xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/neon.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$assert DATATYPE in ["QS8", "QU8"]
7$assert BATCH_TILE % (16 if LD128 else 8) == 0
8$assert BATCH_TILE >= 8
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/vadd.h>
15
16
17$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
18$XINT8X8_T = {"QS8": "int8x8_t", "QU8": "uint8x8_t"}[DATATYPE]
19$XINT8X16_T = {"QS8": "int8x16_t", "QU8": "uint8x16_t"}[DATATYPE]
20$VLD1_X8 = {"QS8": "vld1_s8", "QU8": "vld1_u8"}[DATATYPE]
21$VLD1Q_X8 = {"QS8": "vld1q_s8", "QU8": "vld1q_u8"}[DATATYPE]
22$VLD1_DUP_X8 = {"QS8": "vld1_dup_s8", "QU8": "vld1_dup_u8"}[DATATYPE]
23$VLD1Q_DUP_X8 = {"QS8": "vld1q_dup_s8", "QU8": "vld1q_dup_u8"}[DATATYPE]
24$VST1_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE]
25$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE]
26$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE]
27$VMIN_X8 = {"QS8": "vmin_s8", "QU8": "vmin_u8"}[DATATYPE]
28$VMAX_X8 = {"QS8": "vmax_s8", "QU8": "vmax_u8"}[DATATYPE]
29$VMINQ_X8 = {"QS8": "vminq_s8", "QU8": "vminq_u8"}[DATATYPE]
30$VMAXQ_X8 = {"QS8": "vmaxq_s8", "QU8": "vmaxq_u8"}[DATATYPE]
31$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE]
32$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[DATATYPE]
33$VGET_LOW_X8 = {"QS8": "vget_low_s8", "QU8": "vget_low_u8"}[DATATYPE]
34$VCOMBINE_X8 = {"QS8": "vcombine_s8", "QU8": "vcombine_u8"}[DATATYPE]
35$VREINTERPRET_U32_X8 = {"QS8": "vreinterpret_u32_s8", "QU8": "vreinterpret_u32_u8"}[DATATYPE]
36$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE]
37void xnn_${DATATYPE.lower()}_vadd_minmax_ukernel__neon_${"ld128" if LD128 else "ld64"}_x${BATCH_TILE}(
38    size_t n,
39    const ${XINT8_T}* input_a,
40    const ${XINT8_T}* input_b,
41    ${XINT8_T}* output,
42    const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
43{
44  $if LD128:
45    #if XNN_ARCH_ARM64
46      const ${XINT8X16_T} va_zero_point = ${VLD1Q_DUP_X8}(&params->neon.a_zero_point);
47      const ${XINT8X16_T} vb_zero_point = ${VLD1Q_DUP_X8}(&params->neon.b_zero_point);
48    #else
49      const ${XINT8X8_T} va_zero_point = ${VLD1_DUP_X8}(&params->neon.a_zero_point);
50      const ${XINT8X8_T} vb_zero_point = ${VLD1_DUP_X8}(&params->neon.b_zero_point);
51    #endif
52  $else:
53    const ${XINT8X8_T} va_zero_point = ${VLD1_DUP_X8}(&params->neon.a_zero_point);
54    const ${XINT8X8_T} vb_zero_point = ${VLD1_DUP_X8}(&params->neon.b_zero_point);
55  const int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
56  const int32x4_t vb_multiplier = vld1q_dup_s32(&params->neon.b_multiplier);
57  const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
58  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
59  $if BATCH_TILE >= 16:
60    const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->neon.output_min);
61    const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->neon.output_max);
62  $else:
63    const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->neon.output_min);
64    const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->neon.output_max);
65
66  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
67    $if LD128:
68      $for N in range(0, BATCH_TILE, 16):
69        const ${XINT8X16_T} va${ABC[N:N+16]} = ${VLD1Q_X8}(input_a); input_a += 16;
70        const ${XINT8X16_T} vb${ABC[N:N+16]} = ${VLD1Q_X8}(input_b); input_b += 16;
71
72      #if XNN_ARCH_ARM64
73        $for N in range(0, BATCH_TILE, 16):
74          $if DATATYPE == "QU8":
75            const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va${ABC[N:N+16]}), vget_low_u8(va_zero_point)));
76            const int16x8_t vxa${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_high_u8(va${ABC[N:N+16]}, va_zero_point));
77            const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb${ABC[N:N+16]}), vget_low_u8(vb_zero_point)));
78            const int16x8_t vxb${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_high_u8(vb${ABC[N:N+16]}, vb_zero_point));
79          $else:
80            const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(vget_low_s8(va${ABC[N:N+16]}), vget_low_s8(va_zero_point));
81            const int16x8_t vxa${ABC[N+8:N+16]} = vsubl_high_s8(va${ABC[N:N+16]}, va_zero_point);
82            const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vget_low_s8(vb${ABC[N:N+16]}), vget_low_s8(vb_zero_point));
