Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-1x8c2s4-minmax-neonfp16arith-mlal.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-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
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x8c2s4__neonfp16arith(
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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 uint16_t* c0 = (uint16_t*) c;
46
47 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
48 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
53 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 } else {
62 a0 = zero_data;
63 }
64 a += 1;
65
66 size_t k = kc;
67 while (k >= 16 * sizeof(int8_t)) {
68 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
69 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
70
71 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79
80 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
81 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
82 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
83 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
84 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
85 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
86 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
87 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
88 va0x0 = vext_s8(va0x0, va0x0, 2);
89 va0x1 = vext_s8(va0x1, va0x1, 2);
90 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
91 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
93 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
94 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
95 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
96 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
97 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
98 va0x0 = vext_s8(va0x0, va0x0, 2);
99 va0x1 = vext_s8(va0x1, va0x1, 2);
100 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
101 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
102 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
103 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
104 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
105 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
107 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
108 va0x0 = vext_s8(va0x0, va0x0, 2);
109 va0x1 = vext_s8(va0x1, va0x1, 2);
110 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
111 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
112 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
113 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
114 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
115 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
116 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
117 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
118
119 k -= 16 * sizeof(int8_t);
120 }
121 if (k != 0) {
122 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
123
124 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
125 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
126 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
128 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
130 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
131 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
132
133 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
134 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
135 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
136 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
137 va0x0 = vext_s8(va0x0, va0x0, 2);
138 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
139 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
140 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
142 va0x0 = vext_s8(va0x0, va0x0, 2);
143 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
144 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
145 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
147 va0x0 = vext_s8(va0x0, va0x0, 2);
148 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
149 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
150 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
151 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
152
153 }
154
155 p -= 1 * sizeof(void*);
156 } while (p != 0);
157
158 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
159 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
160
161 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
162 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
163
164 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
165 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
166
167 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
168 #if XNN_ARCH_ARM64
169 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
170 #else
171 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
172 #endif
173 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
174 #if XNN_ARCH_ARM64
175 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
176 #else
177 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
178 #endif
179
180 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
181 #if XNN_ARCH_ARM64
182 const uint16x8x2_t voutput_minmax = vld2q_dup_u16((const uint16_t*) &params->scalar.min);
183 const float16x8_t voutput_min = vreinterpretq_f16_u16(voutput_minmax.val[0]);
184 const float16x8_t voutput_max = vreinterpretq_f16_u16(voutput_minmax.val[1]);
185 #else
186 // vld2_dup is to work around aarch32 clang bug with vld1q_dup
187 const uint16x4x2_t vminmax = vld2_dup_u16((const uint16_t*) &params->scalar.min);
188 const float16x8_t voutput_min = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[0], vminmax.val[0]));
189 const float16x8_t voutput_max = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[1], vminmax.val[1]));
190 #endif
191 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
192 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
193 if XNN_LIKELY(nc >= 8) {
194 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
195
196 a = (const int8_t**restrict) ((uintptr_t) a - ks);
197
198 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
199
200 nc -= 8;
201 } else {
202 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
203 if (nc & 4) {
204 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
205 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
206 }
207 if (nc & 2) {
208 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
209 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
210 }
211 if (nc & 1) {
212 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
213 }
214 nc = 0;
215 }
216 } while (nc != 0);
217 }