Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x16c8-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_qc4w_gemm_minmax_ukernel_1x16c8__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 struct xnn_f16_qc4w_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 <= 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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
44
45 // Loop over groups of 16 columns.
46 do {
47 // Initialize accumulators with bias. 16 bias values are loaded from the
48 // weight matrix, at the start of the group of 16 columns.
49 const int32x4_t vinput_zero_point01 = vld1q_dup_s32(&quantization_params[0].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_point01);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point01);
54 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
55 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point01);
56 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point01);
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 // Inner accumulation loop along the 16 columns.
69 size_t k = kc;
70 // 2x partial unrolled loop to load 8 bytes at a time.
71
72 uint64x2x2_t va01x0123456789ABCDEF;
73 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
74 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
75
76 while (k >= 16 * sizeof(int8_t)) {
77 // Load a 1x16 block of activations.
78 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
79
80 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
81 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
82
83 // Load a 16x16 block of weights.
84 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
90 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
91 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
93 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
94 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
95 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
96 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
97 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
98 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
99 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
100 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
101 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
102 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
103 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
104 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
105 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
106 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
107 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
108
109 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
110 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
111 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
112 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
113 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
114 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
115 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
116 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
117 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
118 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
119 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
120 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
121 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
122 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
123 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
124 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
125 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
126
127 k -= 16 * sizeof(int8_t);
128 }
129 // Handle up to 8 final positions of `k`
130 if XNN_UNLIKELY(k != 0) {
131 // Load a 1x8 block of activations.
132 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
133
134 // Load a 16x16 block of weights.
135 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
137 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
138 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
141 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
142 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
143 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
144 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
145 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
146 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
147 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
148 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
149 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
150 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
151
152 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
153 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
154 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
155 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
156 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
157 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
158 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
159 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
160 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
161 }
162
163 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
164 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
165 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
166 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
167 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
168 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
169 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
170 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
171
172 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
173 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
174 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
175 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
176 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
177
178 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
179 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
180 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
181 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
182
183 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
184 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
185 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
186 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
187 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
188 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
189 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
190 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
191
192 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
193 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
194
195 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
196 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
197 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
198 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
199 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
200 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
201 if XNN_LIKELY(nc >= 16) {
202 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
203 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
204
205 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
206
207 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
208
209 nc -= 16;
210 } else {
211 if (nc & 8) {
212 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
213 vfp16out0x01234567 = vfp16out0x89ABCDEF;
214 }
215 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
216 if (nc & 4) {
217 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
218 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
219 }
220 if (nc & 2) {
221 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
222 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
223 }
224 if (nc & 1) {
225 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
226 }
227 nc = 0;
228 }
229 } while (nc != 0);
230 }