more intrinsics

add transform plane
added first batch function for mulp to intrinsics
This commit is contained in:
marauder2k7 2026-03-05 18:54:28 +00:00
parent ac6ec05690
commit add7f2a5d7
14 changed files with 710 additions and 113 deletions

View file

@ -57,7 +57,7 @@ namespace
// Shuffle helpers
//------------------------------------------------------
inline f32x4 v_shuffle_mask(f32x4 v1, f32x4 v2, const int x, const int y, const int z, const int w)
inline f32x4 v_shuffle_mask(f32x4 v1, f32x4 v2, int x, int y, int z, int w)
{
f32x4 lo = v1;
f32x4 hi = v2;
@ -160,13 +160,13 @@ namespace
// full dot4
inline f32x4 v_dot4(f32x4 a, f32x4 b)
{
return _mm_dp_ps(a, b, 0xFF); // f32x4, 4 lanes into lane 1
return _mm_dp_ps(a, b, 0xFF); // f32x4, 4 lanes into all lanes
}
// dot3 (ignores w)
inline f32x4 v_dot3(f32x4 a, f32x4 b)
{
return _mm_dp_ps(a, b, 0x7F); // f32x4, 3 last lanes into lane 1
return _mm_dp_ps(a, b, 0x7F); // f32x4, 3 last lanes into all lanes
}
// cross product xyz only.
@ -182,7 +182,7 @@ namespace
inline f32x4 v_normalize3(f32x4 v)
{
const f32x4 zero = _mm_setzero_ps();
const f32x4 zero = _mm_setzero_ps();
const f32x4 fallback = _mm_set_ps(0.0f, 1.0f, 0.0f, 0.0f); // {0,0,1,0}
f32x4 dot = v_dot3(v, v);
@ -405,13 +405,6 @@ namespace
f32x8 w;
};
struct vec3_batch8
{
f32x8 x;
f32x8 y;
f32x8 z;
};
struct vec4_batch4
{
f32x4 x;
@ -420,47 +413,37 @@ namespace
f32x4 w;
};
struct vec3_batch4
inline vec4_batch8 load_vec3_batch8(const float* ptr, float w, bool fillW)
{
f32x4 x;
f32x4 y;
f32x4 z;
};
vec4_batch8 r;
inline vec3_batch8 load_vec3_batch8(const float* ptr)
{
// ptr points to x0
// layout: x y z x y z ...
r.x = _mm256_set_ps(ptr[0], ptr[3], ptr[6], ptr[9],
ptr[12], ptr[15], ptr[18], ptr[21]);
r.y = _mm256_set_ps(ptr[1], ptr[4], ptr[7], ptr[10],
ptr[13], ptr[16], ptr[19], ptr[22]);
r.z = _mm256_set_ps(ptr[2], ptr[5], ptr[8], ptr[11],
ptr[14], ptr[17], ptr[20], ptr[23]);
vec3_batch8 r;
r.x = _mm256_set_ps(
ptr[21], ptr[18], ptr[15], ptr[12],
ptr[9], ptr[6], ptr[3], ptr[0]);
r.y = _mm256_set_ps(
ptr[22], ptr[19], ptr[16], ptr[13],
ptr[10], ptr[7], ptr[4], ptr[1]);
r.z = _mm256_set_ps(
ptr[23], ptr[20], ptr[17], ptr[14],
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm256_set1_ps(w);
}
return r;
}
inline vec3_batch4 load_vec3_batch4(const float* ptr)
inline vec4_batch4 load_vec3_batch4(const float* ptr, float w, bool fillW)
{
vec3_batch4 r;
vec4_batch4 r;
r.x = _mm_set_ps(
ptr[9], ptr[6], ptr[3], ptr[0]);
r.x = _mm_set_ps(ptr[0], ptr[3], ptr[6], ptr[9]);
r.y = _mm_set_ps(ptr[1], ptr[4], ptr[7], ptr[10]);
r.z = _mm_set_ps(ptr[2], ptr[5], ptr[8], ptr[11]);
r.y = _mm_set_ps(
ptr[10], ptr[7], ptr[4], ptr[1]);
r.z = _mm_set_ps(
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm_set1_ps(w);
}
return r;
}
@ -478,7 +461,7 @@ namespace
}
// Store the result back to a float3 array with size of 8
inline void store_vec3_batch8(float* out, const vec3_batch8& v)
