1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/math.h>
16
17
xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal(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)])18 void xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 do {
50 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54 int32x4_t vacc1x01 = vacc0x01;
55 int32x4_t vacc1x23 = vacc0x23;
56 int32x4_t vacc1x45 = vacc0x45;
57 int32x4_t vacc1x67 = vacc0x67;
58
59 size_t k = kc;
60 while (k >= 16 * sizeof(int8_t)) {
61 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
62 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
63 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
64 int8x8_t va1x1 = vld1_s8(a1); a1 += 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 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
77 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
79 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
80 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
81 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
82 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
83 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
84 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
85 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
86 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
87 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
88 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
89 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
90 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
91 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
93 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
94 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
95 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
96 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
97 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
98 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
100 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
101 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
102 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
103 va0x0 = vext_s8(va0x0, va0x0, 4);
104 va0x1 = vext_s8(va0x1, va0x1, 4);
105 va1x0 = vext_s8(va1x0, va1x0, 4);
106 va1x1 = vext_s8(va1x1, va1x1, 4);
107 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
108 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
109 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
110 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
111 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
112 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
113 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
114 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
115 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
116 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
117 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
118 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
119 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
120 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
121 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
122 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
123 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
124 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
125 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
126 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
127 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
128 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
129 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
130 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
132 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
133 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
134 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
135
136 k -= 16 * sizeof(int8_t);
137 }
138 if (k != 0) {
139 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
140 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
141
142 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
143 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
144 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
145 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
146 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
149 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
150
151 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
152 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
153 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
154 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
155 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
156 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
157 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
158 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
159 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
160 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
161 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
162 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
163 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
164 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
165 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
166 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
167 va0x0 = vext_s8(va0x0, va0x0, 4);
168 va1x0 = vext_s8(va1x0, va1x0, 4);
169 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
170 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
171 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
172 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
173 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
174 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
175 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
176 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
177 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
178 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
179 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
180 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
181 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
182 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
183 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
184 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
185
186 }
187
188 #if XNN_ARCH_ARM64
189 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
190 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
191 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
192 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
193 #else
194 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
195 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
196 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
197 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
198 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
199 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
200 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
201 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
202 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
203 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
204 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
205 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
206 #endif
207
208 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
209 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
210 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
211 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
212
213 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
214 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
215 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
216 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
217 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
218
219 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
220 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
221 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
222 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
223 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
224
225 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
226 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
227 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
228 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
229 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
230
231 #if XNN_ARCH_ARM64
232 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
233 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
234
235
236 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
237 #else
238 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
239 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
240
241
242 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
243 #endif
244
245 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
246 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
247
248 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
249 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
250
251 if (nc >= 8) {
252 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
253 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
254
255 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
256 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
257
258 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
259 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
260
261 nc -= 8;
262 } else {
263 // Final case where not all of the 8 columns fit in the destination.
264 if (nc & 4) {
265 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
266 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
267 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
268 }
269 if (nc & 2) {
270 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
271 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
272 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
273 }
274 if (nc & 1) {
275 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
276 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
277 }
278
279 nc = 0;
280 }
281 } while (nc != 0);
282 }
283