mul(avx_simd_complex_float<T> lhs, avx_simd_complex_float<T> rhs) {
//lhs = [x1.real, x1.img, x2.real, x2.img, ...]
//rhs = [y1.real, y1.img, y2.real, y2.img, ...]
//ymm1 = [y1.real, y1.real, y2.real, y2.real, ...]
__m256 ymm1 = _mm256_moveldup_ps(rhs.value);
//ymm2 = [x1.img, x1.real, x2.img, x2.real]
__m256 ymm2 = _mm256_permute_ps(lhs.value, 0b10110001);
//ymm3 = [y1.imag, y1.imag, y2.imag, y2.imag]
__m256 ymm3 = _mm256_movehdup_ps(rhs.value);
//ymm4 = ymm2 * ymm3
__m256 ymm4 = _mm256_mul_ps(ymm2, ymm3);
//result = [(lhs * ymm1) -+ ymm4];
#ifdef __FMA__
return _mm256_fmaddsub_ps(lhs.value, ymm1, ymm4);
#elif defined(__FMA4__)
return _mm256_maddsub_ps(lhs.value, ymm1, ymm4);
#else
__m256 tmp = _mm256_mul_ps(lhs.value, ymm1);
return _mm256_addsub_ps(tmp, ymm4);
#endif
}
/****************************************************************
* This technique for efficient SIMD complex-complex multiplication was found at
* https://software.intel.com/file/1000
*****************************************************************/
inline __m256 avx_multiply_float_complex_(const __m256& vecA, const __m256& vecB) {
__m256 vec1 = _mm256_moveldup_ps(vecB);
__m256 vec2 = _mm256_movehdup_ps(vecB);
vec1 = _mm256_mul_ps(vecA,vec1);
vec2 = _mm256_mul_ps(vecA,vec2);
vec2 = _mm256_permute_ps(vec2,0xB1);
return _mm256_addsub_ps(vec1,vec2);
}
// =============================================================
// ====================== RGBX2BGRX_32F ========================
// =============================================================
void _rgbx2bgrx_32f(const float* _src, float* _dest, unsigned int _width,
unsigned int _pitchs, unsigned int _pitchd,
unsigned int _start, unsigned int _stop) {
#ifdef USE_SSE
const unsigned int widthz = (_pitchs/8);
// Get start positions for buffers
const float* tsrc;
float* tdest;
for( unsigned int y=_start; y<=_stop; ++y ) {
tsrc = _src+(y*_pitchs);
tdest = _dest+(y*_pitchd);
for( unsigned int x=0; x<widthz; ++x ) {
#ifdef USE_AVX1
const __m256 v0 = _mm256_load_ps(tsrc);
tsrc+=8;
__m256 r0 = _mm256_permute_ps(v0,0xc6);
_mm256_store_ps(tdest, r0 );
tdest+=8;
#else // NOT TESTED
const __m128 v0 = _mm_load_ps(tsrc);
tsrc+=4;
const __m128 v1 = _mm_load_ps(tsrc);
tsrc+=4;
//__m128 r0 = _mm_shuffle_ps(v0,0xc6);
//__m128 r1 = _mm_shuffle_ps(v1,0xc6);
//_mm_store_ps(tdest, r0 ); tdest+=4;
//_mm_store_ps(tdest, r1 ); tdest+=4;
#endif
}
}
#else
const float* tsrc;
float* tdest;
for( unsigned int y=_start; y<=_stop; ++y ) {
tsrc = _src+(y*_pitchs);
tdest = _dest+(y*_pitchd);
for( unsigned int x=0; x<_width; ++x ) {
float t = tsrc[4*x];
tdest[4*x] = tsrc[4*x+2];
tdest[4*x+2] = t;
}
}
#endif
}
bool PaTriList2(PA_STATE &pa, UINT slot, simdvector tri[3])
{
simdvector &a = PaGetSimdVector(pa, 0, slot);
simdvector &b = PaGetSimdVector(pa, 1, slot);
simdvector &c = PaGetSimdVector(pa, 2, slot);
simdscalar s;
for (int i = 0; i < 4; ++i)
{
simdvector &v0 = tri[0];
v0[i] = _simd_blend_ps(a[i], b[i], 0x92);
v0[i] = _simd_blend_ps(v0[i], c[i], 0x24);
v0[i] = _mm256_permute_ps(v0[i], 0x6C);
s = _mm256_permute2f128_ps(v0[i], v0[i], 0x21);
v0[i] = _simd_blend_ps(v0[i], s, 0x44);
simdvector &v1 = tri[1];
v1[i] = _simd_blend_ps(a[i], b[i], 0x24);
v1[i] = _simd_blend_ps(v1[i], c[i], 0x49);
v1[i] = _mm256_permute_ps(v1[i], 0xB1);
s = _mm256_permute2f128_ps(v1[i], v1[i], 0x21);
v1[i] = _simd_blend_ps(v1[i], s, 0x66);
simdvector &v2 = tri[2];
v2[i] = _simd_blend_ps(a[i], b[i], 0x49);
v2[i] = _simd_blend_ps(v2[i], c[i], 0x92);
v2[i] = _mm256_permute_ps(v2[i], 0xC6);
s = _mm256_permute2f128_ps(v2[i], v2[i], 0x21);
v2[i] = _simd_blend_ps(v2[i], s, 0x22);
}
SetNextPaState(pa, PaTriList0, PaTriListSingle0);
pa.reset = true;
pa.numPrimsComplete += KNOB_VS_SIMD_WIDTH;
return true;
}
div(avx_simd_complex_float<T> lhs, avx_simd_complex_float<T> rhs) {
//lhs = [x1.real, x1.img, x2.real, x2.img ...]