inline void store_vec3_batch8(float* out, const vec4_batch8& v)
{
alignas(32) float xs[8];
alignas(32) float ys[8];
@ -497,7 +480,7 @@ namespace
}
// Store the result back to a float3 array with size of 4
inline void store_vec3_batch4(float* out, const vec3_batch4& v)
inline void store_vec3_batch4(float* out, const vec4_batch4& v)
{
alignas(16) float xs[8];
alignas(16) float ys[8];
@ -515,7 +498,7 @@ namespace
}
}
inline f32x4 vec3_batch4_dot(const vec3_batch4& a, const vec3_batch4& b)
inline f32x4 vec3_batch4_dot(const vec4_batch4& a, const vec4_batch4& b)
{
f32x4 mulx = _mm_mul_ps(a.x, b.x);
f32x4 muly = _mm_mul_ps(a.y, b.y);
@ -524,7 +507,7 @@ namespace
return _mm_add_ps(_mm_add_ps(mulx, muly), mulz);
}
inline f32x8 vec3_batch8_dot(const vec3_batch8& a, const vec3_batch8& b)
inline f32x8 vec3_batch8_dot(const vec4_batch8& a, const vec4_batch8& b)
{
f32x8 mulx = _mm256_mul_ps(a.x, b.x);
f32x8 muly = _mm256_mul_ps(a.y, b.y);
@ -533,4 +516,110 @@ namespace
return _mm256_add_ps(_mm256_add_ps(mulx, muly), mulz);
}
// Batch 8 mul_Vec4.
inline vec4_batch8 m_mul_pos3_batch8(const f32x4x4& m, const vec4_batch8& v)
{
vec4_batch8 r;
__m256 m00 = _mm256_set1_ps(m.r0.m128_f32[0]);
__m256 m01 = _mm256_set1_ps(m.r0.m128_f32[1]);
__m256 m02 = _mm256_set1_ps(m.r0.m128_f32[2]);
__m256 m03 = _mm256_set1_ps(m.r0.m128_f32[3]);
__m256 m10 = _mm256_set1_ps(m.r1.m128_f32[0]);
__m256 m11 = _mm256_set1_ps(m.r1.m128_f32[1]);
__m256 m12 = _mm256_set1_ps(m.r1.m128_f32[2]);
__m256 m13 = _mm256_set1_ps(m.r1.m128_f32[3]);
__m256 m20 = _mm256_set1_ps(m.r2.m128_f32[0]);
__m256 m21 = _mm256_set1_ps(m.r2.m128_f32[1]);
__m256 m22 = _mm256_set1_ps(m.r2.m128_f32[2]);
__m256 m23 = _mm256_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m00),
_mm256_mul_ps(v.y, m01)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m10),
_mm256_mul_ps(v.y, m11)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m20),
_mm256_mul_ps(v.y, m21)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m22), m23));
return r;
}
// Batch 4 mul_Vec4.
inline vec4_batch4 m_mul_pos3_batch4(const f32x4x4& m, const vec4_batch4& v)
{
vec4_batch4 r;
f32x4 m00 = _mm_set1_ps(m.r0.m128_f32[0]);
f32x4 m01 = _mm_set1_ps(m.r0.m128_f32[1]);
f32x4 m02 = _mm_set1_ps(m.r0.m128_f32[2]);
f32x4 m03 = _mm_set1_ps(m.r0.m128_f32[3]);
f32x4 m10 = _mm_set1_ps(m.r1.m128_f32[0]);
f32x4 m11 = _mm_set1_ps(m.r1.m128_f32[1]);
f32x4 m12 = _mm_set1_ps(m.r1.m128_f32[2]);
f32x4 m13 = _mm_set1_ps(m.r1.m128_f32[3]);
f32x4 m20 = _mm_set1_ps(m.r2.m128_f32[0]);
f32x4 m21 = _mm_set1_ps(m.r2.m128_f32[1]);
f32x4 m22 = _mm_set1_ps(m.r2.m128_f32[2]);
f32x4 m23 = _mm_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m00),
_mm_mul_ps(v.y, m01)),
_mm_add_ps(
_mm_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m10),
_mm_mul_ps(v.y, m11)),
_mm_add_ps(
_mm_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m20),
_mm_mul_ps(v.y, m21)),
_mm_add_ps(
_mm_mul_ps(v.z, m22), m23));
return r;
}
}

