1
0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-14 01:35:36 +00:00

Snippets from Guido to optimise Reduce

This commit is contained in:
Peter Boyle 2015-04-23 08:31:40 +01:00
parent 73c0db82d5
commit a9e574dd27

View File

@ -213,19 +213,44 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
friend inline ComplexF Reduce(const vComplexF & in) friend inline ComplexF Reduce(const vComplexF & in)
{ {
#ifdef SSE4 #ifdef SSE4
#error union {
__m128 v1; // SSE 4 x float vector
float f[4]; // scalar array of 4 floats
} u128;
u128.v1= _mm_add_ps(v, _mm_shuffle_ps(v, v, 0b01001110)); // FIXME Prefer to use _MM_SHUFFLE macros
return ComplexF(u128.f[0], u128.f[1]);
#endif #endif
#if defined (AVX1) || defined(AVX2) #ifdef AVX1
// FIXME this is inefficient and use //it would be better passing 2 arguments to saturate the vector lanes
__attribute__ ((aligned(32))) float c_[8]; union {
_mm256_store_ps(c_,in.v); __m256 v1;
return ComplexF(c_[0]+c_[2]+c_[4]+c_[6],c_[1]+c_[3]+c_[5]+c_[7]); float f[8];
} u256;
//SWAP lanes
__m256 t0 = _mm256_permute2f128_ps(in.v, in.v, 1);
__m256 t1 = _mm256_permute_ps(in.v , 0b11011000);//real (0,2,1,3)
__m256 t2 = _mm256_permute_ps(t0 , 0b10001101);//imag (1,3,0,2)
t0 = _mm256_blend_ps(t1, t2, 0b0101000001010000);// (0,0,1,1,0,0,1,1)
t1 = _mm256_hadd_ps(t0,t0);
u256.v1 = _mm256_hadd_ps(t1, t1);
return ComplexF(u256.f[0], u256.f[4]);
#endif
#ifdef AVX2
union {
__m256 v1;
float f[8];
} u256;
const __m256i mask= _mm256_set_epi32( 7, 5, 3, 1, 6, 4, 2, 0);
__m256 tmp1 = _mm256_permutevar8x32_ps(in.v, mask);
__m256 tmp2 = _mm256_hadd_ps(tmp1, tmp1);
u256.v1 = _mm256_hadd_ps(tmp2, tmp2);
return ComplexF(u256.f[0], u256.f[4]);
#endif #endif
#ifdef AVX512 #ifdef AVX512
return ComplexF(_mm512_mask_reduce_add_ps(0x5555, in.v),_mm512_mask_reduce_add_ps(0xAAAA, in.v)); return ComplexF(_mm512_mask_reduce_add_ps(0x5555, in.v),_mm512_mask_reduce_add_ps(0xAAAA, in.v));
#endif #endif
#ifdef QPX #ifdef QPX
#error
#endif #endif
} }