//rhs = [y1.real, y1.img, y2.real, y2.img ...]
//ymm0 = [y1.real, y1.real, y2.real, y2.real, ...]
__m256 ymm0 = _mm256_moveldup_ps(rhs.value);
//ymm1 = [y1.imag, y1.imag, y2.imag, y2.imag]
__m256 ymm1 = _mm256_movehdup_ps(rhs.value);
//ymm2 = [x1.img, x1.real, x2.img, x2.real]
__m256 ymm2 = _mm256_permute_ps(lhs.value, 0b10110001);
//ymm4 = [x.img * y.img, x.real * y.img]
__m256 ymm4 = _mm256_mul_ps(ymm2, ymm1);
//ymm5 = subadd((lhs * ymm0), ymm4)
#ifdef __FMA__
__m256 ymm5 = _mm256_fmsubadd_ps(lhs.value, ymm0, ymm4);
#else
__m256 t1 = _mm256_mul_ps(lhs.value, ymm0);
__m256 t2 = _mm256_sub_ps(_mm256_set1_ps(0.0), ymm4);
__m256 ymm5 = _mm256_addsub_ps(t1, t2);
#endif
//ymm3 = [y.imag^2, y.imag^2]
__m256 ymm3 = _mm256_mul_ps(ymm1, ymm1);
//ymm0 = (ymm0 * ymm0 + ymm3)
#ifdef __FMA__
ymm0 = _mm256_fmadd_ps(ymm0, ymm0, ymm3);
#else
__m256 t3 = _mm256_mul_ps(ymm0, ymm0);
ymm0 = _mm256_add_ps(t3, ymm3);
#endif
//result = ymm5 / ymm0
return _mm256_div_ps(ymm5, ymm0);
}
void
test8bit (void)
{
i1 = _mm_cmpistrm (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistri (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistra (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistrc (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistro (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistrs (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
k1 = _mm_cmpistrz (i2, i3, k4); /* { dg-error "the third argument must be an 8-bit immediate" } */
i1 = _mm_cmpestrm (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestri (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestra (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestrc (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestro (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestrs (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
k1 = _mm_cmpestrz (i2, k2, i3, k3, k4); /* { dg-error "the fifth argument must be an 8-bit immediate" } */
b1 = _mm256_blend_ps (b2, b3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
k1 = _cvtss_sh (f1, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm256_cvtps_ph (b2, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
b1 = _mm256_dp_ps (b2, b3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
e1 = _mm256_permute2f128_pd (e2, e3, k4);/* { dg-error "the last argument must be an 8-bit immediate" } */
b1 = _mm256_permute2f128_ps (b2, b3, k4);/* { dg-error "the last argument must be an 8-bit immediate" } */
l1 = _mm256_permute2f128_si256 (l2, l3, k4);/* { dg-error "the last argument must be an 8-bit immediate" } */
b1 = _mm256_permute_ps (b2, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_aeskeygenassist_si128 (i2, k4);/* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_blend_epi16 (i2, i3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_clmulepi64_si128 (i2, i3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_cvtps_ph (a1, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
d1 = _mm_dp_pd (d2, d3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
a1 = _mm_dp_ps (a2, a3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
a1 = _mm_insert_ps (a2, a3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_mpsadbw_epu8 (i2, i3, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
a1 = _mm_permute_ps (a2, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_slli_si128 (i2, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
i1 = _mm_srli_si128 (i2, k4); /* { dg-error "the last argument must be an 8-bit immediate" } */
}
INLINE avxb shuffle(const avxb& a) {
return _mm256_permute_ps(a, _MM_SHUFFLE(i3, i2, i1, i0));
}
template<index_t index_0, index_t index_1, index_t index_2, index_t index_3> INLINE const avxi shuffle( const avxi& a ) {
return _mm256_castps_si256(_mm256_permute_ps(_mm256_castsi256_ps(a), _MM_SHUFFLE(index_3, index_2, index_1, index_0)));
}
M_ALWAYS_INLINE
static void bar(float (& input)[8])
{
/*
static constexpr uint_fast8_t idx[][2] = {
{0, 1}, {3, 2}, {4, 5}, {7, 6}, // (1)
{0, 2}, {1, 3}, {6, 4}, {7, 5}, // (2)
{0, 1}, {2, 3}, {5, 4}, {7, 6}, // (3)
{0, 4}, {1, 5}, {2, 6}, {3, 7}, // (4)
{0, 2}, {1, 3}, {4, 6}, {5, 7}, // (5)
{0, 1}, {2, 3}, {4, 5}, {6, 7} // (6)
};
*/
// Индекса трябва да представим в по удобен вид за
// AVX инструкциите. Няма смисъл от цикъл и после развиване
// защото (4)-тия случай е специален... По добре на ръка.
