xref: /aosp_15_r20/external/XNNPACK/src/qs8-vcvt/ssse3.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker// Copyright 2022 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker//
3*4bdc9457SAndroid Build Coastguard Worker// This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker// LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE >= 16
7*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE % 16 == 0
8*4bdc9457SAndroid Build Coastguard Worker$SIMD_TILE = BATCH_TILE // 16
9*4bdc9457SAndroid Build Coastguard Worker$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10*4bdc9457SAndroid Build Coastguard Worker#include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker#include <tmmintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/vcvt.h>
16*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/unaligned.h>
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Worker
19*4bdc9457SAndroid Build Coastguard Worker$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
20*4bdc9457SAndroid Build Coastguard Worker$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE]
21*4bdc9457SAndroid Build Coastguard Workervoid xnn_${DATATYPE.lower()}_vcvt_ukernel__ssse3_x${BATCH_TILE}(
22*4bdc9457SAndroid Build Coastguard Worker    size_t n,
23*4bdc9457SAndroid Build Coastguard Worker    const ${XINT8_T}* x,
24*4bdc9457SAndroid Build Coastguard Worker    ${XINT8_T}* y,
25*4bdc9457SAndroid Build Coastguard Worker    const union xnn_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26*4bdc9457SAndroid Build Coastguard Worker{
27*4bdc9457SAndroid Build Coastguard Worker  assert(n != 0);
28*4bdc9457SAndroid Build Coastguard Worker  assert(n % sizeof(${XINT8_T}) == 0);
29*4bdc9457SAndroid Build Coastguard Worker  assert(x != NULL);
30*4bdc9457SAndroid Build Coastguard Worker  assert(y != NULL);
31*4bdc9457SAndroid Build Coastguard Worker
32*4bdc9457SAndroid Build Coastguard Worker  const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
33*4bdc9457SAndroid Build Coastguard Worker  const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
34*4bdc9457SAndroid Build Coastguard Worker  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
35*4bdc9457SAndroid Build Coastguard Worker  $if DATATYPE == "QU8":
36*4bdc9457SAndroid Build Coastguard Worker    const __m128i vzero = _mm_setzero_si128();
37*4bdc9457SAndroid Build Coastguard Worker  $if BATCH_TILE > 8:
38*4bdc9457SAndroid Build Coastguard Worker    for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
39*4bdc9457SAndroid Build Coastguard Worker      const __m128i vx${ABC[0]} = _mm_loadu_si128((const __m128i*) x);
40*4bdc9457SAndroid Build Coastguard Worker      $for N in range(1, SIMD_TILE):
41*4bdc9457SAndroid Build Coastguard Worker        const __m128i vx${ABC[N]} = _mm_loadu_si128((const __m128i*) (x + ${N * 16}));
42*4bdc9457SAndroid Build Coastguard Worker      x += ${BATCH_TILE};
43*4bdc9457SAndroid Build Coastguard Worker
44*4bdc9457SAndroid Build Coastguard Worker      $for N in range(SIMD_TILE):
45*4bdc9457SAndroid Build Coastguard Worker        $if DATATYPE == "QU8":
46*4bdc9457SAndroid Build Coastguard Worker          __m128i vacc${ABC[2*N]} = _mm_unpacklo_epi8(vx${ABC[N]}, vzero);
47*4bdc9457SAndroid Build Coastguard Worker          __m128i vacc${ABC[2*N+1]} = _mm_unpackhi_epi8(vx${ABC[N]}, vzero);
48*4bdc9457SAndroid Build Coastguard Worker        $else:
49*4bdc9457SAndroid Build Coastguard Worker          const __m128i vm${ABC[N]} = _mm_cmpgt_epi8(_mm_setzero_si128(), vx${ABC[N]});
50*4bdc9457SAndroid Build Coastguard Worker          __m128i vacc${ABC[2*N]} = _mm_unpacklo_epi8(vx${ABC[N]}, vm${ABC[N]});
51*4bdc9457SAndroid Build Coastguard Worker          __m128i vacc${ABC[2*N+1]} = _mm_unpackhi_epi8(vx${ABC[N]}, vm${ABC[N]});
52*4bdc9457SAndroid Build Coastguard Worker
53*4bdc9457SAndroid Build Coastguard Worker      $for N in range(2*SIMD_TILE):
54*4bdc9457SAndroid Build Coastguard Worker        vacc${ABC[N]} = _mm_sub_epi16(vinput_zero_point, vacc${ABC[N]});
55*4bdc9457SAndroid Build Coastguard Worker
56*4bdc9457SAndroid Build Coastguard Worker      $for N in range(2*SIMD_TILE):
57*4bdc9457SAndroid Build Coastguard Worker        vacc${ABC[N]} = _mm_slli_epi16(vacc${ABC[N]}, 7);
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Worker      $for N in range(2*SIMD_TILE):
60*4bdc9457SAndroid Build Coastguard Worker        vacc${ABC[N]} = _mm_mulhrs_epi16(vacc${ABC[N]}, vmultiplier);
61*4bdc9457SAndroid Build Coastguard Worker
62*4bdc9457SAndroid Build Coastguard Worker      $for N in range(2*SIMD_TILE):
63*4bdc9457SAndroid Build Coastguard Worker        vacc${ABC[N]} = _mm_adds_epi16(vacc${ABC[N]}, voutput_zero_point);
64*4bdc9457SAndroid Build Coastguard Worker
65*4bdc9457SAndroid Build Coastguard Worker      $for N in range(SIMD_TILE):
66*4bdc9457SAndroid Build Coastguard Worker        const __m128i vy${ABC[N]} = ${_MM_PACKXS_EPI16}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]});
67*4bdc9457SAndroid Build Coastguard Worker
68*4bdc9457SAndroid Build Coastguard Worker      _mm_storeu_si128((__m128i*) y, vy${ABC[0]});