View file

@ -15,9 +15,12 @@ namespace math_backend::mat44::dispatch
gMat44.mul_pos3 = mat44_mul_pos3_impl;
gMat44.mul_vec3 = mat44_mul_vec3_impl;
gMat44.mul_float4 = mat44_mul_float4_impl;
gMat44.transform_plane = mat44_transform_plane_impl;
gMat44.scale = mat44_scale_impl;
gMat44.get_scale = mat44_get_scale_impl;
gMat44.normalize = mat44_normalize_impl;
gMat44.determinant = mat44_get_determinant;
gMat44.batch_mul_pos3 = mat44_batch_mul_pos3;
}
}

View file

@ -405,13 +405,6 @@ namespace
f32x8 w;
};
struct vec3_batch8
{
f32x8 x;
f32x8 y;
f32x8 z;
};
struct vec4_batch4
{
f32x4 x;
@ -420,47 +413,37 @@ namespace
f32x4 w;
};
struct vec3_batch4
inline vec4_batch8 load_vec3_batch8(const float* ptr, float w, bool fillW)
{
f32x4 x;
f32x4 y;
f32x4 z;
};
vec4_batch8 r;
inline vec3_batch8 load_vec3_batch8(const float* ptr)
{
// ptr points to x0
// layout: x y z x y z ...
r.x = _mm256_set_ps(ptr[0], ptr[3], ptr[6], ptr[9],
ptr[12], ptr[15], ptr[18], ptr[21]);
r.y = _mm256_set_ps(ptr[1], ptr[4], ptr[7], ptr[10],
ptr[13], ptr[16], ptr[19], ptr[22]);
r.z = _mm256_set_ps(ptr[2], ptr[5], ptr[8], ptr[11],
ptr[14], ptr[17], ptr[20], ptr[23]);
vec3_batch8 r;
r.x = _mm256_set_ps(
ptr[21], ptr[18], ptr[15], ptr[12],
ptr[9], ptr[6], ptr[3], ptr[0]);
r.y = _mm256_set_ps(
ptr[22], ptr[19], ptr[16], ptr[13],
ptr[10], ptr[7], ptr[4], ptr[1]);
r.z = _mm256_set_ps(
ptr[23], ptr[20], ptr[17], ptr[14],
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm256_set1_ps(w);
}
return r;
}
inline vec3_batch4 load_vec3_batch4(const float* ptr)
inline vec4_batch4 load_vec3_batch4(const float* ptr, float w, bool fillW)
{
vec3_batch4 r;
vec4_batch4 r;
r.x = _mm_set_ps(
ptr[9], ptr[6], ptr[3], ptr[0]);
r.x = _mm_set_ps(ptr[0], ptr[3], ptr[6], ptr[9]);
r.y = _mm_set_ps(ptr[1], ptr[4], ptr[7], ptr[10]);
r.z = _mm_set_ps(ptr[2], ptr[5], ptr[8], ptr[11]);
r.y = _mm_set_ps(
ptr[10], ptr[7], ptr[4], ptr[1]);
r.z = _mm_set_ps(
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm_set1_ps(w);
}
return r;
}
@ -478,7 +461,7 @@ namespace
}
// Store the result back to a float3 array with size of 8
inline void store_vec3_batch8(float* out, const vec3_batch8& v)
inline void store_vec3_batch8(float* out, const vec4_batch8& v)
{
alignas(32) float xs[8];
alignas(32) float ys[8];
@ -497,7 +480,7 @@ namespace
}
// Store the result back to a float3 array with size of 4
inline void store_vec3_batch4(float* out, const vec3_batch4& v)
inline void store_vec3_batch4(float* out, const vec4_batch4& v)
{
alignas(16) float xs[8];
alignas(16) float ys[8];
@ -515,7 +498,7 @@ namespace
