1// Copyright 2020 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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11$if SSE == 5:
12  #ifdef __GNUC__
13    #include <x86intrin.h>
14  #else
15    #include <immintrin.h>
16    #include <ammintrin.h>
17  #endif
18$else:
19  #include <immintrin.h>
20
21#include <xnnpack/intrinsics-polyfill.h>
22#include <xnnpack/vadd.h>
23
24
25$ISA = {4: "sse41", 5: "xop"}[SSE]
26void xnn_qs8_vadd_minmax_ukernel__${ISA}_mul32_ld32_x${BATCH_TILE}(
27    size_t n,
28    const int8_t* input_x,
29    const int8_t* input_y,
30    int8_t* output,
31    const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
32{
33  const __m128i vzero_point_product = _mm_load_si128((const __m128i*) params->sse2.zero_point_product);
34  const __m128i vx_multiplier = _mm_load_si128((const __m128i*) params->sse2.x_multiplier);
35  const __m128i vy_multiplier = _mm_load_si128((const __m128i*) params->sse2.y_multiplier);
36  const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
37  const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
38  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
39  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
40  const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
41  const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
42
43  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
44    const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
45    const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
46    $for N in range(4, BATCH_TILE, 4):
47      const __m128i vx${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + ${N}));
48      const __m128i vy${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + ${N}));
49    input_x += ${BATCH_TILE};
50    input_y += ${BATCH_TILE};
51
52    $if SSE == 5:
53      $for N in range(0, BATCH_TILE, 4):
54        __m128i vacc${ABC[N:N+4]} = _mm_macc_epi32(vx${ABC[N:N+4]}, vx_multiplier, vzero_point_product);
55
56      $for N in range(0, BATCH_TILE, 4):
57        vacc${ABC[N:N+4]} = _mm_macc_epi32(vy${ABC[N:N+4]}, vy_multiplier, vacc${ABC[N:N+4]});
58    $else:
59      $for N in range(0, BATCH_TILE, 4):
60        __m128i vacc${ABC[N:N+4]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[N:N+4]}, vx_multiplier));
61
62      $for N in range(0, BATCH_TILE, 4):
63        vacc${ABC[N:N+4]} = _mm_add_epi32(vacc${ABC[N:N+4]}, _mm_mullo_epi32(vy${ABC[N:N+4]}, vy_multiplier));
64
65    $for N in range(0, BATCH_TILE, 4):
66      const __m128i vrem${ABC[N:N+4]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[N:N+4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[N:N+4]}));
67
68    $for N in range(0, BATCH_TILE, 4):
69      vacc${ABC[N:N+4]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[N:N+4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[N:N+4]}, vremainder_threshold));
70
71    $for N in range(0, BATCH_TILE, 8):
72      __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);
73
74    $for N in range(0, BATCH_TILE, 8):
75      vout${ABC[N:N+8]} = _mm_max_epi16(vout${ABC[N:N+8]}, voutput_min);
76
77    $for N in range(0, BATCH_TILE, 8):
78      vout${ABC[N:N+8]} = _mm_min_epi16(vout${ABC[N:N+8]}, voutput_max);
79
80    $for N in range(0, BATCH_TILE, 16):
81      $if N + 8 < BATCH_TILE:
82        const __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
83      $else:
84        const __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
85
86    $if BATCH_TILE >= 16:
87      _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
88    $else:
89      _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
90    $for N in range(16, BATCH_TILE, 16):
91      $if N + 8 < BATCH_TILE:
92        _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]});
93      $else:
94        _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]});
95    output += ${BATCH_TILE};
96  }
97  if XNN_UNLIKELY(n != 0) {
98    ${"do " if BATCH_TILE > 8 else ""}{
99      const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
100      const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
101      const __m128i vx${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
102      const __m128i vy${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4));
103      $if BATCH_TILE > 8:
104        input_x += 8;
105        input_y += 8;
106
107      $if SSE == 5:
108        __m128i vacc${ABC[0:4]} = _mm_macc_epi32(vx${ABC[0:4]}, vx_multiplier, vzero_point_product);
109        __m128i vacc${ABC[4:8]} = _mm_macc_epi32(vx${ABC[4:8]}, vx_multiplier, vzero_point_product);
110
111        vacc${ABC[0:4]} = _mm_macc_epi32(vy${ABC[0:4]}, vy_multiplier, vacc${ABC[0:4]});
112        vacc${ABC[4:8]} = _mm_macc_epi32(vy${ABC[4:8]}, vy_multiplier, vacc${ABC[4:8]});
113      $else:
114        __m128i vacc${ABC[0:4]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[0:4]}, vx_multiplier));
115        __m128i vacc${ABC[4:8]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[4:8]}, vx_multiplier));
116
117        vacc${ABC[0:4]} = _mm_add_epi32(vacc${ABC[0:4]}, _mm_mullo_epi32(vy${ABC[0:4]}, vy_multiplier));
118        vacc${ABC[4:8]} = _mm_add_epi32(vacc${ABC[4:8]}, _mm_mullo_epi32(vy${ABC[4:8]}, vy_multiplier));
119
120      const __m128i vrem${ABC[0:4]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[0:4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[0:4]}));
121      const __m128i vrem${ABC[4:8]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[4:8]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[4:8]}));
122
123      vacc${ABC[0:4]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[0:4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[0:4]}, vremainder_threshold));
124      vacc${ABC[4:8]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[4:8]}, vshift), _mm_cmpgt_epi32(vrem${ABC[4:8]}, vremainder_threshold));
125
126      __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);
127      vout${ABC[0:8]} = _mm_max_epi16(vout${ABC[0:8]}, voutput_min);
128      vout${ABC[0:8]} = _mm_min_epi16(vout${ABC[0:8]}, voutput_max);
129
130      __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
131
132      $if BATCH_TILE > 8:
133        if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
134          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
135          output += 8;
136          n -= 8 * sizeof(int8_t);
137        } else {
138          if (n & (4 * sizeof(int8_t))) {
139            *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
140            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
141            output += 4;
142          }
143          if (n & (2 * sizeof(int8_t))) {
144            *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
145            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
146            output += 2;
147          }
148          if (n & (1 * sizeof(int8_t))) {
149            $if SSE >= 4:
150              *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
151            $else:
152              *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
153          }
154          n = 0;
155        }
156      $else:
157        if (n & (4 * sizeof(int8_t))) {
158          *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
159          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
160          output += 4;
161        }
162        if (n & (2 * sizeof(int8_t))) {
163          *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
164          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
165          output += 2;
166        }
167        if (n & (1 * sizeof(int8_t))) {
168          $if SSE >= 4:
169            *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
170          $else:
171            *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
172        }
173    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
174  }
175}
176