xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/unipass-neon-mul8.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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$assert REQUANTIZATION == "RNDNU"
7$assert DATATYPE == "QU8"
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9$assert CHANNEL_TILE % 8 == 0
10$assert CHANNEL_TILE >= 8
11$assert KERNEL_TILE >= 2
12#include <assert.h>
13
14#include <arm_neon.h>
15
16#include <xnnpack/dwconv.h>
17
18
19void xnn_qu8_dwconv_minmax_rndnu_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__neon_mul8(
20    size_t channels,
21    size_t output_width,
22    const uint8_t** input,
23    const void* weights,
24    uint8_t* output,
25    size_t input_stride,
26    size_t output_increment,
27    size_t input_offset,
28    const uint8_t* zero,
29    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30{
31  assert(channels != 0);
32  assert(output_width != 0);
33
34  const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point);
35  const uint16x8_t vkernel_zero_point16 = vmovl_u8(vkernel_zero_point);
36  const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
37  const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
38  const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
39  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
40  $if CHANNEL_TILE == 8:
41    const uint8x8_t voutput_min = vld1_dup_u8(&params->rndnu_neon.output_min);
42    const uint8x8_t voutput_max = vld1_dup_u8(&params->rndnu_neon.output_max);
43  $else:
44    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
45    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
46  do {
47    $for K in range(KERNEL_TILE):
48      const uint8_t* i${K} = input[${K}];
49      assert(i${K} != NULL);
50      if XNN_UNPREDICTABLE(i${K} != zero) {
51        i${K} = (const uint8_t*) ((uintptr_t) i${K} + input_offset);
52      }
53    input = (const uint8_t**) ((uintptr_t) input + input_stride);
54
55
56    size_t c = channels;
57    const void* w = weights;
58    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
59      $for C in range(0, CHANNEL_TILE, 4):
60        int32x4_t vacc${ABC[C:C+4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61
62
63      $for K in range(KERNEL_TILE):
64        $for C in range(0, CHANNEL_TILE, 8):
65          const uint8x8_t vi${K}x${ABC[C:C+8]} = vld1_u8(i${K}); i${K} += 8;
66          const uint8x8_t vk${K}x${ABC[C:C+8]} = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
67
68        $for C in range(0, CHANNEL_TILE, 8):
69          $if K == 0:
70            uint16x8_t vprod${ABC[C:C+8]} = vmull_u8(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
71          $else:
72            vprod${ABC[C:C+8]} = vmull_u8(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
73          $if KERNEL_TILE == 1:
74            uint16x8_t vsum${ABC[0:8]} = vmovl_u8(vi${K}x${ABC[0:8]});
75          $if K == 1:
76            uint16x8_t vsum${ABC[C:C+8]} = vaddl_u8(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});
77          $elif K > 1:
78            vsum${ABC[C:C+8]} = vaddw_u8(vsum${ABC[C:C+8]}, vi${K}x${ABC[C:C+8]});
79
80        $for C in range(0, CHANNEL_TILE, 8):
81          vacc${ABC[C:C+4]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C:C+4]}), vget_low_u16(vprod${ABC[C:C+8]})));
82          vacc${ABC[C+4:C+8]} = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc${ABC[C+4:C+8]}), vget_high_u16(vprod${ABC[C:C+8]})));
83
84      $for C in range(0, CHANNEL_TILE, 8):
85        vacc${ABC[C:C+4]} = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc${ABC[C:C+4]}), vget_low_u16(vsum${ABC[C:C+8]}), vget_low_u16(vkernel_zero_point16)));
86        vacc${ABC[C+4:C+8]} = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc${ABC[C+4:C+8]}), vget_high_u16(vsum${ABC[C:C+8]}), vget_high_u16(vkernel_zero_point16)));
87
88      $for C in range(0, CHANNEL_TILE, 4):
89        vacc${ABC[C:C+4]} = vshlq_s32(vacc${ABC[C:C+4]}, vright_pre_shift);
90
91      $for C in range(0, CHANNEL_TILE, 4):
92        vacc${ABC[C:C+4]} = vqdmulhq_s32(vacc${ABC[C:C+4]}, vmultiplier);
93
94      $for C in range(0, CHANNEL_TILE, 4):
95        vacc${ABC[C:C+4]} = vrshlq_s32(vacc${ABC[C:C+4]}, vright_post_shift);
96
97#if XNN_ARCH_ARM64
98      $for C in range(0, CHANNEL_TILE, 8):
99        const int16x8_t vacc${ABC[C:C+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]}), voutput_zero_point);
100
101      $for C in range(0, CHANNEL_TILE, 16):
102        $if C + 8 < CHANNEL_TILE:
103          uint8x16_t vout${ABC[C:C+16]} = vqmovun_high_s16(vqmovun_s16(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
104        $else:
105          uint8x8_t vout${ABC[C:C+8]} = vqmovun_s16(vacc${ABC[C:C+8]});
106#else
107      $for C in range(0, CHANNEL_TILE, 8):
108        const int16x8_t vacc${ABC[C:C+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]})), voutput_zero_point);
109
110      $for C in range(0, CHANNEL_TILE, 16):
111        $if C + 8 < CHANNEL_TILE:
112          uint8x16_t vout${ABC[C:C+16]} = vcombine_u8(vqmovun_s16(vacc${ABC[C:C+8]}), vqmovun_s16(vacc${ABC[C+8:C+16]}));
113        $else:
114          uint8x8_t vout${ABC[C:C+8]} = vqmovun_s16(vacc${ABC[C:C+8]});
115#endif
116
117      $for C in range(0, CHANNEL_TILE, 16):
118        $if C + 8 < CHANNEL_TILE:
119          vout${ABC[C:C+16]} = vmaxq_u8(vout${ABC[C:C+16]}, voutput_min);
120        $else:
121          $if CHANNEL_TILE == 8:
122            vout${ABC[C:C+8]} = vmax_u8(vout${ABC[C:C+8]}, voutput_min);
123          $else:
124            vout${ABC[C:C+8]} = vmax_u8(vout${ABC[C:C+8]}, vget_low_u8(voutput_min));
125
126      $for C in range(0, CHANNEL_TILE, 16):
127        $if C + 8 < CHANNEL_TILE:
128          vout${ABC[C:C+16]} = vminq_u8(vout${ABC[C:C+16]}, voutput_max);
129        $else:
130          $if CHANNEL_TILE == 8:
131            vout${ABC[C:C+8]} = vmin_u8(vout${ABC[C:C+8]}, voutput_max);
132          $else:
133            vout${ABC[C:C+8]} = vmin_u8(vout${ABC[C:C+8]}, vget_low_u8(voutput_max));
134
135      $for C in range(0, CHANNEL_TILE, 16):
136        $if C + 8 < CHANNEL_TILE:
137          vst1q_u8(output, vout${ABC[C:C+16]}); output += 16;
138        $else:
139          vst1_u8(output, vout${ABC[C:C+8]}); output += 8;
140    }
141    if XNN_UNLIKELY(c != 0) {
142      $if CHANNEL_TILE > 8:
143        const uint8_t* k = (const uint8_t*) ((const int32_t*) w + ${CHANNEL_TILE});
144      ${"do " if CHANNEL_TILE > 8 else ""}{
145        int32x4_t vacc${ABC[0:4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
146        int32x4_t vacc${ABC[4:8]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
147
148        $for K in range(KERNEL_TILE):
149          $if CHANNEL_TILE > 8:
150            const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K}))); i${K} += 8;
151          $else:
152            const int16x8_t vi${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i${K})));
153          $if CHANNEL_TILE > 8:
154            $if K == 0:
155              const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
156            $else:
157              const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + ${K * CHANNEL_TILE - 8})), vkernel_zero_point));
158          $else:
159            $if K == 0:
160              const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point));
161            $else:
162              const int16x8_t vk${K}x${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + ${K * CHANNEL_TILE})), vkernel_zero_point));
163
164          vacc${ABC[0:4]} = vmlal_s16(vacc${ABC[0:4]}, vget_low_s16(vi${K}x${ABC[0:8]}), vget_low_s16(vk${K}x${ABC[0:8]}));
165          vacc${ABC[4:8]} = vmlal_s16(vacc${ABC[4:8]}, vget_high_s16(vi${K}x${ABC[0:8]}), vget_high_s16(vk${K}x${ABC[0:8]}));
166
167        vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_pre_shift);
168        vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_pre_shift);
169
170        vacc${ABC[0:4]} = vqdmulhq_s32(vacc${ABC[0:4]}, vmultiplier);
171        vacc${ABC[4:8]} = vqdmulhq_s32(vacc${ABC[4:8]}, vmultiplier);
172
173        vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_post_shift);
174        vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_post_shift);
175
176#if XNN_ARCH_ARM64
177        const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]}), voutput_zero_point);
178        uint8x8_t vout${ABC[0:8]} = vqmovun_s16(vacc${ABC[0:8]});
179#else
180        const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]})), voutput_zero_point);
181        uint8x8_t vout${ABC[0:8]} = vqmovun_s16(vacc${ABC[0:8]});
182#endif
183
184        $if CHANNEL_TILE == 8:
185          vout${ABC[0:8]} = vmax_u8(vout${ABC[0:8]}, voutput_min);
186          vout${ABC[0:8]} = vmin_u8(vout${ABC[0:8]}, voutput_max);
187        $else:
188          vout${ABC[0:8]} = vmax_u8(vout${ABC[0:8]}, vget_low_u8(voutput_min));
189          vout${ABC[0:8]} = vmin_u8(vout${ABC[0:8]}, vget_low_u8(voutput_max));
190
191        $if CHANNEL_TILE > 8:
192          if XNN_LIKELY(c >= 8) {
193            vst1_u8(output, vout${ABC[0:8]}); output += 8;
194            c -= 8;
195          } else {
196            if (c & 4) {
197              vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout${ABC[0:8]}), 0); output += 4;
198              vout${ABC[0:8]} = vext_u8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
199            }
200            if (c & 2) {
201              vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout${ABC[0:8]}), 0); output += 2;
202              vout${ABC[0:8]} = vext_u8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
203            }
204            if (c & 1) {
205              vst1_lane_u8(output, vout${ABC[0:8]}, 0); output += 1;
206            }
207            c = 0;
208          }
209        $else:
210          if (c & 4) {
211            vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout${ABC[0:8]}), 0); output += 4;
212            vout${ABC[0:8]} = vext_u8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
213          }
214          if (c & 2) {
215            vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout${ABC[0:8]}), 0); output += 2;
216            vout${ABC[0:8]} = vext_u8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
217          }
218          if (c & 1) {
219            vst1_lane_u8(output, vout${ABC[0:8]}, 0); output += 1;
220          }
221      }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
222    }
223
224    output = (uint8_t*) ((uintptr_t) output + output_increment);
225  } while (--output_width != 0);
226}
227