83            const int16x8_t vxb${ABC[N+8:N+16]} = vsubl_high_s8(vb${ABC[N:N+16]}, vb_zero_point);
84      #else  // !XNN_ARCH_ARM64
85        $for N in range(0, BATCH_TILE, 16):
86          $if DATATYPE == "QU8":
87            const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va${ABC[N:N+16]}), va_zero_point));
88            const int16x8_t vxa${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va${ABC[N:N+16]}), va_zero_point));
89            const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb${ABC[N:N+16]}), vb_zero_point));
90            const int16x8_t vxb${ABC[N+8:N+16]} = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(vb${ABC[N:N+16]}), vb_zero_point));
91          $else:
92            const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(vget_low_s8(va${ABC[N:N+16]}), va_zero_point);
93            const int16x8_t vxa${ABC[N+8:N+16]} = vsubl_s8(vget_high_s8(va${ABC[N:N+16]}), va_zero_point);
94            const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vget_low_s8(vb${ABC[N:N+16]}), vb_zero_point);
95            const int16x8_t vxb${ABC[N+8:N+16]} = vsubl_s8(vget_high_s8(vb${ABC[N:N+16]}), vb_zero_point);
96      #endif  // XNN_ARCH_ARM64
97    $else:
98      $for N in range(0, BATCH_TILE, 8):
99        const ${XINT8X8_T} va${ABC[N:N+8]} = ${VLD1_X8}(input_a); input_a += 8;
100        const ${XINT8X8_T} vb${ABC[N:N+8]} = ${VLD1_X8}(input_b); input_b += 8;
101
102      $for N in range(0, BATCH_TILE, 8):
103        $if DATATYPE == "QU8":
104          const int16x8_t vxa${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[N:N+8]}, va_zero_point));
105          const int16x8_t vxb${ABC[N:N+8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[N:N+8]}, vb_zero_point));
106        $else:
107          const int16x8_t vxa${ABC[N:N+8]} = vsubl_s8(va${ABC[N:N+8]}, va_zero_point);
108          const int16x8_t vxb${ABC[N:N+8]} = vsubl_s8(vb${ABC[N:N+8]}, vb_zero_point);
109
110    $for N in range(0, BATCH_TILE, 8):
111      int32x4_t vacc${ABC[N:N+4]} = vmulq_s32(vmovl_s16(vget_low_s16(vxa${ABC[N:N+8]})), va_multiplier);
112      int32x4_t vacc${ABC[N+4:N+8]} = vmulq_s32(vmovl_s16(vget_high_s16(vxa${ABC[N:N+8]})), va_multiplier);
113
114    $for N in range(0, BATCH_TILE, 8):
115      vacc${ABC[N:N+4]} = vmlaq_s32(vacc${ABC[N:N+4]}, vmovl_s16(vget_low_s16(vxb${ABC[N:N+8]})), vb_multiplier);
116      vacc${ABC[N+4:N+8]} = vmlaq_s32(vacc${ABC[N+4:N+8]}, vmovl_s16(vget_high_s16(vxb${ABC[N:N+8]})), vb_multiplier);
117
118    $for N in range(0, BATCH_TILE, 4):
119      vacc${ABC[N:N+4]} = vrshlq_s32(vacc${ABC[N:N+4]}, vright_shift);
120
121    $for N in range(0, BATCH_TILE, 8):
122      const int16x8_t vacc${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[N:N+4]}), vqmovn_s32(vacc${ABC[N+4:N+8]})), voutput_zero_point);
123
124    $for N in range(0, BATCH_TILE, 16):
125      $if N + 8 < BATCH_TILE:
126        ${XINT8X16_T} vout${ABC[N:N+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${ABC[N+8:N+16]}));
127      $else:
128        ${XINT8X8_T} vout${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${ABC[N:N+8]});
129
130    $for N in range(0, BATCH_TILE, 16):
131      $if N + 8 < BATCH_TILE:
132        vout${ABC[N:N+16]} = ${VMAXQ_X8}(vout${ABC[N:N+16]}, voutput_min);
133      $elif BATCH_TILE >= 16:
134        vout${ABC[N:N+8]} = ${VMAX_X8}(vout${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_min));
135      $else:
136        vout${ABC[N:N+8]} = ${VMAX_X8}(vout${ABC[N:N+8]}, voutput_min);
137
138    $for N in range(0, BATCH_TILE, 16):
139      $if N + 8 < BATCH_TILE:
140        vout${ABC[N:N+16]} = ${VMINQ_X8}(vout${ABC[N:N+16]}, voutput_max);
141      $elif BATCH_TILE >= 16:
142        vout${ABC[N:N+8]} = ${VMIN_X8}(vout${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_max));
143      $else:
144        vout${ABC[N:N+8]} = ${VMIN_X8}(vout${ABC[N:N+8]}, voutput_max);
145
146    $for N in range(0, BATCH_TILE, 16):
147      $if N + 8 < BATCH_TILE:
148        ${VST1Q_X8}(output, vout${ABC[N:N+16]}); output += 16;
149      $else:
150        ${VST1_X8}(output, vout${ABC[N:N+8]}); output += 8;
151  }
152  if XNN_UNLIKELY(n != 0) {
153    ${"do " if BATCH_TILE > 8 else ""}{
154      $if BATCH_TILE > 8:
155        const ${XINT8X8_T} va${ABC[0:8]} = ${VLD1_X8}(input_a); input_a += 8;
