Skip to content

Commit 7ab1e73

Browse files
author
Raghuveer Devulapalli
committed
Add AVX-512 sort for float16 dtype
1 parent eda9e45 commit 7ab1e73

File tree

4 files changed

+284
-66
lines changed

4 files changed

+284
-66
lines changed

‎src/avx512-16bit-qsort.hpp‎

Lines changed: 221 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -29,10 +29,148 @@ static const uint16_t network[6][32]
2929
{16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
3030
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}};
3131

32+
struct float16 {
33+
uint16_t val;
34+
};
35+
36+
template <>
37+
struct zmm_vector<float16> {
38+
using type_t = uint16_t;
39+
using zmm_t = __m512i;
40+
using ymm_t = __m256i;
41+
using opmask_t = __mmask32;
42+
static const uint8_t numlanes = 32;
43+
44+
static zmm_t get_network(int index)
45+
{
46+
return _mm512_loadu_si512(&network[index - 1][0]);
47+
}
48+
static type_t type_max()
49+
{
50+
return X86_SIMD_SORT_INFINITYH;
51+
}
52+
static type_t type_min()
53+
{
54+
return X86_SIMD_SORT_NEGINFINITYH;
55+
}
56+
static zmm_t zmm_max()
57+
{
58+
return _mm512_set1_epi16(type_max());
59+
}
60+
static opmask_t knot_opmask(opmask_t x)
61+
{
62+
return _knot_mask32(x);
63+
}
64+
65+
static opmask_t ge(zmm_t x, zmm_t y)
66+
{
67+
zmm_t sign_x = _mm512_and_si512(x, _mm512_set1_epi16(0x8000));
68+
zmm_t sign_y = _mm512_and_si512(y, _mm512_set1_epi16(0x8000));
69+
zmm_t exp_x = _mm512_and_si512(x, _mm512_set1_epi16(0x7c00));
70+
zmm_t exp_y = _mm512_and_si512(y, _mm512_set1_epi16(0x7c00));
71+
zmm_t mant_x = _mm512_and_si512(x, _mm512_set1_epi16(0x3ff));
72+
zmm_t mant_y = _mm512_and_si512(y, _mm512_set1_epi16(0x3ff));
73+
74+
__mmask32 mask_ge = _mm512_cmp_epu16_mask(sign_x, sign_y, _MM_CMPINT_LT); // only greater than
75+
__mmask32 sign_eq = _mm512_cmpeq_epu16_mask(sign_x, sign_y);
76+
__mmask32 neg = _mm512_mask_cmpeq_epu16_mask(sign_eq, sign_x, _mm512_set1_epi16(0x8000)); // both numbers are -ve
77+
78+
// compare exponents only if signs are equal:
79+
mask_ge = mask_ge | _mm512_mask_cmp_epu16_mask(sign_eq, exp_x, exp_y, _MM_CMPINT_NLE);
80+
// get mask for elements for which both sign and exponents are equal:
81+
__mmask32 exp_eq = _mm512_mask_cmpeq_epu16_mask(sign_eq, exp_x, exp_y);
82+
83+
// compare mantissa for elements for which both sign and expponent are equal:
84+
mask_ge = mask_ge | _mm512_mask_cmp_epu16_mask(exp_eq, mant_x, mant_y, _MM_CMPINT_NLT);
85+
return _kxor_mask32(mask_ge, neg);
86+
}
87+
static zmm_t loadu(void const *mem)
88+
{
89+
return _mm512_loadu_si512(mem);
90+
}
91+
static zmm_t max(zmm_t x, zmm_t y)
92+
{
93+
return _mm512_mask_mov_epi16(y, ge(x, y), x);
94+
}
95+
static void mask_compressstoreu(void *mem, opmask_t mask, zmm_t x)
96+
{
97+
// AVX512_VBMI2
98+
return _mm512_mask_compressstoreu_epi16(mem, mask, x);
99+
}
100+
static zmm_t mask_loadu(zmm_t x, opmask_t mask, void const *mem)
101+
{
102+
// AVX512BW
103+
return _mm512_mask_loadu_epi16(x, mask, mem);
104+
}
105+
static zmm_t mask_mov(zmm_t x, opmask_t mask, zmm_t y)
106+
{
107+
return _mm512_mask_mov_epi16(x, mask, y);
108+
}
109+
static void mask_storeu(void *mem, opmask_t mask, zmm_t x)
110+
{
111+
return _mm512_mask_storeu_epi16(mem, mask, x);
112+
}
113+
static zmm_t min(zmm_t x, zmm_t y)
114+
{
115+
return _mm512_mask_mov_epi16(x, ge(x, y), y);
116+
}
117+
static zmm_t permutexvar(__m512i idx, zmm_t zmm)
118+
{
119+
return _mm512_permutexvar_epi16(idx, zmm);
120+
}
121+
// Apparently this is a terrible for perf, npy_half_to_float seems to work
122+
// better
123+
//static float uint16_to_float(uint16_t val)
124+
//{
125+
// // Ideally use _mm_loadu_si16, but its only gcc > 11.x
126+
// // TODO: use inline ASM? https://godbolt.org/z/aGYvh7fMM
127+
// __m128i xmm = _mm_maskz_loadu_epi16(0x01, &val);
128+
// __m128 xmm2 = _mm_cvtph_ps(xmm);
129+
// return _mm_cvtss_f32(xmm2);
130+
//}
131+
static type_t float_to_uint16(float val)
132+
{
133+
__m128 xmm = _mm_load_ss(&val);
134+
__m128i xmm2 = _mm_cvtps_ph(xmm, _MM_FROUND_NO_EXC);
135+
return _mm_extract_epi16(xmm2, 0);
136+
}
137+
static type_t reducemax(zmm_t v)
138+
{
139+
__m512 lo = _mm512_cvtph_ps(_mm512_extracti64x4_epi64(v, 0));
140+
__m512 hi = _mm512_cvtph_ps(_mm512_extracti64x4_epi64(v, 1));
141+
float lo_max = _mm512_reduce_max_ps(lo);
142+
float hi_max = _mm512_reduce_max_ps(hi);
143+
return float_to_uint16(std::max(lo_max, hi_max));
144+
}
145+
static type_t reducemin(zmm_t v)
146+
{
147+
__m512 lo = _mm512_cvtph_ps(_mm512_extracti64x4_epi64(v, 0));
148+
__m512 hi = _mm512_cvtph_ps(_mm512_extracti64x4_epi64(v, 1));
149+
float lo_max = _mm512_reduce_min_ps(lo);
150+
float hi_max = _mm512_reduce_min_ps(hi);
151+
return float_to_uint16(std::min(lo_max, hi_max));
152+
}
153+
static zmm_t set1(type_t v)
154+
{
155+
return _mm512_set1_epi16(v);
156+
}
157+
template <uint8_t mask>
158+
static zmm_t shuffle(zmm_t zmm)
159+
{
160+
zmm = _mm512_shufflehi_epi16(zmm, (_MM_PERM_ENUM)mask);
161+
return _mm512_shufflelo_epi16(zmm, (_MM_PERM_ENUM)mask);
162+
}
163+
static void storeu(void *mem, zmm_t x)
164+
{
165+
return _mm512_storeu_si512(mem, x);
166+
}
167+
};
168+
32169
template <>
33170
struct zmm_vector<int16_t> {
34171
using type_t = int16_t;
35172
using zmm_t = __m512i;
173+
using ymm_t = __m256i;
36174
using opmask_t = __mmask32;
37175
static const uint8_t numlanes = 32;
38176

