Debian Codegraph @

src/f16-f32acc-rsum/f16c.c.in

raw
1 // Copyright 2023 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 $assert BATCH_TILE % 8 == 0
7 $assert BATCH_TILE >= 8
8 $SIMD_TILE = BATCH_TILE // 8
9 $assert ACCUMULATORS <= SIMD_TILE
10 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
11 #include <assert.h>
12
13 #include <immintrin.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/reduce.h"
18 #include "xnnpack/unaligned.h"
19
20
21 $ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
22 void xnn_f16_f32acc_rsum_ukernel__f16c_u${BATCH_TILE}${ACC_SUFFIX}(
23 size_t batch,
24 const xnn_float16* input,
25 float* output,
26 const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
29
30 assert(batch != 0);
31 assert(batch % sizeof(uint16_t) == 0);
32 assert(input != NULL);
33 assert(output != NULL);
34
35 const uint16_t* i = (const uint16_t*) input;
36 $for A in range(ACCUMULATORS):
37 __m256 vacc${A} = _mm256_setzero_ps();
38 $if BATCH_TILE > 8:
39 for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
40 const __m256 vt0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
41 $for N in range(1, SIMD_TILE):
42 const __m256 vt${N} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
43 i += ${BATCH_TILE};
44
45 $for N in range(SIMD_TILE):
46 vacc${N % ACCUMULATORS} = _mm256_add_ps(vacc${N % ACCUMULATORS}, vt${N});
47 }
48 $if ACCUMULATORS > 1:
49 $ACC_SLICE = 1
50 $while ACC_SLICE < ACCUMULATORS:
51 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
52 $if A + ACC_SLICE < ACCUMULATORS:
53 vacc${A} = _mm256_add_ps(vacc${A}, vacc${A + ACC_SLICE});
54 $ACC_SLICE *= 2
55 for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
56 const __m256 vt = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
57 i += 8;
58
59 vacc0 = _mm256_add_ps(vacc0, vt);
60 }
61 if XNN_UNLIKELY(batch != 0) {
62 assert(batch >= 1 * sizeof(uint16_t));
63 assert(batch <= 7 * sizeof(uint16_t));
64 const __m128i vmask = _mm_loadu_si128((const __m128i*) ((uintptr_t) &mask_table[7] - batch));
65 const __m128i vh = _mm_castps_si128(_mm_maskload_ps((const float*) i, vmask));
66 const __m256 vt = _mm256_cvtph_ps(vh);
67 vacc0 = _mm256_add_ps(vacc0, vt);
68 i = (const void*) ((uintptr_t) i + batch);
69 if (batch & (1 * sizeof(uint16_t))) {
70 const __m128i vh = _mm_insert_epi16(_mm_setzero_si128(), (int) unaligned_load_u16(i - 1), 0);
71 const __m256 vt = _mm256_zextps128_ps256(_mm_cvtph_ps(vh));
72 vacc0 = _mm256_add_ps(vacc0, vt);
73 }
74 }
75 __m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc0), _mm256_extractf128_ps(vacc0, 1));
76 vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
77 vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
78 vacc = _mm_mul_ss(vacc, _mm_load_ss(&params->scalar.scale));
79
80 float vout = _mm_cvtss_f32(vacc);
81 *output += vout;
82 }