156        const ${XINT8X8_T} vb${ABC[0:8]} = ${VLD1_X8}(input_b); input_b += 8;
157      $else:
158        const ${XINT8X8_T} va${ABC[0:8]} = ${VLD1_X8}(input_a);
159        const ${XINT8X8_T} vb${ABC[0:8]} = ${VLD1_X8}(input_b);
160
161      $if LD128:
162        $if DATATYPE == "QU8":
163          #if XNN_ARCH_ARM64
164            const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, vget_low_u8(va_zero_point)));
165            const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vget_low_u8(vb_zero_point)));
166          #else  // !XNN_ARCH_ARM64
167            const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, va_zero_point));
168            const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vb_zero_point));
169          #endif  // XNN_ARCH_ARM64
170        $else:
171          #if XNN_ARCH_ARM64
172            const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, vget_low_s8(va_zero_point));
173            const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vget_low_s8(vb_zero_point));
174          #else  // !XNN_ARCH_ARM64
175            const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, va_zero_point);
176            const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vb_zero_point);
177          #endif  // XNN_ARCH_ARM64
178      $else:
179        $if DATATYPE == "QU8":
180          const int16x8_t vxa${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(va${ABC[0:8]}, va_zero_point));
181          const int16x8_t vxb${ABC[0:8]} = vreinterpretq_s16_u16(vsubl_u8(vb${ABC[0:8]}, vb_zero_point));
182        $else:
183          const int16x8_t vxa${ABC[0:8]} = vsubl_s8(va${ABC[0:8]}, va_zero_point);
184          const int16x8_t vxb${ABC[0:8]} = vsubl_s8(vb${ABC[0:8]}, vb_zero_point);
185
186      int32x4_t vacc${ABC[0:4]} = vmulq_s32(vmovl_s16(vget_low_s16(vxa${ABC[0:8]})), va_multiplier);
187      int32x4_t vacc${ABC[4:8]} = vmulq_s32(vmovl_s16(vget_high_s16(vxa${ABC[0:8]})), va_multiplier);
188
189      vacc${ABC[0:4]} = vmlaq_s32(vacc${ABC[0:4]}, vmovl_s16(vget_low_s16(vxb${ABC[0:8]})), vb_multiplier);
190      vacc${ABC[4:8]} = vmlaq_s32(vacc${ABC[4:8]}, vmovl_s16(vget_high_s16(vxb${ABC[0:8]})), vb_multiplier);
191
192      vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_shift);
193      vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_shift);
194
195      const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]})), voutput_zero_point);
196
197      ${XINT8X8_T} vout${ABC[0:8]} = ${VQMOVXN_S16}(vacc${ABC[0:8]});
198      $if BATCH_TILE >= 16:
199        vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_min));
200        vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, ${VGET_LOW_X8}(voutput_max));
201      $else:
202        vout${ABC[0:8]} = ${VMAX_X8}(vout${ABC[0:8]}, voutput_min);
203        vout${ABC[0:8]} = ${VMIN_X8}(vout${ABC[0:8]}, voutput_max);
204
205      $if BATCH_TILE > 8:
206        if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) {
207          ${VST1_X8}(output, vout${ABC[0:8]}); output += 8;
208          n -= 8 * sizeof(${XINT8_T});
209        } else {
210          if (n & (4 * sizeof(${XINT8_T}))) {
211            vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
212            vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
213          }
214          if (n & (2 * sizeof(${XINT8_T}))) {
215            vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
216            vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
217          }
218          if (n & (1 * sizeof(${XINT8_T}))) {
219            ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0);
220          }
221          n = 0;
222        }
223      $else:
224        if (n & (4 * sizeof(${XINT8_T}))) {
225          vst1_lane_u32((void*) output, ${VREINTERPRET_U32_X8}(vout${ABC[0:8]}), 0); output += 4;
226          vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
227        }
228        if (n & (2 * sizeof(${XINT8_T}))) {
229          vst1_lane_u16((void*) output, ${VREINTERPRET_U16_X8}(vout${ABC[0:8]}), 0); output += 2;
230          vout${ABC[0:8]} = ${VEXT_X8}(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
231        }
232        if (n & (1 * sizeof(${XINT8_T}))) {
233          ${VST1_LANE_X8}(output, vout${ABC[0:8]}, 0);
234        }
235    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
236  }
237}
238