1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neon-mull-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51
52 size_t p = ks;
53 do {
54 const int8_t* restrict a0 = a[0];
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61
62 while (k >= 16 * sizeof(int8_t)) {
63 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
64 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
65
66 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74
75 const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
76 const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
77
78 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
79 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
81 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
82 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
83 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
85 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
86 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
87 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
89 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
90 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
91 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
93 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
94 const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
95 const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
96
97 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
98 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
100 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
101 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
102 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
104 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
105 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
106 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
108 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
109 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
110 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
112 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
113
114 k -= 16 * sizeof(int8_t);
115 }
116
117 if (k >= 8 * sizeof(int8_t)) {
118 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
119
120 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128
129 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
130
131 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
132 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
133 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
134 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
135 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
136 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
137 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
138 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
139 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
140
141 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
142 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
143 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
144 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
145 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
146 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
147 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
148 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
149
150 k -= 8 * sizeof(int8_t);
151 }
152
153 if XNN_UNLIKELY(k != 0) {
154 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
155
156 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160
161 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
162 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
163 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
164 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
165 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
166 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
167 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
168 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
169 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
170 }
171 p -= 1 * sizeof(void*);
172 } while (p != 0);
173
174 #if XNN_ARCH_ARM64
175 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
176 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
177 #else
178 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
179 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
180 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
181 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
182 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
183 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
184 #endif
185
186 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
187 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
188
189 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
190 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
191 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
192
193 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
194 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
195 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
196
197 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
198 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
199 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
200
201 #if XNN_ARCH_ARM64
202 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
203
204
205 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
206 #else
207 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
208
209
210 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
211 #endif
212
213 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
214 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
215
216 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
217 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
218
219 if (nc >= 8) {
220 vst1_s8(c0 + 0, vout0x01234567);
221
222 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
223
224 a = (const int8_t**restrict) ((uintptr_t) a - ks);
225
226 nc -= 8;
227 } else {
228 if (nc & 4) {
229 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
230 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
231 }
232 if (nc & 2) {
233 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
234 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
235 }
236 if (nc & 1) {
237 vst1_lane_s8(c0, vout0x01234567, 0);
238 }
239
240 nc = 0;
241 }
242 } while (nc != 0);
243 }
244