xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up8x3-minmax-fp32-neonv8-mla8-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul8.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
34   const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     input = (const int8_t**) ((uintptr_t) input + input_stride);
53 
54     size_t c = channels;
55     const void* w = weights;
56     for (; c >= 8; c -= 8) {
57       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
58       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
59 
60       const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
61       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
62 
63       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
64 
65       const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
66       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
67 
68       vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
69 
70       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
71       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
72       const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
73       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74 
75       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
76 
77       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
78       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
79 
80       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
81       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
82 
83       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
84       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
85 
86       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
87       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
88 
89       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
90       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
91 
92 #if XNN_ARCH_ARM64
93       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
94 
95       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
96 
97       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
98 #else  // !XNN_ARCH_ARM64
99       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
100 
101       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
102 
103       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
104 #endif  // !XNN_ARCH_ARM64
105 
106       vout01234567 = vmax_s8(vout01234567, voutput_min);
107 
108       vout01234567 = vmin_s8(vout01234567, voutput_max);
109 
110       vst1_s8(output, vout01234567); output += 8;
111     }
112     if XNN_UNLIKELY(c != 0) {
113       {
114         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
115         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
116 
117         const int8x8_t vi0x01234567 = vld1_s8(i0);
118         const int8x8_t vk0x01234567 = vld1_s8(w);
119 
120         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
121 
122         const int8x8_t vi1x01234567 = vld1_s8(i1);
123         const int8x8_t vk1x01234567 = vld1_s8((const void*) ((const int8_t*) w + 8));
124 
125         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
126 
127         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
128         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
129         const int8x8_t vi2x01234567 = vld1_s8(i2);
130         const int8x8_t vk2x01234567 = vld1_s8((const void*) ((const int8_t*) w + 16));
131 
132         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
133 
134         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
135         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
136 
137         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
138         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
139 
140         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 24 * sizeof(int8_t)));
141         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 24 * sizeof(int8_t) + 4 * sizeof(float)));
142         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
143         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
144 
145         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
146         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
147 
148 #if XNN_ARCH_ARM64
149         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
150 #else
151         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
152 #endif
153         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
154 
155         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
156         vout01234567 = vmax_s8(vout01234567, voutput_min);
157         vout01234567 = vmin_s8(vout01234567, voutput_max);
158 
159         if (c & 4) {
160           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
161           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
162         }
163         if (c & 2) {
164           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
165           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
166         }
167         if (c & 1) {
168           vst1_lane_s8(output, vout01234567, 0); output += 1;
169         }
170       }
171     }
172 
173     output = (int8_t*) ((uintptr_t) output + output_increment);
174   } while (--output_width != 0);
175 }
176