Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-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_qb4w_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_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, 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 size_t bl = params->scalar.blocksize;
51 assert(bl <= kc);
52 assert(bl != 0);
53 assert(bl % 32 == 0);
54 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
55 // Loop over groups of 16 columns.
56 do {
57 // Initialize accumulators with bias. 16 bias 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 for (size_t kb=0; kb < kc; kb += bl) {
74 int32x4_t vacc0x0123 = vdupq_n_s32(0);
75 int32x4_t vacc1x0123 = vdupq_n_s32(0);
76 int32x4_t vacc0x4567 = vdupq_n_s32(0);
77 int32x4_t vacc1x4567 = vdupq_n_s32(0);
78 int32x4_t vacc0x89AB = vdupq_n_s32(0);
79 int32x4_t vacc1x89AB = vdupq_n_s32(0);
80 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
81 int32x4_t vacc1xCDEF = vdupq_n_s32(0);
82 // Inner accumulation loop along the 16 columns.
83 size_t k = bl;
84 // 2x partial unrolled loop to load 8 bytes at a time.
85 while (k >= 8 * sizeof(int8_t)) {
86 // Load a 2x8 block of activations.
87 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
88 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
89
90 // Load a 8x16 block of weights.
91 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
93 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
96 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
97 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
98 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
99 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
100 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
101 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
102 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
103
104 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
105 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
106 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
107 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
108 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
109 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
110 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
111 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
112 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
113 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
114 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
115 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
116 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
117 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
118 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
119 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
120 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
121
122 k -= 8 * sizeof(int8_t);
123 }
124 // Handle up to 4 final positions of `k`
125 if XNN_UNLIKELY(k != 0) {
126 // Load a 2x4 block of activations.
127 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
128 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
129
130 // Load a 4x16 block of weights.
131 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
135
136 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
137 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
138 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
139 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
140 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
141 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
142 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
143 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
144 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
145 }
146 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
147 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
148 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
149 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
150
151 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
152 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
153 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
154 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
155 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
156 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
157 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
158 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
159 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
160 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
161 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
162 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
163 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
164 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
165 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
166 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
167 }
168
169 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
170 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
171 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
172 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
173 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
174 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
175 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
176 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
177 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
178
179
180 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
181 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
182 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
183 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
184 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
185 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
186 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
187 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
188 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
189 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
190 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
191 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
192
193 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
194 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
195 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
196 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
197
198 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
199 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
200 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
201 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
202 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
203 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
204 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
205 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
206 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
207 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
208 if XNN_LIKELY(nc >= 16) {
209 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
210 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
211 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
212 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
213
214 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
215 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
216
217 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
218 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
219
220 nc -= 16;
221 } else {
222 if (nc & 8) {
223 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
224 vfp16out0x01234567 = vfp16out0x89ABCDEF;
225 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
226 vfp16out1x01234567 = vfp16out1x89ABCDEF;
227 }
228 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
229 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
230 if (nc & 4) {
231 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
232 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
233 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
234 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
235 }
236 if (nc & 2) {
237 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
238 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
239 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
240 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
241 }
242 if (nc & 1) {
243 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
244 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
245 }
246 nc = 0;
247 }
248 } while (nc != 0);
249 }