@@ -130,6 +268,7 @@ template <>
130268
struct zmm_vector<uint16_t> {
131269
using type_t = uint16_t;
132270
using zmm_t = __m512i;
271+
using ymm_t = __m256i;
133272
using opmask_t = __mmask32;
134273
static const uint8_t numlanes = 32;
135274

@@ -227,7 +366,7 @@ struct zmm_vector<uint16_t> {
227366
* https://en.wikipedia.org/wiki/Bitonic_sorter#/media/File:BitonicSort.svg
228367
*/
229368
template <typename vtype, typename zmm_t = typename vtype::zmm_t>
230-
X86_SIMD_SORT_FORCEINLINE zmm_t sort_zmm_16bit(zmm_t zmm)
369+
X86_SIMD_SORT_FINLINE zmm_t sort_zmm_16bit(zmm_t zmm)
231370
{
232371
// Level 1
233372
zmm = cmp_merge<vtype>(
@@ -287,7 +426,7 @@ X86_SIMD_SORT_FORCEINLINE zmm_t sort_zmm_16bit(zmm_t zmm)
287426

288427
// Assumes zmm is bitonic and performs a recursive half cleaner
289428
template <typename vtype, typename zmm_t = typename vtype::zmm_t>
290-
X86_SIMD_SORT_FORCEINLINE zmm_t bitonic_merge_zmm_16bit(zmm_t zmm)
429+
X86_SIMD_SORT_FINLINE zmm_t bitonic_merge_zmm_16bit(zmm_t zmm)
291430
{
292431
// 1) half_cleaner[32]: compare 1-17, 2-18, 3-19 etc ..
293432
zmm = cmp_merge<vtype>(
@@ -313,8 +452,7 @@ X86_SIMD_SORT_FORCEINLINE zmm_t bitonic_merge_zmm_16bit(zmm_t zmm)
313452