}
}
inline f32x4 vec3_batch4_dot(const vec3_batch4& a, const vec3_batch4& b)
inline f32x4 vec3_batch4_dot(const vec4_batch4& a, const vec4_batch4& b)
{
f32x4 mulx = _mm_mul_ps(a.x, b.x);
f32x4 muly = _mm_mul_ps(a.y, b.y);
@ -524,7 +507,7 @@ namespace
return _mm_add_ps(_mm_add_ps(mulx, muly), mulz);
}
inline f32x8 vec3_batch8_dot(const vec3_batch8& a, const vec3_batch8& b)
inline f32x8 vec3_batch8_dot(const vec4_batch8& a, const vec4_batch8& b)
{
f32x8 mulx = _mm256_mul_ps(a.x, b.x);
f32x8 muly = _mm256_mul_ps(a.y, b.y);
@ -533,4 +516,110 @@ namespace
return _mm256_add_ps(_mm256_add_ps(mulx, muly), mulz);
}
// Batch 8 mul_Vec4.
inline vec4_batch8 m_mul_pos3_batch8(const f32x4x4& m, const vec4_batch8& v)
{
vec4_batch8 r;
__m256 m00 = _mm256_set1_ps(m.r0.m128_f32[0]);
__m256 m01 = _mm256_set1_ps(m.r0.m128_f32[1]);
__m256 m02 = _mm256_set1_ps(m.r0.m128_f32[2]);
__m256 m03 = _mm256_set1_ps(m.r0.m128_f32[3]);
__m256 m10 = _mm256_set1_ps(m.r1.m128_f32[0]);
__m256 m11 = _mm256_set1_ps(m.r1.m128_f32[1]);
__m256 m12 = _mm256_set1_ps(m.r1.m128_f32[2]);
__m256 m13 = _mm256_set1_ps(m.r1.m128_f32[3]);
__m256 m20 = _mm256_set1_ps(m.r2.m128_f32[0]);
__m256 m21 = _mm256_set1_ps(m.r2.m128_f32[1]);
__m256 m22 = _mm256_set1_ps(m.r2.m128_f32[2]);
__m256 m23 = _mm256_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m00),
_mm256_mul_ps(v.y, m01)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m10),
_mm256_mul_ps(v.y, m11)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm256_add_ps(
_mm256_add_ps(
_mm256_mul_ps(v.x, m20),
_mm256_mul_ps(v.y, m21)),
_mm256_add_ps(
_mm256_mul_ps(v.z, m22), m23));
return r;
}
// Batch 4 mul_Vec4.
inline vec4_batch4 m_mul_pos3_batch4(const f32x4x4& m, const vec4_batch4& v)
{
vec4_batch4 r;
f32x4 m00 = _mm_set1_ps(m.r0.m128_f32[0]);
f32x4 m01 = _mm_set1_ps(m.r0.m128_f32[1]);
f32x4 m02 = _mm_set1_ps(m.r0.m128_f32[2]);
f32x4 m03 = _mm_set1_ps(m.r0.m128_f32[3]);
f32x4 m10 = _mm_set1_ps(m.r1.m128_f32[0]);
f32x4 m11 = _mm_set1_ps(m.r1.m128_f32[1]);
f32x4 m12 = _mm_set1_ps(m.r1.m128_f32[2]);
f32x4 m13 = _mm_set1_ps(m.r1.m128_f32[3]);
f32x4 m20 = _mm_set1_ps(m.r2.m128_f32[0]);
f32x4 m21 = _mm_set1_ps(m.r2.m128_f32[1]);
f32x4 m22 = _mm_set1_ps(m.r2.m128_f32[2]);
f32x4 m23 = _mm_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m00),
_mm_mul_ps(v.y, m01)),
_mm_add_ps(
_mm_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m10),
_mm_mul_ps(v.y, m11)),
_mm_add_ps(
_mm_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m20),
_mm_mul_ps(v.y, m21)),
_mm_add_ps(
_mm_mul_ps(v.z, m22), m23));
return r;
}
}

