1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 do {
47 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
48 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
49 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
50 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51
52 size_t p = ks;
53 do {
54 const int8_t* restrict a0 = a[0];
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61 while (k >= 16 * sizeof(int8_t)) {
62 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
64
65 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73
74 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
75 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
76 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
77 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
78 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
79 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
80 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
81 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
82 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
83 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
84 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
85 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
86 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
87 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
89 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
90 va0x0 = vext_s8(va0x0, va0x0, 4);
91 va0x1 = vext_s8(va0x1, va0x1, 4);
92 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
93 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
94 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
95 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
96 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
97 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
98 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
99 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
100 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
101 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
102 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
103 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
104 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
105 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
107 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
108
109 k -= 16 * sizeof(int8_t);
110 }
111 if (k != 0) {
112 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
113
114 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
122
123 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
124 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
125 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
126 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
127 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
128 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
129 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
130 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
131 va0x0 = vext_s8(va0x0, va0x0, 4);
132 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
133 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
134 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
135 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
136 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
137 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
138 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
139 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
140
141 }
142
143 p -= 1 * sizeof(void*);
144 } while (p != 0);
145
146 #if XNN_ARCH_ARM64
147 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
148 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
149 #else
150 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
151 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
152 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
153 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
154 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
155 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
156 #endif
157
158 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
159 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
160
161 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
162 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
163 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
164
165 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
166 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
167 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
168
169 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
170 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
171 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
172
173 #if XNN_ARCH_ARM64
174 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
175
176
177 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
178 #else
179 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
180
181
182 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
183 #endif
184
185 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
186 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
187
188 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
189 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
190
191 if (nc >= 8) {
192 vst1_s8(c0 + 0, vout0x01234567);
193
194 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
195
196 a = (const int8_t**restrict) ((uintptr_t) a - ks);
197
198 nc -= 8;
199 } else {
200 if (nc & 4) {
201 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
202 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
203 }
204 if (nc & 2) {
205 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
206 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
207 }
208 if (nc & 1) {
209 vst1_lane_s8(c0, vout0x01234567, 0);
210 }
211
212 nc = 0;
213 }
214 } while (nc != 0);
215 }
216