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_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r(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)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r(
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_qc8_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 int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
54 const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
55 const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
56 const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
57
58 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66
67 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
68 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
69
70 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
71 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
73 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
74 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
75 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
77 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
78 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
79 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
81 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
82 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
83 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
85 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
86 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
87 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
88
89 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
90 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
92 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
93 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
94 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
96 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
97 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
98 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
100 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
101 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
102 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
104 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
105
106 k -= 16 * sizeof(int8_t);
107 }
108
109 if (k >= 8 * sizeof(int8_t)) {
110 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
111 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
112
113 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121
122 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
123
124 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
125 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
126 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
127 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
128 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
129 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
130 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
131 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
132 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
133
134 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
135 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
136 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
137 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
138 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
139 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
140 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
141 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
142
143 k -= 8 * sizeof(int8_t);
144 }
145
146 if XNN_UNLIKELY(k != 0) {
147 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
148
149 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153
154 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
155 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
156 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
157 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
158 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
159 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
160 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
161 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
162 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
163 }
164
165 #if XNN_ARCH_ARM64
166 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
167 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
168 #else
169 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
170 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
171 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
172 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
173 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
174 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
175 #endif
176
177 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
178 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
179
180 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
181 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
182 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
183 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
184
185 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
186 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
187
188 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
189 #if XNN_ARCH_ARM64
190 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
191
192 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
193
194 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
195 #else
196 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
197
198 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
199
200 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
201 #endif
202
203 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
204 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
205
206 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
207 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
208
209 if (nc >= 8) {
210 vst1_s8(c0 + 0, vout0x01234567);
211
212 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
213
214 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
215
216 nc -= 8;
217 } else {
218 // Final case where not all of the 8 columns fit in the destination.
219 if (nc & 4) {
220 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
221 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
222 }
223 if (nc & 2) {
224 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
225 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
226 }
227 if (nc & 1) {
228 vst1_lane_s8(c0, vout0x01234567, 0);
229 }
230
231 nc = 0;
232 }
233 } while (nc != 0);
234 }
235