View file

@ -15,9 +15,11 @@ namespace math_backend::mat44::dispatch
gMat44.mul_pos3 = mat44_mul_pos3_impl;
gMat44.mul_vec3 = mat44_mul_vec3_impl;
gMat44.mul_float4 = mat44_mul_float4_impl;
gMat44.transform_plane = mat44_transform_plane_impl;
gMat44.scale = mat44_scale_impl;
gMat44.get_scale = mat44_get_scale_impl;
gMat44.normalize = mat44_normalize_impl;
gMat44.determinant = mat44_get_determinant;
gMat44.batch_mul_pos3 = mat44_batch_mul_pos3;
}
}

View file

@ -410,4 +410,95 @@ namespace
return m_transpose(adj);
}
struct vec4_batch4
{
f32x4 x;
f32x4 y;
f32x4 z;
f32x4 w;
};
inline vec4_batch4 load_vec3_batch4(const float* ptr, float w, bool fillW)
{
vec4_batch4 r;
r.x = (f32x4){ ptr[9], ptr[6], ptr[3], ptr[0] };
r.y = (f32x4){ ptr[10], ptr[7], ptr[4], ptr[1] };
r.z = (f32x4){ ptr[11], ptr[8], ptr[5], ptr[2] };
if (fillW)
{
r.w = vdupq_n_f32(w);
}
return r;
}
inline void store_vec3_batch4(float* out, const vec4_batch4& v)
{
alignas(16) float xs[4];
alignas(16) float ys[4];
alignas(16) float zs[4];
vst1q_f32(xs, v.x);
vst1q_f32(ys, v.y);
vst1q_f32(zs, v.z);
for (int i = 0; i < 4; ++i)
{
out[i * 3 + 0] = xs[i];
out[i * 3 + 1] = ys[i];
out[i * 3 + 2] = zs[i];
}
}
inline vec4_batch4 m_mul_pos3_batch4(const f32x4x4& m, const vec4_batch4& v)
{
vec4_batch4 r;
float32x4_t m00 = vdupq_n_f32(m.r0.m128_f32[0]);
float32x4_t m01 = vdupq_n_f32(m.r0.m128_f32[1]);
float32x4_t m02 = vdupq_n_f32(m.r0.m128_f32[2]);
float32x4_t m03 = vdupq_n_f32(m.r0.m128_f32[3]);
float32x4_t m10 = vdupq_n_f32(m.r1.m128_f32[0]);
float32x4_t m11 = vdupq_n_f32(m.r1.m128_f32[1]);
float32x4_t m12 = vdupq_n_f32(m.r1.m128_f32[2]);
float32x4_t m13 = vdupq_n_f32(m.r1.m128_f32[3]);
float32x4_t m20 = vdupq_n_f32(m.r2.m128_f32[0]);
float32x4_t m21 = vdupq_n_f32(m.r2.m128_f32[1]);
float32x4_t m22 = vdupq_n_f32(m.r2.m128_f32[2]);
float32x4_t m23 = vdupq_n_f32(m.r2.m128_f32[3]);
// row0 dot
r.x = vaddq_f32(
vaddq_f32(
vmulq_f32(v.x, m00),
vmulq_f32(v.y, m01)),
vaddq_f32(
vmulq_f32(v.z, m02),
m03));
// row1 dot
r.y = vaddq_f32(
vaddq_f32(
vmulq_f32(v.x, m10),
vmulq_f32(v.y, m11)),
vaddq_f32(
vmulq_f32(v.z, m12),
m13));
// row2 dot
r.z = vaddq_f32(
vaddq_f32(
vmulq_f32(v.x, m20),
vmulq_f32(v.y, m21)),
vaddq_f32(
vmulq_f32(v.z, m22),
m23));
return r;
}
}

