 +22,19 @@ namespace bx
 	///
 	simd128_t simd_mul(simd128_t _a, const float4x4_t* _b);
 
-	///
+	/// Multiplies two matrices
 	void float4x4_mul(float4x4_t* _result, const float4x4_t* _a, const float4x4_t* _b);
 
+	/// Multiplies two 3x4 affine matrices (i.e. "model" or "world" matrices).
+	/// This function is a micro-optimized version of float4x4_mul() in the case
+	/// when the last row of the both input matrices are (0, 0, 0, 1).
+	void model4x4_mul(float4x4_t* _result, const float4x4_t* _a, const float4x4_t* _b);
+
+	/// Multiplies a 3x4 affine matrix with a general 4x4 matrix.
+	/// This function is a micro-optimized version of float4x4_mul() in the case
+	/// when the last row of the _model input matrix is (0, 0, 0, 1).
+	void model4x4_mul_viewproj4x4(float4x4_t* _result, const float4x4_t* _model, const float4x4_t* _viewProj);
+
 	///
 	void float4x4_transpose(float4x4_t* _result, const float4x4_t* _mtx);
 

include/bx/inline/float4x4_t.inl

@@ -40,10 +40,97 @@ namespace bx
 
 	BX_SIMD_INLINE void float4x4_mul(float4x4_t* _result, const float4x4_t* _a, const float4x4_t* _b)
 	{
+#if !BX_CONFIG_SUPPORTS_SIMD
+		const float *a = (const float*)_a;
+		const float *b = (const float*)_b;
+		float *r = (float*)_result;
+		r[0] = a[0]*b[0] + a[1]*b[4] + a[2]*b[8] + a[3]*b[12];
+		r[1] = a[0]*b[1] + a[1]*b[5] + a[2]*b[9] + a[3]*b[13];
+		r[2] = a[0]*b[2] + a[1]*b[6] + a[2]*b[10] + a[3]*b[14];
+		r[3] = a[0]*b[3] + a[1]*b[7] + a[2]*b[11] + a[3]*b[15];
+
+		r[4] = a[4]*b[0] + a[5]*b[4] + a[6]*b[8] + a[7]*b[12];
+		r[5] = a[4]*b[1] + a[5]*b[5] + a[6]*b[9] + a[7]*b[13];
+		r[6] = a[4]*b[2] + a[5]*b[6] + a[6]*b[10] + a[7]*b[14];
+		r[7] = a[4]*b[3] + a[5]*b[7] + a[6]*b[11] + a[7]*b[15];
+
+		r[8] = a[8]*b[0] + a[9]*b[4] + a[10]*b[8] + a[11]*b[12];
+		r[9] = a[8]*b[1] + a[9]*b[5] + a[10]*b[9] + a[11]*b[13];
+		r[10] = a[8]*b[2] + a[9]*b[6] + a[10]*b[10] + a[11]*b[14];
+		r[11] = a[8]*b[3] + a[9]*b[7] + a[10]*b[11] + a[11]*b[15];
+
+		r[12] = a[12]*b[0] + a[13]*b[4] + a[14]*b[8] + a[15]*b[12];
+		r[13] = a[12]*b[1] + a[13]*b[5] + a[14]*b[9] + a[15]*b[13];
+		r[14] = a[12]*b[2] + a[13]*b[6] + a[14]*b[10] + a[15]*b[14];
+		r[15] = a[12]*b[3] + a[13]*b[7] + a[14]*b[11] + a[15]*b[15];
+#else
 		_result->col[0] = simd_mul(_a->col[0], _b);
 		_result->col[1] = simd_mul(_a->col[1], _b);
 		_result->col[2] = simd_mul(_a->col[2], _b);
 		_result->col[3] = simd_mul(_a->col[3], _b);
