1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43
44 do {
45 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49
50 size_t k = kc;
51
52 while (k >= 16 * sizeof(int8_t)) {
53 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
54 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
55
56 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64
65 const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
66 const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
67
68 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
69 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
71 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
72 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
73 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
75 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
76 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
77 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
79 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
80 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
81 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
83 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
84 const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
85 const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
86
87 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
88 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
90 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
91 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
92 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
94 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
95 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
96 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
98 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
99 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
100 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
102 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
103
104 k -= 16 * sizeof(int8_t);
105 }
106
107 if (k >= 8 * sizeof(int8_t)) {
108 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
109
110 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118
119 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
120
121 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
122 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
123 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
124 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
125 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
126 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
127 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
128 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
129 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
130
131 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
132 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
133 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
134 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
135 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
136 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
137 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
138 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
139
140 k -= 8 * sizeof(int8_t);
141 }
142
143 if XNN_UNLIKELY(k != 0) {
144 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
145
146 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150
151 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
152 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
153 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
154 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
155 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
156 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
157 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
158 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
159 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
160 }
161
162 #if XNN_ARCH_ARM64
163 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
164 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
165 #else
166 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
167 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
168 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
169 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
170 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
171 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
172 #endif
173
174 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
175 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
176
177 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
178 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
179 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
180
181 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
182 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
183
184 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
185 #if XNN_ARCH_ARM64
186 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
187
188 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
189
190 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
191 #else
192 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
193
194 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
195
196 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
197 #endif
198
199 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
200 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
201
202 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
203 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
204
205 if (nc >= 8) {
206 vst1_s8(c0 + 0, vout0x01234567);
207
208 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
209
210 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
211
212 nc -= 8;
213 } else {
214 // Final case where not all of the 8 columns fit in the destination.
215 if (nc & 4) {
216 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
217 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
218 }
219 if (nc & 2) {
220 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
221 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
222 }
223 if (nc & 1) {
224 vst1_lane_s8(c0, vout0x01234567, 0);
225 }
226
227 nc = 0;
228 }
229 } while (nc != 0);
230 }
231