314453
// Assumes zmm1 and zmm2 are sorted and performs a recursive half cleaner
315454
template <typename vtype, typename zmm_t = typename vtype::zmm_t>
316-
X86_SIMD_SORT_FORCEINLINE void bitonic_merge_two_zmm_16bit(zmm_t &zmm1,
317-
zmm_t &zmm2)
455+
X86_SIMD_SORT_FINLINE void bitonic_merge_two_zmm_16bit(zmm_t &zmm1, zmm_t &zmm2)
318456
{
319457
// 1) First step of a merging network: coex of zmm1 and zmm2 reversed
320458
zmm2 = vtype::permutexvar(vtype::get_network(4), zmm2);
@@ -328,7 +466,7 @@ X86_SIMD_SORT_FORCEINLINE void bitonic_merge_two_zmm_16bit(zmm_t &zmm1,
328466
// Assumes [zmm0, zmm1] and [zmm2, zmm3] are sorted and performs a recursive
329467
// half cleaner
330468
template <typename vtype, typename zmm_t = typename vtype::zmm_t>
331-
X86_SIMD_SORT_FORCEINLINE void bitonic_merge_four_zmm_16bit(zmm_t *zmm)
469+
X86_SIMD_SORT_FINLINE void bitonic_merge_four_zmm_16bit(zmm_t *zmm)
332470
{
333471
zmm_t zmm2r = vtype::permutexvar(vtype::get_network(4), zmm[2]);
334472
zmm_t zmm3r = vtype::permutexvar(vtype::get_network(4), zmm[3]);
@@ -349,7 +487,7 @@ X86_SIMD_SORT_FORCEINLINE void bitonic_merge_four_zmm_16bit(zmm_t *zmm)
349487
}
350488

351489
template <typename vtype, typename type_t>
352-
X86_SIMD_SORT_FORCEINLINE void sort_32_16bit(type_t *arr, int32_t N)
490+
X86_SIMD_SORT_FINLINE void sort_32_16bit(type_t *arr, int32_t N)
353491
{
354492
typename vtype::opmask_t load_mask = ((0x1ull << N) - 0x1ull) & 0xFFFFFFFF;
355493
typename vtype::zmm_t zmm
@@ -358,7 +496,7 @@ X86_SIMD_SORT_FORCEINLINE void sort_32_16bit(type_t *arr, int32_t N)
358496
}
359497

360498
template <typename vtype, typename type_t>
361-
X86_SIMD_SORT_FORCEINLINE void sort_64_16bit(type_t *arr, int32_t N)
499+
X86_SIMD_SORT_FINLINE void sort_64_16bit(type_t *arr, int32_t N)
362500
{
363501
if (N <= 32) {
364502
sort_32_16bit<vtype>(arr, N);
@@ -377,7 +515,7 @@ X86_SIMD_SORT_FORCEINLINE void sort_64_16bit(type_t *arr, int32_t N)
377515
}
378516

379517
template <typename vtype, typename type_t>
380-
X86_SIMD_SORT_FORCEINLINE void sort_128_16bit(type_t *arr, int32_t N)
518+
X86_SIMD_SORT_FINLINE void sort_128_16bit(type_t *arr, int32_t N)
381519
{
382520
if (N <= 64) {
383521
sort_64_16bit<vtype>(arr, N);
@@ -410,9 +548,9 @@ X86_SIMD_SORT_FORCEINLINE void sort_128_16bit(type_t *arr, int32_t N)
410548
}
411549

412550
template <typename vtype, typename type_t>
413-
X86_SIMD_SORT_FORCEINLINE type_t get_pivot_16bit(type_t *arr,
414-
const int64_t left,
415-
const int64_t right)
551+
X86_SIMD_SORT_FINLINE type_t get_pivot_16bit(type_t *arr,
552+
const int64_t left,
553+
const int64_t right)
416554
{
417555
// median of 32
418556
int64_t size = (right - left) / 32;
@@ -453,6 +591,38 @@ X86_SIMD_SORT_FORCEINLINE type_t get_pivot_16bit(type_t *arr,
453591
return ((type_t *)&sort)[16];
454592
}
455593

