1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neon-mull-shuffle.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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal(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)])19 void xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 int8_t* c0 = c;
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 do {
48 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
49 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
50 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52
53 size_t p = ks;
54 do {
55 const int8_t* restrict a0 = a[0];
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = kc;
62 while (k >= 16 * sizeof(int8_t)) {
63 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
64 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
65
66 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74
75 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
76 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
77 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
78 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
79 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
80 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
81 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
82 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
83 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
84 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
85 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
86 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
87 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
88 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
89 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
90 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
91 va0x0 = vext_s8(va0x0, va0x0, 4);
92 va0x1 = vext_s8(va0x1, va0x1, 4);
93 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
94 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
95 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
96 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
97 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
98 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
100 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
101 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
102 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
103 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
104 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
105 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
106 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
107 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
108 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
109
110 k -= 16 * sizeof(int8_t);
111 }
112 if (k != 0) {
113 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
114
115 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
123
124 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
125 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
126 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
127 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
128 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
129 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
130 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
131 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
132 va0x0 = vext_s8(va0x0, va0x0, 4);
133 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
134 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
135 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
136 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
137 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
138 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
139 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
140 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
141
142 }
143
144 p -= 1 * sizeof(void*);
145 } while (p != 0);
146
147 #if XNN_ARCH_ARM64
148 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
149 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
150 #else
151 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
152 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
153 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
154 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
155 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
156 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
157 #endif
158
159 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
160 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
161
162 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
163 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
164 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
165
166 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
167 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
168
169 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
170 #if XNN_ARCH_ARM64
171 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
172
173 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
174
175 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
176 #else
177 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
178
179 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
180
181 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
182 #endif
183
184 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
185 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
186
187 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
188 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
189
190 if (nc >= 8) {
191 vst1_s8(c0 + 0, vout0x01234567);
192
193 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
194
195 a = (const int8_t**restrict) ((uintptr_t) a - ks);
196
197 nc -= 8;
198 } else {
199 if (nc & 4) {
200 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
201 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
202 }
203 if (nc & 2) {
204 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
205 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
206 }
207 if (nc & 1) {
208 vst1_lane_s8(c0, vout0x01234567, 0);
209 }
210
211 nc = 0;
212 }
213 } while (nc != 0);
214 }
215