View file

@ -15,9 +15,11 @@ namespace math_backend::mat44::dispatch
gMat44.mul_pos3 = mat44_mul_pos3_impl;
gMat44.mul_vec3 = mat44_mul_vec3_impl;
gMat44.mul_float4 = mat44_mul_float4_impl;
gMat44.transform_plane = mat44_transform_plane_impl;
gMat44.scale = mat44_scale_impl;
gMat44.get_scale = mat44_get_scale_impl;
gMat44.normalize = mat44_normalize_impl;
gMat44.determinant = mat44_get_determinant;
gMat44.batch_mul_pos3 = mat44_batch_mul_pos3;
}
}

View file

@ -544,17 +544,10 @@ namespace
f32x4 w;
};
struct vec3_batch4
{
f32x4 x;
f32x4 y;
f32x4 z;
};
inline vec3_batch4 load_vec3_batch4(const float* ptr)
inline vec4_batch4 load_vec3_batch4(const float* ptr, float w, bool fillW)
{
vec3_batch4 r;
vec4_batch4 r;
r.x = _mm_set_ps(
ptr[9], ptr[6], ptr[3], ptr[0]);
@ -565,6 +558,11 @@ namespace
r.z = _mm_set_ps(
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm_set1_ps(w);
}
return r;
}
@ -575,7 +573,7 @@ namespace
}
// Store the result back to a float3 array with size of 4
inline void store_vec3_batch4(float* out, const vec3_batch4& v)
inline void store_vec3_batch4(float* out, const vec4_batch4& v)
{
alignas(16) float xs[8];
alignas(16) float ys[8];
@ -593,7 +591,7 @@ namespace
}
}
inline f32x4 vec3_batch4_dot(const vec3_batch4& a, const vec3_batch4& b)
inline f32x4 vec3_batch4_dot(const vec4_batch4& a, const vec4_batch4& b)
{
f32x4 mulx = _mm_mul_ps(a.x, b.x);
f32x4 muly = _mm_mul_ps(a.y, b.y);
@ -601,4 +599,57 @@ namespace
return _mm_add_ps(_mm_add_ps(mulx, muly), mulz);
}
// Batch 4 mul_Vec4.
inline vec4_batch4 m_mul_pos3_batch4(const f32x4x4& m, const vec4_batch4& v)
{
vec4_batch4 r;
f32x4 m00 = _mm_set1_ps(m.r0.m128_f32[0]);
f32x4 m01 = _mm_set1_ps(m.r0.m128_f32[1]);
f32x4 m02 = _mm_set1_ps(m.r0.m128_f32[2]);
f32x4 m03 = _mm_set1_ps(m.r0.m128_f32[3]);
f32x4 m10 = _mm_set1_ps(m.r1.m128_f32[0]);
f32x4 m11 = _mm_set1_ps(m.r1.m128_f32[1]);
f32x4 m12 = _mm_set1_ps(m.r1.m128_f32[2]);
f32x4 m13 = _mm_set1_ps(m.r1.m128_f32[3]);
f32x4 m20 = _mm_set1_ps(m.r2.m128_f32[0]);
f32x4 m21 = _mm_set1_ps(m.r2.m128_f32[1]);
f32x4 m22 = _mm_set1_ps(m.r2.m128_f32[2]);
f32x4 m23 = _mm_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m00),
_mm_mul_ps(v.y, m01)),
_mm_add_ps(
_mm_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m10),
_mm_mul_ps(v.y, m11)),
_mm_add_ps(
_mm_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m20),
_mm_mul_ps(v.y, m21)),
_mm_add_ps(
_mm_mul_ps(v.z, m22), m23));
return r;
}
}

