1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mlal-lane.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/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane_prfm(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane_prfm(
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_qc8_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 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51 int32x4_t vacc1x0123 = vacc0x0123;
52 int32x4_t vacc1x4567 = vacc0x4567;
53
54 size_t k = kc;
55 while (k >= 8 * sizeof(int8_t)) {
56 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
57 const int16x8_t vxa0 = vmovl_s8(va0);
58 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
59 const int16x8_t vxa1 = vmovl_s8(va1);
60
61 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
62 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
63
64 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
65 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
66 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
67 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
68 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
69 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
70
71 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
72 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
73 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
74 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
75 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
76 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
77
78 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
79 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
80 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
81 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
82 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
83 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
84
85 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
86 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
87 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
88 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
89
90 __builtin_prefetch((const int8_t*) w + 448);
91
92 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
93 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
94
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
96 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
97 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
98 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
99 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
100 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
101
102 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
103 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
104 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
105 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
106 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
107 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
108
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
111 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
112 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
113 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
114 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
115
116 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
117 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
118 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
119 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
120
121 k -= 8 * sizeof(int8_t);
122 }
123 if XNN_UNLIKELY(k != 0) {
124 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
125 const int16x8_t vxa0 = vmovl_s8(va0);
126 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
127 const int16x8_t vxa1 = vmovl_s8(va1);
128
129 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
130 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
131
132 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
133 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
134 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
135 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
136
137 if (k >= 2 * sizeof(int8_t)) {
138 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
139 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
140
141 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
142 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
143 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
144 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
145
146 if (k > 2 * sizeof(int8_t)) {
147 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
148 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
149
150 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
151 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
152 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
153 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
154
155 if (k >= 4 * sizeof(int8_t)) {
156 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
157 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
161 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
162 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
163
164 if (k > 4 * sizeof(int8_t)) {
165 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
166 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
167
168 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
169 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
170 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
171 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
172
173 if (k >= 6 * sizeof(int8_t)) {
174 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
175 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
176
177 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
178 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
179 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
180 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
181
182 if (k > 6 * sizeof(int8_t)) {
183 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
184 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
185
186 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
187 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
188 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
189 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
190 }
191 }
192 }
193 }
194 }
195 }
196 }
197
198 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
199 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
200 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
201 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
202
203 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
204 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
205 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
206 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
207 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
208 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
209
210 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
211 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
212 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
213 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
214 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
215
216 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
217 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
218 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
219 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
220 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
221
222 #if XNN_ARCH_ARM64
223 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
224 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
225
226
227 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
228 #else
229 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
230 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
231
232
233 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
234 #endif
235
236 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
237 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
238
239 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
240 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
241
242 if (nc >= 8) {
243 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
244 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
245
246 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
247 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
248
249 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
250 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
251
252 nc -= 8;
253 } else {
254 if (nc & 4) {
255 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
256 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
257 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
258 }
259 if (nc & 2) {
260 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
261 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
262 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
263 }
264 if (nc & 1) {
265 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
266 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
267 }
268
269 nc = 0;
270 }
271 } while (nc != 0);
272 }
273