Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x16c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_2x16c4__neondotfp16arith(
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 <= 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, 4 * 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 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
51 // Loop over groups of 16 columns.
52 do {
53 // Initialize accumulators with bias. 16 bias values are loaded from the
54 // weight matrix, at the start of the group of 16 columns.
55 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
56 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
58 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
59 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
61 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
62 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
64 int32x4_t vacc1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
65 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
66 int32x4_t vacc0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
67 int32x4_t vacc1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
68
69 // Inner accumulation loop along the 16 columns.
70 size_t k = kc;
71 // 2x partial unrolled loop to load 8 bytes at a time.
72 while (k >= 8 * sizeof(int8_t)) {
73 // Load a 2x8 block of activations.
74 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
75 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
76
77 // Load a 8x16 block of weights.
78 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
79 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
80 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
81 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
83 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
84 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
85 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
86 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
87 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
88 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
89 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
90
91 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
92 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
93 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
94 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
95 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
96 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
97 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
98 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
99 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
100 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
101 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
102 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
103 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
104 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
105 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
106 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
107 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
108
109 k -= 8 * sizeof(int8_t);
110 }
111 // Handle up to 4 final positions of `k`
112 if XNN_UNLIKELY(k != 0) {
113 // Load a 2x4 block of activations.
114 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
115 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
116
117 // Load a 4x16 block of weights.
118 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
123 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
124 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
125 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
126
127 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
128 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
129 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
130 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
131 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
132 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
133 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
134 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
135 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
136 }
137
138 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
139 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
140 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
141 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
142 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
143 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
144 float32x4_t vout1x89AB = vcvtq_n_f32_s32(vacc1x89AB, 4);
145 float32x4_t vout1xCDEF = vcvtq_n_f32_s32(vacc1xCDEF, 4);
146 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
147 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
148 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
149 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
150 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
151 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
152 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
153 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
154 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
155
156 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
157 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
158 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
159 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
160
161 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
162 #if XNN_ARCH_ARM64
163 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
164 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
165 #else
166 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
167 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
168 #endif
169 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
170 #if XNN_ARCH_ARM64
171 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
172 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
173 #else
174 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
175 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
176 #endif
177 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
178 #if XNN_ARCH_ARM64
179 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
180 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
181 #else
182 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
183 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
184 #endif
185 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
186 #if XNN_ARCH_ARM64
187 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
188 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
189 #else
190 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
191 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
192 #endif
193
194 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
195 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
196 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
197 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
198
199 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
200 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
201 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
202 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
203 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
204 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
205 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
206 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
207 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
208 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
209 if XNN_LIKELY(nc >= 16) {
210 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
211 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
212 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
213 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
214
215 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
216 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
217
218 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
219 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
220
221 nc -= 16;
222 } else {
223 if (nc & 8) {
224 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
225 vfp16out0x01234567 = vfp16out0x89ABCDEF;
226 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
227 vfp16out1x01234567 = vfp16out1x89ABCDEF;
228 }
229 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
230 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
231 if (nc & 4) {
232 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
233 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
234 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
235 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
236 }
237 if (nc & 2) {
238 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
239 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
240 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
241 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
242 }
243 if (nc & 1) {
244 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
245 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
246 }
247 nc = 0;
248 }
249 } while (nc != 0);
250 }