Skip to content

Commit

Permalink
kram - simd - fix up long, use neon vec2 ops
Browse files Browse the repository at this point in the history
  • Loading branch information
@alecazam
alecazam committed Sep 30, 2024
1 parent 3ae9626 commit 3b9ee68
Show file tree
Hide file tree
Showing 7 changed files with 320 additions and 195 deletions.
121 changes: 85 additions & 36 deletions libkram/vectormath/double234.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,9 @@ typedef struct { double4s columns[3]; } double3x4s;
typedef struct { double4s columns[4]; } double4x4s;

// glue to Accelerate
#if SIMD_RENAME_TO_SIMD_NAMESPACE
#if SIMD_ACCELERATE_MATH_NAMES
macroVector8TypesStorageRenames(double, simd_double)
#endif
#endif // SIMD_ACCELERATE_MATH_NAMES

#ifdef __cplusplus
}
Expand Down Expand Up @@ -76,15 +76,13 @@ SIMD_CALL double4 double4m(double3 v, double w = 1.0) {
SIMD_CALL double reduce_min(double2 x) {
return vminvq_f64(x);
}

SIMD_CALL double reduce_min(double4 x) {
return fmin(reduce_min(x.lo),reduce_min(x.hi));
}

SIMD_CALL double reduce_max(double2 x) {
return vmaxvq_f64(x);
}

SIMD_CALL double reduce_max(double4 x) {
return fmax(reduce_max(x.lo),reduce_max(x.hi));
}
Expand All @@ -102,7 +100,6 @@ SIMD_CALL double2 max(double2 x, double2 y) {
// precise returns x on Nan
return vmaxnmq_f64(x, y);
}

SIMD_CALL double4 max(double4 x, double4 y) {
// precise returns x on Nan
return double4m(max(x.lo,y.lo), max(x.hi,y.hi));
Expand All @@ -125,28 +122,59 @@ SIMD_CALL double4 sqrt(double4 x) {
return double4m(sqrt(x.lo), sqrt(x.hi));
}

// use sse2neon to port this for now
SIMD_CALL double2 reduce_addv(double2 x) {
// 4:1 reduction
x = vpaddq_f64(x, x);
return x.x;
}
SIMD_CALL double4 reduce_addv(double4 x) {
// 4:1 reduction
x.lo = _mm_hadd_pd(x.lo, x.lo);
x.hi = _mm_hadd_pd(x.hi, x.hi);
x.lo = _mm_hadd_pd(x.lo, x.hi);
x.lo = vpaddq_f64(x.lo, x.lo);
x.hi = vpaddq_f64(x.hi, x.hi);
x.lo = vpaddq_f64(x.lo, x.hi);
return x.lo.x; // repeat x to all values
}

SIMD_CALL double reduce_add(double2 x) {
return reduce_addv(x).x;
}
SIMD_CALL double reduce_add(double4 x) {
return reduce_addv(x).x;
}

SIMD_CALL double2 round(double2 vv) {
// round to nearest | exc
return vrndnq_f64(vv); // _MM_FROUND_NO_EXC
}
SIMD_CALL double4 round(double4 vv) {
// round to nearest | exc
return double4m(round(vv.lo),round(vv.hi));
}

SIMD_CALL double2 ceil(double2 vv) {
return vrndpq_f64(vv);
}
SIMD_CALL double4 ceil(double4 vv) {
return double4m(ceil(vv.lo),ceil(vv.hi));
}

SIMD_CALL double2 floor(double2 vv) {
return vrndmq_f64(vv);
}
SIMD_CALL double4 floor(double4 vv) {
return double4m(floor(vv.lo),floor(vv.hi));
}

#endif // SIMD_NEON

//----------------------

#if SIMD_SSE

// x64 doesn't seem to have a simd op for min/max reduce
SIMD_CALL double reduce_min(double4 x) {
return fmin(fmin(x.x,x.y), fmin(x.z,x.w));
}

SIMD_CALL double reduce_max(double4 x) {
return fmax(fmax(x.x,x.y), fmax(x.z,x.w));
}
Expand All @@ -163,7 +191,6 @@ SIMD_CALL double4 min(double4 x, double4 y) {
// precise returns x on Nan
return bitselect_forminmax(_mm_min_pd(x, y), x, y != y);
}

SIMD_CALL double4 max(double4 x, double4 y) {
// precise returns x on Nan
return bitselect_forminmax(_mm_max_pd(x, y), x, y != y);
Expand All @@ -178,6 +205,9 @@ SIMD_CALL double4 muladd(double4 x, double4 y, double4 t) {
return x * y + t;
#endif
}
SIMD_CALL double2 muladd(double2 x, double2 y, double2 t) {
return vec4to2(muladd(vec2to4(x), vec2to4(y), vec2to4(t)));
}

SIMD_CALL double2 sqrt(double2 x) {
return _mm_sqrt_pd(x);
Expand All @@ -190,7 +220,6 @@ SIMD_CALL double2 reduce_addv(double2 x) {
x = _mm_hadd_pd(x.lo, x.lo);
return x.x;
}

SIMD_CALL double4 reduce_addv(double4 x) {
// 4:1 reduction
x.lo = _mm_hadd_pd(x.lo, x.lo); // TODO: fix
Expand All @@ -199,24 +228,42 @@ SIMD_CALL double4 reduce_addv(double4 x) {
return x.lo.x; // repeat x to all values
}

SIMD_CALL double reduce_add(double2 x) {
return reduce_addv(x).x;
}
SIMD_CALL double reduce_add(double4 x) {
return reduce_addv(x).x;
}

