Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-3x8c4-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_3x8c4__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 <= 3);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
57 // Loop over groups of 8 columns.
58 do {
59 // Initialize accumulators with bias. 8 bias values are loaded from the
60 // weight matrix, at the start of the group of 8 columns.
61 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
62 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
64 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
65 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
66 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
67 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
68 const int32x4_t vinput_zero_point2 = vld1q_dup_s32(&quantization_params[2].zero_point);
69 int32x4_t vacc2x0123 = vmulq_s32(vksum0123, vinput_zero_point2);
70 int32x4_t vacc2x4567 = vmulq_s32(vksum4567, vinput_zero_point2);
71
72 // Inner accumulation loop along the 8 columns.
73 size_t k = kc;
74 // 2x partial unrolled loop to load 8 bytes at a time.
75 while (k >= 8 * sizeof(int8_t)) {
76 // Load a 3x8 block of activations.
77 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
78 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
79 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
80
81 // Load a 8x8 block of weights.
82 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
85 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
86 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
87 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
88
89 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
90 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
91 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
92 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
93 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
94 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
95 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
96 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
97 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
98 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
99 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
100 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
101 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
102
103 k -= 8 * sizeof(int8_t);
104 }
105 // Handle up to 4 final positions of `k`
106 if XNN_UNLIKELY(k != 0) {
107 // Load a 3x4 block of activations.
108 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
109 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
110 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
111
112 // Load a 4x8 block of weights.
113 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
116 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
117
118 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
119 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
120 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
121 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
122 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
123 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
124 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
125 }
126
127 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
128 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
129 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
130 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
131 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
132 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
133 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
134 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
135 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
136 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
137 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
138 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
139 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
140 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
141
142 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
143 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
144
145 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
146 #if XNN_ARCH_ARM64
147 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
148 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
149 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
150 #else
151 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
152 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
153 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
154 #endif
155 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
156 #if XNN_ARCH_ARM64
157 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
158 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
159 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
160 #else
161 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
162 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
163 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
164 #endif
165
166 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
167 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
168 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
169
170 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
171 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
172 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
173 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
174 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
175 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
176 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
177 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
178 if XNN_LIKELY(nc >= 8) {
179 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
180 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
181 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
182
183 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
184 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
185 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
186
187 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
188 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
189 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
190
191 nc -= 8;
192 } else {
193 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
194 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
195 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
196 if (nc & 4) {
197 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
198 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
199 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
200 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
201 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
202 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
203 }
204 if (nc & 2) {
205 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
206 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
207 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
208 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
209 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
210 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
211 }
212 if (nc & 1) {
213 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
214 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
215 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
216 }
217 nc = 0;
218 }
219 } while (nc != 0);
220 }