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/math.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup(
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 <= 1);
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 kc = round_up_po2(kc, 4 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48
49 size_t k = kc;
50
51 while (k >= 16 * sizeof(int8_t)) {
52 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
53 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
54
55 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63
64 const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
65 const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
66
67 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
68 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
70 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
71 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
72 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
74 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
75 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
76 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
78 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
79 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
80 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
82 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
83 const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
84 const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
85
86 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
87 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
89 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
90 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
91 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
93 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
94 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
95 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
97 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
98 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
99 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
101 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
102
103 k -= 16 * sizeof(int8_t);
104 }
105
106 if (k >= 8 * sizeof(int8_t)) {
107 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
108
109 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117
118 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
119
120 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
121 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
122 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
123 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
124 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
125 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
126 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
127 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
128 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
129
130 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
131 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
132 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
133 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
134 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
135 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
136 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
137 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
138
139 k -= 8 * sizeof(int8_t);
140 }
141
142 if XNN_UNLIKELY(k != 0) {
143 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
144
145 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149
150 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
151 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
152 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
153 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
154 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
155 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
156 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
157 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
158 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
159 }
160
161 #if XNN_ARCH_ARM64
162 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
163 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
164 #else
165 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
166 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
167 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
168 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
169 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
170 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
171 #endif
172
173 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
174 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
175
176 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
177 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
178 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
179 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
180
181 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
182 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
183 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
184
185 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
186 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
187 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
188
189 #if XNN_ARCH_ARM64
190 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
191
192
193 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
194 #else
195 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
196
197
198 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
199 #endif
200
201 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
202 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
203
204 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
205 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
206
207 if (nc >= 8) {
208 vst1_s8(c0 + 0, vout0x01234567);
209
210 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
211
212 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
213
214 nc -= 8;
215 } else {
216 // Final case where not all of the 8 columns fit in the destination.
217 if (nc & 4) {
218 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
219 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
220 }
221 if (nc & 2) {
222 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
223 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
224 }
225 if (nc & 1) {
226 vst1_lane_s8(c0, vout0x01234567, 0);
227 }
228
229 nc = 0;
230 }
231 } while (nc != 0);
232 }
233