+#endif
+	}
+
+	BX_SIMD_INLINE void model4x4_mul(float4x4_t* _result, const float4x4_t* _a, const float4x4_t* _b)
+	{
+#if !BX_CONFIG_SUPPORTS_SIMD
+		const float *a = (const float*)_a; // a[3]==a[7]==a[11]==0, a[15]=1
+		const float *b = (const float*)_b; // b[3]==b[7]==b[11]==0, b[15]=1
+		float *r = (float*)_result;
+		r[0] = a[0]*b[0] + a[1]*b[4] + a[2]*b[8];
+		r[1] = a[0]*b[1] + a[1]*b[5] + a[2]*b[9];
+		r[2] = a[0]*b[2] + a[1]*b[6] + a[2]*b[10];
+		r[3] = 0.f;
+
+		r[4] = a[4]*b[0] + a[5]*b[4] + a[6]*b[8];
+		r[5] = a[4]*b[1] + a[5]*b[5] + a[6]*b[9];
+		r[6] = a[4]*b[2] + a[5]*b[6] + a[6]*b[10];
+		r[7] = 0.f;
+
+		r[8] = a[8]*b[0] + a[9]*b[4] + a[10]*b[8];
+		r[9] = a[8]*b[1] + a[9]*b[5] + a[10]*b[9];
+		r[10] = a[8]*b[2] + a[9]*b[6] + a[10]*b[10];
+		r[11] = 0.f;
+
+		r[12] = a[12]*b[0] + a[13]*b[4] + a[14]*b[8] + b[12];
+		r[13] = a[12]*b[1] + a[13]*b[5] + a[14]*b[9] + b[13];
+		r[14] = a[12]*b[2] + a[13]*b[6] + a[14]*b[10] + b[14];
+		r[15] = 1.f;
+#else
+		// With SIMD faster to do the general 4x4 form:
+		float4x4_mul(_result, _a, _b);
+#endif
+	}
+
+	BX_SIMD_INLINE void model4x4_mul_viewproj4x4(float4x4_t* _result, const float4x4_t* _model, const float4x4_t* _viewProj)
+	{
+#if !BX_CONFIG_SUPPORTS_SIMD
+		const float *a = (const float*)_model; // a[3]==a[7]==a[11]==0, a[15]=1
+		const float *b = (const float*)_viewProj;
+		float *r = (float*)_result;
+		r[0] = a[0]*b[0] + a[1]*b[4] + a[2]*b[8];
+		r[1] = a[0]*b[1] + a[1]*b[5] + a[2]*b[9];
+		r[2] = a[0]*b[2] + a[1]*b[6] + a[2]*b[10];
+		r[3] = a[0]*b[3] + a[1]*b[7] + a[2]*b[11];
+
+		r[4] = a[4]*b[0] + a[5]*b[4] + a[6]*b[8];
+		r[5] = a[4]*b[1] + a[5]*b[5] + a[6]*b[9];
+		r[6] = a[4]*b[2] + a[5]*b[6] + a[6]*b[10];
+		r[7] = a[4]*b[3] + a[5]*b[7] + a[6]*b[11];
+
+		r[8] = a[8]*b[0] + a[9]*b[4] + a[10]*b[8];
+		r[9] = a[8]*b[1] + a[9]*b[5] + a[10]*b[9];
+		r[10] = a[8]*b[2] + a[9]*b[6] + a[10]*b[10];
+		r[11] = a[8]*b[3] + a[9]*b[7] + a[10]*b[11];
+
+		r[12] = a[12]*b[0] + a[13]*b[4] + a[14]*b[8] + b[12];
+		r[13] = a[12]*b[1] + a[13]*b[5] + a[14]*b[9] + b[13];
+		r[14] = a[12]*b[2] + a[13]*b[6] + a[14]*b[10] + b[14];
+		r[15] = a[12]*b[3] + a[13]*b[7] + a[14]*b[11] + b[15];
+#else
+		// With SIMD faster to do the general 4x4 form:
+		float4x4_mul(_result, _model, _viewProj);
+#endif
 	}
 
 	BX_SIMD_FORCE_INLINE void float4x4_transpose(float4x4_t* _result, const float4x4_t* _mtx)