1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-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
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r(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)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45
46 size_t k = kc;
47
48 while (k >= 16 * sizeof(int8_t)) {
49 const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
50 const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
51 const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
52 const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
53 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
54 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61
62 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
63 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
64
65 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
66 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
68 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
69 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
70 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
72 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
73 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
74 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
75
76 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
77 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
79 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
80 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
81 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
83 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
84 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
85 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
86
87 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
88 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
90 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
91 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
92 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
94 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
95 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
96 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
97
98 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
99 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
101 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
102 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
103 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
105 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
106
107 k -= 16 * sizeof(int8_t);
108 }
109
110 if (k >= 8 * sizeof(int8_t)) {
111 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
112 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
113
114 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122
123 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
124
125 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
126 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
127 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
128 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
129 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
130
131 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
132 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
133 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
134 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
135 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
136
137 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
138 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
139 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
140 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
141 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
142
143 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
144 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
145 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
147
148 k -= 8 * sizeof(int8_t);
149 }
150
151 if XNN_UNLIKELY(k != 0) {
152 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
153
154 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156
157 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
158 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
159 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
160 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
161 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
162
163 if (k > 2 * sizeof(int8_t)) {
164 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166
167 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
168 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
169 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
170 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
171 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
172
173 if (k > 4 * sizeof(int8_t)) {
174 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176
177 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
178 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
179 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
180 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
182 }
183 }
184 }
185
186 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
187 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
188
189 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
190 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
191 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
192 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
193
194 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
195 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
196 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
197
198 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
199 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
200 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
201
202 #if XNN_ARCH_ARM64
203 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
204
205
206 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
207 #else
208 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
209
210
211 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
212 #endif
213
214 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
215 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
216
217 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
218 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
219
220 if (nc >= 8) {
221 vst1_s8(c0 + 0, vout0x01234567);
222
223 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
224
225 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
226
227 nc -= 8;
228 } else {
229 // Final case where not all of the 8 columns fit in the destination.
230 if (nc & 4) {
231 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
232 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
233 }
234 if (nc & 2) {
235 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
236 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
237 }
238 if (nc & 1) {
239 vst1_lane_s8(c0, vout0x01234567, 0);
240 }
241
242 nc = 0;
243 }
244 } while (nc != 0);
245 }
246