Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-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_f32_qb4w_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const struct xnn_f32_qb4w_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 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 float* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
54
55 // Loop over groups of 16 columns.
56 do {
57 // Initialize accumulators with scaled vksum. 16 scaled vksum values are loaded from the
58 // weight matrix, at the start of the group of 16 columns.
59 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
60 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
61 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
62 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
63 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
64 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
65 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
66 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
67 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
68 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
69 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
70 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
71 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);
72
73
74 for (size_t kb=0; kb < kc; kb += bl) {
75 int32x4_t vacc01x01 = vdupq_n_s32(0);
76 int32x4_t vacc01x23 = vdupq_n_s32(0);
77 int32x4_t vacc01x45 = vdupq_n_s32(0);
78 int32x4_t vacc01x67 = vdupq_n_s32(0);
79 int32x4_t vacc01x89 = vdupq_n_s32(0);
80 int32x4_t vacc01xAB = vdupq_n_s32(0);
81 int32x4_t vacc01xCD = vdupq_n_s32(0);
82 int32x4_t vacc01xEF = vdupq_n_s32(0);
83
84 size_t k = bl;
85 // 2x partial unrolled loop to load 8 bytes at a time.
86
87 uint64x2x2_t va01x0123456789ABCDEF;
88 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
89 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
90
91 while (k >= 16 * sizeof(int8_t)) {
92 // Load a 2x16 block of activations.
93 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
94 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
95
96 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
97 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
98
99 // Load a 16x16 block of weights.
100 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
103 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
104 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
107 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
109 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
110 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
111 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
112 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
113 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
114 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
115 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
116 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
117 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
118 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
119 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
120 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
121 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
122 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
123 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
124
125 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
126 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
127 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
128 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
129 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
130 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
131 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
132 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
133 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
134 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
135 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
136 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
137 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
138 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
139 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
140 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
141 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
142
143 k -= 16 * sizeof(int8_t);
144 }
145 // Handle up to 8 final positions of `k`
146 if XNN_UNLIKELY(k != 0) {
147 // Load a 2x8 block of activations.
148 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
149 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
150
151 // Load a 16x16 block of weights.
152 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
156 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
161 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
162 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
163 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
164 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
165 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
166 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
167 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
168
169 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
170 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
171 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
172 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
173 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
174 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
175 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
176 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
177 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
178 }
179
180 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
181 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
182 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
183 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
184 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
185 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
186 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
187 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
188 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
189 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
190 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
191 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
192 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
193 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
194 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
195 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
196 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
197 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
198 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
199 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
200 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
201 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
202 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
203 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
204 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
205 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
206 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
207 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
208 }
209
210 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
211 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
212 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
213 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
214 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
215 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
216 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
217 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
218 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
219
220
221 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
222 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
223 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
224 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
225 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
226 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
227 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
228 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
229 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
230 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
231 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
232 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
233
234 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
235 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
236 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
237 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
238 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
239 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
240 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
241 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
242 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
243
244 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
245 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
246 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
247 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
248 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
249 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
250 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
251 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
252 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
253
254 if XNN_LIKELY(nc >= 16) {
255 vst1q_f32(c0, vout0x0123);
256 vst1q_f32(c0 + 4, vout0x4567);
257 vst1q_f32(c0 + 8, vout0x89AB);
258 vst1q_f32(c0 + 12, vout0xCDEF);
259 vst1q_f32(c1, vout1x0123);
260 vst1q_f32(c1 + 4, vout1x4567);
261 vst1q_f32(c1 + 8, vout1x89AB);
262 vst1q_f32(c1 + 12, vout1xCDEF);
263
264 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
265 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
266
267 c0 = (float*) ((uintptr_t) c0 + cn_stride);
268 c1 = (float*) ((uintptr_t) c1 + cn_stride);
269
270 nc -= 16;
271 } else {
272 if (nc & 8) {
273 vst1q_f32(c0, vout0x0123); c0 += 4;
274 vout0x0123 = vout0x89AB;
275 vst1q_f32(c1, vout1x0123); c1 += 4;
276 vout1x0123 = vout1x89AB;
277 vst1q_f32(c0, vout0x4567); c0 += 4;
278 vout0x4567 = vout0xCDEF;
279 vst1q_f32(c1, vout1x4567); c1 += 4;
280 vout1x4567 = vout1xCDEF;
281 }
282 if (nc & 4) {
283 vst1q_f32(c0, vout0x0123); c0 += 4;
284 vout0x0123 = vout0x4567;
285 vst1q_f32(c1, vout1x0123); c1 += 4;
286 vout1x0123 = vout1x4567;
287 }
288 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
289 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
290 if (nc & 2) {
291 vst1_f32(c0, vout0x01); c0 += 2;
292 vst1_f32(c1, vout1x01); c1 += 2;
293 vout0x01 = vget_high_f32(vout0x0123);
294 vout1x01 = vget_high_f32(vout1x0123);
295 }
296 if (nc & 1) {
297 vst1_lane_f32(c0, vout0x01, 0);
298 vst1_lane_f32(c1, vout1x01, 0);
299 }
300 nc = 0;
301 }
302 } while (nc != 0);
303 }