Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-1x16c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x16c8__neoni8mm(
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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47
48 do {
49 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
50 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
54 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
55 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
56 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
58
59 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
60 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
61 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
62 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
63 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB)));
64 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB)));
65 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF)));
66 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF)));
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 } else {
74 a0 = zero_data;
75 }
76 a += 1;
77
78 uint64x2x2_t va01x0123456789ABCDEF;
79 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
80 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
81
82 // Inner accumulation loop along the 16 columns.
83 size_t k = kc;
84 // 2x partial unrolled loop to load 8 bytes at a time.
85 while (k >= 16 * sizeof(int8_t)) {
86 // Load a 1x16 block of activations.
87 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
88
89 // Load a 16x16 block of weights.
90 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
91 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
93 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
97 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
98 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
99 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
100 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
103 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
104 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106
107 // Multiply-accumulate: 1x16 * 16x16 --> 1x16.
108 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
109 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
110 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
111 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
112 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
113 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
114 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
115 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
116 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
117 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
118 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
119 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
120 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
121 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
122 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
123 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
124
125 k -= 16 * sizeof(int8_t);
126 }
127 // Handle up to 8 final positions of `k`
128 if XNN_UNLIKELY(k != 0) {
129 // Load a 1x8 block of activations.
130 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
131
132 // Load a 16x16 block of weights.
133 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
137 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
138 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
141
142 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
143 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
144 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
145 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
146 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
147 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
148 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
149 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
150 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
151 }
152
153 p -= 1 * sizeof(void*);
154 } while (p != 0);
155
156 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
157 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
158 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
159 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
160
161 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
162 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
163 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
164 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
165 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
166 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
167 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
168 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
169 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
170
171 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
172 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
173 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
174 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
175
176 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
177 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
178 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
179 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
180 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
181 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
182 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
183 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
184
185 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
186 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
187 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
188 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
189 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
190 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
191 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
192 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
193
194 if XNN_LIKELY(nc >= 16) {
195 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
196 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
197
198 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
199
200 a = (const int8_t**restrict) ((uintptr_t) a - ks);
201
202 nc -= 16;
203 } else {
204 if (nc & 8) {
205 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
206 vfp16out0x01234567 = vfp16out0x89ABCDEF;
207 }
208 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
209 if (nc & 4) {
210 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
211 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
212 }
213 if (nc & 2) {
214 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
215 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
216 }
217 if (nc & 1) {
218 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
219 }
220 nc = 0;
221 }
222 } while (nc != 0);
223 }