594+
template <>
595+
bool comparison_func<zmm_vector<float16>>(const uint16_t &a, const uint16_t &b)
596+
{
597+
uint16_t signa = a & 0x8000, signb = b & 0x8000;
598+
uint16_t expa = a & 0x7c00, expb = b & 0x7c00;
599+
uint16_t manta = a & 0x3ff, mantb = b & 0x3ff;
600+
if (signa != signb) {
601+
// opposite signs
602+
return a > b;
603+
}
604+
else if (signa > 0) {
605+
// both -ve
606+
if (expa != expb) {
607+
return expa > expb;
608+
}
609+
else {
610+
return manta > mantb;
611+
}
612+
}
613+
else {
614+
// both +ve
615+
if (expa != expb) {
616+
return expa < expb;
617+
}
618+
else {
619+
return manta < mantb;
620+
}
621+
}
622+
623+
//return npy_half_to_float(a) < npy_half_to_float(b);
624+
}
625+
456626
template <typename vtype, typename type_t>
457627
static void
458628
qsort_16bit_(type_t *arr, int64_t left, int64_t right, int64_t max_iters)
@@ -461,7 +631,7 @@ qsort_16bit_(type_t *arr, int64_t left, int64_t right, int64_t max_iters)
461631
* Resort to std::sort if quicksort isnt making any progress
462632
*/
463633
if (max_iters <= 0) {
464-
std::sort(arr + left, arr + right + 1);
634+
std::sort(arr + left, arr + right + 1, comparison_func<vtype>);
465635
return;
466636
}
467637
/*
@@ -483,12 +653,39 @@ qsort_16bit_(type_t *arr, int64_t left, int64_t right, int64_t max_iters)
483653
qsort_16bit_<vtype>(arr, pivot_index, right, max_iters - 1);
484654
}
485655

656+
X86_SIMD_SORT_FINLINE int64_t replace_nan_with_inf(uint16_t *arr, int64_t arrsize)
657+
{
658+
int64_t nan_count = 0;
659+
__mmask16 loadmask = 0xFFFF;
660+
while (arrsize > 0) {
661+
if (arrsize < 16) { loadmask = (0x0001 << arrsize) - 0x0001; }
662+
__m256i in_zmm = _mm256_maskz_loadu_epi16(loadmask, arr);
663+
__m512 in_zmm_asfloat = _mm512_cvtph_ps(in_zmm);
664+
__mmask16 nanmask = _mm512_cmp_ps_mask(
665+
in_zmm_asfloat, in_zmm_asfloat, _CMP_NEQ_UQ);
666+
nan_count += _mm_popcnt_u32((int32_t)nanmask);
667+
_mm256_mask_storeu_epi16(arr, nanmask, YMM_MAX_HALF);
668+
arr += 16;
669+
arrsize -= 16;
670+
}
671+
return nan_count;
672+
}
673+
674+
X86_SIMD_SORT_FINLINE void
675+
replace_inf_with_nan(uint16_t *arr, int64_t arrsize, int64_t nan_count)
676+
{
677+
for (int64_t ii = arrsize - 1; nan_count > 0; --ii) {
678+
arr[ii] = 0xFFFF;
679+
nan_count -= 1;
680+
}
681+
}
682+
486683
template <>
487684
void avx512_qsort(int16_t *arr, int64_t arrsize)
488685
{
489686
if (arrsize > 1) {
490687
qsort_16bit_<zmm_vector<int16_t>, int16_t>(
491-
arr, 0, arrsize - 1, 2 * (63 - __builtin_clzll(arrsize)));
688+
arr, 0, arrsize - 1, 2 * (int64_t)log2(arrsize));
492689
}
493690
}
494691

@@ -497,7 +694,17 @@ void avx512_qsort(uint16_t *arr, int64_t arrsize)
497694
{
498695
if (arrsize > 1) {
499696
qsort_16bit_<zmm_vector<uint16_t>, uint16_t>(
500-
arr, 0, arrsize - 1, 2 * (63 - __builtin_clzll(arrsize)));
697+
arr, 0, arrsize - 1, 2 * (int64_t)log2(arrsize));
698+
}
699+
}
700+
701+
void avx512_qsort_fp16(uint16_t *arr, int64_t arrsize)
702+
{
703+
if (arrsize > 1) {
704+
int64_t nan_count = replace_nan_with_inf(arr, arrsize);
705+
qsort_16bit_<zmm_vector<float16>, uint16_t>(
706+
arr, 0, arrsize - 1, 2 * (int64_t)log2(arrsize));
707+
replace_inf_with_nan(arr, arrsize, nan_count);
501708
}
502709
}
503710
#endif // AVX512_QSORT_16BIT

0 commit comments

Comments
 (0)