static constexpr int blend_mask_1 =0b10011001;
static constexpr int blend_mask_2=0b11000011;
static constexpr int blend_mask_3 =0b10100101;
static constexpr int blend_mask_4 =0b00001111;
static constexpr int blend_mask_5=0b00110011;
static constexpr int blend_mask_6=0b01010101;
// Отговаря на (1), (3) и (6)
static constexpr int permute_mask_1=0b10110001;
// Отговаря на (2) и (5)
static constexpr int permute_mask_2=0b01001110;
__m256 result= _mm256_load_ps(input);
// (1)
__m256 mapped=_mm256_permute_ps(result,permute_mask_1);
__m256 min=_mm256_min_ps(result,mapped);
__m256 max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_1);
// (2)
mapped=_mm256_permute_ps(result,permute_mask_2);
min=_mm256_min_ps(result,mapped);
max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_2);
// (3)
mapped=_mm256_permute_ps(result,permute_mask_1);
min=_mm256_min_ps(result,mapped);
max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_3);
// (4) Специалния случай тук трябва да пермутираме
// между двете половини на YMM регистъра.
mapped=_mm256_permute2f128_ps(result,result,1);
min=_mm256_min_ps(result,mapped);
max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_4);
// (5)
mapped=_mm256_permute_ps(result,permute_mask_2);
min=_mm256_min_ps(result,mapped);
max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_5);
// (6)
mapped=_mm256_permute_ps(result,permute_mask_1);
min=_mm256_min_ps(result,mapped);
max=_mm256_max_ps(result,mapped);
result=_mm256_blend_ps(max,min,blend_mask_6);
/**/
_mm256_store_ps(input,result);
}
}
nl = (1 << log);
*newLength = nl;
float *ret = mallocf(nl + additionalLength);
memcpy(ret, ptr, length * sizeof(float));
memsetf(ret + length, 0.f, nl - length);
return ret;
}
float *rmemcpyf(float *__restrict dest,
const float *__restrict src, size_t length) {
#ifdef __AVX__
for (int i = 0; i < (int)length - 7; i += 8) {
__m256 vec = _mm256_loadu_ps(src + i);
vec = _mm256_permute2f128_ps(vec, vec, 1);
vec = _mm256_permute_ps(vec, 0x1B);
_mm256_storeu_ps(dest + length - i - 8, vec);
}
for (size_t i = (length & ~0x7); i < length; i++) {
dest[length - i - 1] = src[i];
}
#elif defined(__ARM_NEON__)
for (int i = 0; i < (int)length - 3; i += 4) {
float32x4_t vec = vld1q_f32(src + i);
vec = vrev64q_f32(vec);
vec = vcombine_f32(vget_high_f32(vec), vget_low_f32(vec));
vst1q_f32(dest + length - i - 4, vec);
}
for (size_t i = (length & ~0x3); i < length; i++) {
// PRE: all vectors aligned,
// imag_c = [i1,i1,...,i4,i4]
// vec = [v1r,v1i,...,v4r,v4i]
// component-wise multiplication
// POST: returns [-i1*v1i,i1*v1r,...,-i4*v4i,i4*v4r]
inline __m256 avx_multiply_float_imag_(const __m256& imag_c, const __m256& vec) {
static const __m256 zero = _mm256_setzero_ps();
__m256 vec1 = _mm256_mul_ps(imag_c,vec);
vec1 = _mm256_permute_ps(vec1,0xB1);
return _mm256_addsub_ps(zero,vec1);
}
inline __m256 _mm256_broadcast_3_ss(__m256 a) {
__m256 b = _mm256_permute_ps(a, _MM_SHUFFLE(3, 3, 3, 3));
return _mm256_blend_ps(b, _mm256_permute2f128_ps(b, b, 1), 0xF0);
}
inline __m256 _mm256_broadcast_lo_ss(__m256 a) {
__m256 b = _mm256_permute_ps(a, _MM_SHUFFLE(0, 0, 0, 0)); \
return _mm256_blend_ps(b, _mm256_permute2f128_ps(b, b, 1), 0xF0); \
}
__m256 test_mm256_permute_ps(__m256 a) {
// Check if the mask is correct
// CHECK: shufflevector{{.*}}<i32 3, i32 2, i32 1, i32 0, i32 7, i32 6, i32 5, i32 4>
return _mm256_permute_ps(a, 0x1b);
}