View file

@ -15,9 +15,11 @@ namespace math_backend::mat44::dispatch
gMat44.mul_pos3 = mat44_mul_pos3_impl;
gMat44.mul_vec3 = mat44_mul_vec3_impl;
gMat44.mul_float4 = mat44_mul_float4_impl;
gMat44.transform_plane = mat44_transform_plane_impl;
gMat44.scale = mat44_scale_impl;
gMat44.get_scale = mat44_get_scale_impl;
gMat44.normalize = mat44_normalize_impl;
gMat44.determinant = mat44_get_determinant;
gMat44.batch_mul_pos3 = mat44_batch_mul_pos3;
}
}

View file

@ -514,17 +514,10 @@ namespace
f32x4 w;
};
struct vec3_batch4
{
f32x4 x;
f32x4 y;
f32x4 z;
};
inline vec3_batch4 load_vec3_batch4(const float* ptr)
inline vec4_batch4 load_vec3_batch4(const float* ptr, float w, bool fillW)
{
vec3_batch4 r;
vec4_batch4 r;
r.x = _mm_set_ps(
ptr[9], ptr[6], ptr[3], ptr[0]);
@ -535,6 +528,11 @@ namespace
r.z = _mm_set_ps(
ptr[11], ptr[8], ptr[5], ptr[2]);
if (fillW)
{
r.w = _mm_set1_ps(w);
}
return r;
}
@ -545,7 +543,7 @@ namespace
}
// Store the result back to a float3 array with size of 4
inline void store_vec3_batch4(float* out, const vec3_batch4& v)
inline void store_vec3_batch4(float* out, const vec4_batch4& v)
{
alignas(16) float xs[8];
alignas(16) float ys[8];
@ -563,7 +561,7 @@ namespace
}
}
inline f32x4 vec3_batch4_dot(const vec3_batch4& a, const vec3_batch4& b)
inline f32x4 vec3_batch4_dot(const vec4_batch4& a, const vec4_batch4& b)
{
f32x4 mulx = _mm_mul_ps(a.x, b.x);
f32x4 muly = _mm_mul_ps(a.y, b.y);
@ -571,4 +569,57 @@ namespace
return _mm_add_ps(_mm_add_ps(mulx, muly), mulz);
}
// Batch 4 mul_Vec4.
inline vec4_batch4 m_mul_pos3_batch4(const f32x4x4& m, const vec4_batch4& v)
{
vec4_batch4 r;
f32x4 m00 = _mm_set1_ps(m.r0.m128_f32[0]);
f32x4 m01 = _mm_set1_ps(m.r0.m128_f32[1]);
f32x4 m02 = _mm_set1_ps(m.r0.m128_f32[2]);
f32x4 m03 = _mm_set1_ps(m.r0.m128_f32[3]);
f32x4 m10 = _mm_set1_ps(m.r1.m128_f32[0]);
f32x4 m11 = _mm_set1_ps(m.r1.m128_f32[1]);
f32x4 m12 = _mm_set1_ps(m.r1.m128_f32[2]);
f32x4 m13 = _mm_set1_ps(m.r1.m128_f32[3]);
f32x4 m20 = _mm_set1_ps(m.r2.m128_f32[0]);
f32x4 m21 = _mm_set1_ps(m.r2.m128_f32[1]);
f32x4 m22 = _mm_set1_ps(m.r2.m128_f32[2]);
f32x4 m23 = _mm_set1_ps(m.r2.m128_f32[3]);
//
// row0 dot
//
r.x = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m00),
_mm_mul_ps(v.y, m01)),
_mm_add_ps(
_mm_mul_ps(v.z, m02), m03));
//
// row1 dot
//
r.y = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m10),
_mm_mul_ps(v.y, m11)),
_mm_add_ps(
_mm_mul_ps(v.z, m12), m13));
//
// row2 dot
//
r.z = _mm_add_ps(
_mm_add_ps(
_mm_mul_ps(v.x, m20),
_mm_mul_ps(v.y, m21)),
_mm_add_ps(
_mm_mul_ps(v.z, m22), m23));
return r;
}
}