1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal(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_1x8c4s2__neonv8_mlal(
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 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 do {
45 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
46 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
47 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
48 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
49
50 size_t k = kc;
51 while (k >= 16 * sizeof(int8_t)) {
52 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
53 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
54
55 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
62 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63
64 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
65 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
66 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
67 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
68 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
69 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
70 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
71 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
72 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
73 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
74 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
75 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
76 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
77 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
79 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
80 va0x0 = vext_s8(va0x0, va0x0, 4);
81 va0x1 = vext_s8(va0x1, va0x1, 4);
82 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
83 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
84 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
85 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
86 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
87 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
89 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
90 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
91 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
93 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
94 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
95 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
96 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
97 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
98
99 k -= 16 * sizeof(int8_t);
100 }
101 if (k != 0) {
102 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
103
104 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112
113 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
114 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
115 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
116 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
117 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
118 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
119 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
120 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
121 va0x0 = vext_s8(va0x0, va0x0, 4);
122 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
123 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
124 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
125 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
126 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
127 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
128 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
129 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
130
131 }
132
133 #if XNN_ARCH_ARM64
134 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
135 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
136 #else
137 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
138 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
139 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
140 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
141 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
142 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
143 #endif
144
145 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
146 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
147
148 const float32x4_t vscale0123 = vld1q_f32(w); w = (const float*) w + 4;
149 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
150 const float32x4_t vscale4567 = vld1q_f32(w); w = (const float*) w + 4;
151 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
152
153 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
154 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
155
156 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
157 #if XNN_ARCH_ARM64
158 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
159
160 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
161
162 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
163 #else
164 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
165
166 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
167
168 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
169 #endif
170
171 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
172 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
173
174 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
175 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
176
177 if (nc >= 8) {
178 vst1_s8(c0 + 0, vout0x01234567);
179
180 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
181
182 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
183
184 nc -= 8;
185 } else {
186 // Final case where not all of the 8 columns fit in the destination.
187 if (nc & 4) {
188 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
189 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
190 }
191 if (nc & 2) {
192 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
193 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
194 }
195 if (nc & 1) {
196 vst1_lane_s8(c0, vout0x01234567, 0);
197 }
198
199 nc = 0;
200 }
201 } while (nc != 0);
202 }
203