69*4bdc9457SAndroid Build Coastguard Worker      $for N in range(1, SIMD_TILE):
70*4bdc9457SAndroid Build Coastguard Worker        _mm_storeu_si128((__m128i*) (y + ${N * 16}), vy${ABC[N]});
71*4bdc9457SAndroid Build Coastguard Worker      y += ${BATCH_TILE};
72*4bdc9457SAndroid Build Coastguard Worker    }
73*4bdc9457SAndroid Build Coastguard Worker  for (; n >= 16 * sizeof(${XINT8_T}); n -= 16 * sizeof(${XINT8_T})) {
74*4bdc9457SAndroid Build Coastguard Worker    const __m128i vx = _mm_loadu_si128((const __m128i*) x);
75*4bdc9457SAndroid Build Coastguard Worker    x += 16;
76*4bdc9457SAndroid Build Coastguard Worker
77*4bdc9457SAndroid Build Coastguard Worker    $if DATATYPE == "QU8":
78*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
79*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
80*4bdc9457SAndroid Build Coastguard Worker    $else:
81*4bdc9457SAndroid Build Coastguard Worker      const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
82*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
83*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
84*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
85*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
86*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_slli_epi16(vacc_lo, 7);
87*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_slli_epi16(vacc_hi, 7);
88*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
89*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
90*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
91*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
92*4bdc9457SAndroid Build Coastguard Worker
93*4bdc9457SAndroid Build Coastguard Worker    const __m128i vy = ${_MM_PACKXS_EPI16}(vacc_lo, vacc_hi);
94*4bdc9457SAndroid Build Coastguard Worker    _mm_storeu_si128((__m128i*) y, vy);
95*4bdc9457SAndroid Build Coastguard Worker    y += 16;
96*4bdc9457SAndroid Build Coastguard Worker  }
97*4bdc9457SAndroid Build Coastguard Worker  if XNN_UNLIKELY(n != 0) {
98*4bdc9457SAndroid Build Coastguard Worker    assert(n >= 1 * sizeof(${XINT8_T}));
99*4bdc9457SAndroid Build Coastguard Worker    assert(n <= 15 * sizeof(${XINT8_T}));
100*4bdc9457SAndroid Build Coastguard Worker
101*4bdc9457SAndroid Build Coastguard Worker    const __m128i vx = _mm_loadu_si128((const __m128i*) x);
102*4bdc9457SAndroid Build Coastguard Worker
103*4bdc9457SAndroid Build Coastguard Worker    $if DATATYPE == "QU8":
104*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
105*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
106*4bdc9457SAndroid Build Coastguard Worker    $else:
107*4bdc9457SAndroid Build Coastguard Worker      const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
108*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
109*4bdc9457SAndroid Build Coastguard Worker      __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
110*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
111*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
112*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_slli_epi16(vacc_lo, 7);
113*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_slli_epi16(vacc_hi, 7);
114*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
115*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
116*4bdc9457SAndroid Build Coastguard Worker    vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
117*4bdc9457SAndroid Build Coastguard Worker    vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
118*4bdc9457SAndroid Build Coastguard Worker
119*4bdc9457SAndroid Build Coastguard Worker    __m128i vy = ${_MM_PACKXS_EPI16}(vacc_lo, vacc_hi);
120*4bdc9457SAndroid Build Coastguard Worker    if (n & (8 * sizeof(${XINT8_T}))) {
121*4bdc9457SAndroid Build Coastguard Worker      _mm_storel_epi64((__m128i*) y, vy);
122*4bdc9457SAndroid Build Coastguard Worker      vy = _mm_unpackhi_epi64(vy, vy);
123*4bdc9457SAndroid Build Coastguard Worker      y += 8;
124*4bdc9457SAndroid Build Coastguard Worker    }
125*4bdc9457SAndroid Build Coastguard Worker    if (n & (4 * sizeof(${XINT8_T}))) {
126*4bdc9457SAndroid Build Coastguard Worker      unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
127*4bdc9457SAndroid Build Coastguard Worker      vy = _mm_srli_epi64(vy, 32);
128*4bdc9457SAndroid Build Coastguard Worker      y += 4;
129*4bdc9457SAndroid Build Coastguard Worker    }
130*4bdc9457SAndroid Build Coastguard Worker    uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
131*4bdc9457SAndroid Build Coastguard Worker    if (n & (2 * sizeof(${XINT8_T}))) {
132*4bdc9457SAndroid Build Coastguard Worker      unaligned_store_u16(y, (uint16_t) vy_lo);
133*4bdc9457SAndroid Build Coastguard Worker      vy_lo >>= 16;
134*4bdc9457SAndroid Build Coastguard Worker      y += 2;
135*4bdc9457SAndroid Build Coastguard Worker    }
136*4bdc9457SAndroid Build Coastguard Worker    if (n & (1 * sizeof(${XINT8_T}))) {
137*4bdc9457SAndroid Build Coastguard Worker      *y = (${XINT8_T}) vy_lo;
138*4bdc9457SAndroid Build Coastguard Worker    }
139*4bdc9457SAndroid Build Coastguard Worker  }
140*4bdc9457SAndroid Build Coastguard Worker}
141