#endif // SIMD_LONG && SIMD_SSE

// SSE4.1
// single ops in AVX/2

SIMD_CALL double4 round(double4 vv) {
return double4m(_mm_round_pd(vv.lo, 0x8),_mm_round_pd(vv.hi, 0x8)); // round to nearest | exc
// round to nearest | exc
return double4m(_mm_round_pd(vv.lo, _MM_FROUND_NO_EXC),
_mm_round_pd(vv.hi, _MM_FROUND_NO_EXC));
}
SIMD_CALL double2 round(double2 x) {
return vec4to2(round(vec2to4(x)));
}

SIMD_CALL double4 ceil(double4 vv) {
return double4m(_mm_ceil_pd(vv.lo),_mm_ceil_pd(vv.hi));
}
SIMD_CALL double2 ceil(double2 x) {
return vec4to2(ceil(vec2to4(x)));
}

SIMD_CALL double4 floor(double4 vv) {
return double4m(_mm_floor_pd(vv.lo),_mm_floor_pd(vv.hi));
}
SIMD_CALL double2 floor(double2 x) {
return vec4to2(floor(vec2to4(x)));
}


#endif // SIMD_SSE


// end of implementation
//-----------------------------------
Expand Down Expand Up @@ -249,10 +296,10 @@ SIMD_CALL double4 select(double4 x, double4 y, long4 mask) {

// zeroext - internal helper
SIMD_CALL double4 zeroext(double2 x) {
return (double4){x.x,x.y,0,0};
double4 v = 0; v.xy = x; return v;
}
SIMD_CALL double4 zeroext(double3 x) {
return (double4){x.x,x.y,x.z,0};
double4 v = 0; v.xyz = x; return v;
}

// min
Expand Down Expand Up @@ -307,9 +354,9 @@ SIMD_CALL double3 recip(double3 x) {
}

// reduce_add
SIMD_CALL double reduce_add(double2 x) {
return reduce_add(zeroext(x));
}
//SIMD_CALL double reduce_add(double2 x) {
// return reduce_add(zeroext(x));
//}
SIMD_CALL double reduce_add(double3 x) {
return reduce_add(zeroext(x));
}
Expand All @@ -332,16 +379,14 @@ SIMD_CALL double reduce_max(double3 x) {
return reduce_max(vec3to4(x));
}


// round (to nearest)
SIMD_CALL double2 round(double2 x) { return vec4to2(round(vec2to4(x))); }
SIMD_CALL double3 round(double3 x) { return vec4to3(round(vec3to4(x))); }

// ceil
SIMD_CALL double2 ceil(double2 x) { return vec4to2(ceil(vec2to4(x))); }
SIMD_CALL double3 ceil(double3 x) { return vec4to3(ceil(vec3to4(x))); }

// floor
SIMD_CALL double2 floor(double2 x) { return vec4to2(floor(vec2to4(x))); }
SIMD_CALL double3 floor(double3 x) { return vec4to3(floor(vec3to4(x))); }

// clamp
Expand All @@ -356,14 +401,18 @@ SIMD_CALL double4 clamp(double4 x, double4 minv, double4 maxv) {
return min(maxv, max(minv, x));
}

double2 saturate(double2 x);
double3 saturate(double3 x);
double4 saturate(double4 x);
// saturate
SIMD_CALL double2 saturate(double2 x) {
return clamp(x, 0, (double2)1);
}
SIMD_CALL double3 saturate(double3 x) {
return clamp(x, 0, (double3)1);
}
SIMD_CALL double4 saturate(double4 x) {
return clamp(x, 0, (double4)1);
}

// muladd - arm has double2 op
//SIMD_CALL double2 muladd(double2 x, double2 y, double2 t) {
// return vec4to2(muladd(vec2to4(x), vec2to4(y), vec2to4(t)));
//}
// muladd
SIMD_CALL double3 muladd(double3 x, double3 y, double3 t) {
return vec4to3(muladd(vec3to4(x), vec3to4(y), vec3to4(t)));
}
Expand Down Expand Up @@ -440,13 +489,13 @@ SIMD_CALL double3 normalize(double3 x) {

// abs
SIMD_CALL double2 abs(double2 x) {
return bitselect(0.0, x, 0x7fffffff); // TODO: fp64
return bitselect(0.0, x, 0x7fffffffffffffff);
}
SIMD_CALL double3 abs(double3 x) {
return bitselect(0.0, x, 0x7fffffff);
return bitselect(0.0, x, 0x7fffffffffffffff);
}
SIMD_CALL double4 abs(double4 x) {
return bitselect(0.0, x, 0x7fffffff);
return bitselect(0.0, x, 0x7fffffffffffffff);
}

// cross
Expand Down Expand Up @@ -506,13 +555,13 @@ SIMD_CALL double4 smoothstep(double4 edge0, double4 edge1, double4 x) {

// fract
SIMD_CALL double2 fract(double2 x) {
return min(x - floor(x), 0x1.fffffep-1f); // TODO: fp64
return min(x - floor(x), 0x1.fffffffffffffp-1);
}
SIMD_CALL double3 fract(double3 x) {
return min(x - floor(x), 0x1.fffffep-1f);
return min(x - floor(x), 0x1.fffffffffffffp-1);
}
SIMD_CALL double4 fract(double4 x) {
return min(x - floor(x), 0x1.fffffep-1f);
return min(x - floor(x), 0x1.fffffffffffffp-1);
}

//-------------------
Expand Down
Loading

0 comments on commit 3b9ee68

Please sign in to comment.