Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-2x16c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_gemm_minmax_ukernel_2x16c8__neoni8mm(
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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 // Loop over groups of 16 columns.
51 do {
52 // Initialize accumulators with bias. 16 bias values are loaded from the
53 // weight matrix, at the start of the group of 16 columns.
54 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
55 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
57 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
58 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
60 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
61 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
62 const int32x4_t vksumzp0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
63 const int32x4_t vksumzp1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
64 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
65 const int32x4_t vksumzp0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
66 const int32x4_t vksumzp1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
67
68 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
69 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
70 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
71 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
72 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
73 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
74 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
75 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
76
77 // Inner accumulation loop along the 16 columns.
78 size_t k = kc;
79 // 2x partial unrolled loop to load 8 bytes at a time.
80
81 uint64x2x2_t va01x0123456789ABCDEF;
82 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
83 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
84
85 while (k >= 16 * sizeof(int8_t)) {
86 // Load a 2x16 block of activations.
87 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
88 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
89
90 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
91 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
92
93 // Load a 16x16 block of weights.
94 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
97 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
98 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
99 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
100 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
103 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
104 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
107 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
109 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
110
111 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
112 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
113 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
114 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
115 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
116 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
117 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
118 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
119 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
120 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
121 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
122 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
123 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
124 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
125 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
126 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
127 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
128
129 k -= 16 * sizeof(int8_t);
130 }
131 // Handle up to 8 final positions of `k`
132 if XNN_UNLIKELY(k != 0) {
133 // Load a 2x8 block of activations.
134 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
135 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
136
137 // Load a 16x16 block of weights.
138 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
141 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
142 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
143 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
144 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
145 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
146
147 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
148 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
149 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
150 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
151 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
152 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
153 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
154 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
155 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
156 }
157
158 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
159 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
160 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
161 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
162 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
163 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
164 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
165 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
166 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
167 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
168 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
169 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
170 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
171 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
172 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
173 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
174
175 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
176 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
177 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
178 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
179 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
180 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
181 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
182 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
183 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
184
185 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
186 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
187 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
188 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
189
190 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
191 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
192 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
193 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
194 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
195 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
196 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
197 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
198 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
199 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
200 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
201 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
202
203 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
204 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
205 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
206 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
207
208 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
209 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
210 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
211 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
212 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
213 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
214 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
215 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
216 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
217 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
218 if XNN_LIKELY(nc >= 16) {
219 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
220 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
221 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
222 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
223
224 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
225 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
226
227 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
228 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
229
230 nc -= 16;
231 } else {
232 if (nc & 8) {
233 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
234 vfp16out0x01234567 = vfp16out0x89ABCDEF;
235 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
236 vfp16out1x01234567 = vfp16out1x89ABCDEF;
237 }
238 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
239 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
240 if (nc & 4) {
241 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
242 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
243 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
244 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
245 }
246 if (nc & 2) {
247 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
248 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
249 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
250 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
251 }
252 if (nc & 1) {
253 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
254 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
255 }
256 nc = 0;
257 }
258 } while (nc != 0);
259 }