xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x3-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_up16x3__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_up16x3__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 int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x16_t voutput_max = vld1q_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 >= 16; c -= 16) {
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       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
60       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61 
62       const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
63       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
64       const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
65       const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
66 
67       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
68       int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
69 
70       const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
71       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
72       const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
73       const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74 
75       vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
76       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi1x89ABCDEF, vk1x89ABCDEF);
77 
78       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
79       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
80       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
81       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
82       const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
83       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
84       const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
85       const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
86 
87       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
88       vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
89 
90       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
91       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
92       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
93       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
94 
95       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
96       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
97       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
98       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
99 
100       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
101       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
102       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
103       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
104 
105       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
106       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
107       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
108       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
109 
110       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
111       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
112       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
113       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
114 
115 #if XNN_ARCH_ARM64
116       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
117       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
118 
119       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
120       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
121 
122       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
123 #else  // !XNN_ARCH_ARM64
124       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
125       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
126 
127       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
128       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
129 
130       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
131 #endif  // !XNN_ARCH_ARM64
132 
133       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
134 
135       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
136 
137       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
138     }
139     if XNN_UNLIKELY(c != 0) {
140       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
141       do {
142         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
143         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
144 
145         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
146         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
147 
148         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
149 
150         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
151         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
152 
153         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
154 
155         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
156         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
157         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
158         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
159 
160         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
161 
162         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
163         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
164 
165         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
166         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
167 
168         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
169         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t) + 4 * sizeof(float)));
170         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
171         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
172 
173         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
174         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
175 
176 #if XNN_ARCH_ARM64
177         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
178 #else
179         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
180 #endif
181         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
182 
183         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
184         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
185         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
186 
187         if XNN_LIKELY(c >= 8) {
188           vst1_s8(output, vout01234567); output += 8;
189           c -= 8;
190         } else {
191           if (c & 4) {
192             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
193             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
194           }
195           if (c & 2) {
196             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
197             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
198           }
199           if (c & 1) {
200             vst1_lane_s8(output, vout01234567, 0); output += 1;
201           }
202           c = 0;
203         }
204       } while (c != 0);
205     }
206 
207     output = (int8_t*) ((uintptr_t) output + output_increment);
208   } while (--output_width != 0);
209 }
210