Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x16c4-minmax-neondot.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_f32_qb4w_gemm_minmax_ukernel_1x16c4__neondot(
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 <= 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, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43
44 size_t bl = params->scalar.blocksize;
45 assert(bl <= kc);
46 assert(bl != 0);
47 assert(bl % 32 == 0);
48 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
49 // Loop over groups of 16 columns.
50 do {
51 // Initialize accumulators with bias. 16 bias values are loaded from the
52 // weight matrix, at the start of the group of 16 columns.
53 const float32x4_t vinput_zero_point0 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[0].zero_point));
54 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
55 float32x4_t vout0x0123 = vmulq_f32(vksum0123, vinput_zero_point0);
56 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
57 float32x4_t vout0x4567 = vmulq_f32(vksum4567, vinput_zero_point0);
58 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
59 float32x4_t vout0x89AB = vmulq_f32(vksum89AB, vinput_zero_point0);
60 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
61 float32x4_t vout0xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point0);
62
63 for (size_t kb=0; kb < kc; kb += bl) {
64 int32x4_t vacc0x0123 = vdupq_n_s32(0);
65 int32x4_t vacc0x4567 = vdupq_n_s32(0);
66 int32x4_t vacc0x89AB = vdupq_n_s32(0);
67 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
68 // Inner accumulation loop along the 16 columns.
69 size_t k = bl;
70 // 2x partial unrolled loop to load 8 bytes at a time.
71 while (k >= 8 * sizeof(int8_t)) {
72 // Load a 1x8 block of activations.
73 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
74
75 // Load a 8x16 block of weights.
76 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
77 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
78 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
79 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
80 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
81 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
82 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
83 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
84 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
85 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
86 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
87 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
88
89 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
90 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
91 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
92 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
93 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
94 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
95 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
96 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
97 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
98
99 k -= 8 * sizeof(int8_t);
100 }
101 // Handle up to 4 final positions of `k`
102 if XNN_UNLIKELY(k != 0) {
103 // Load a 1x4 block of activations.
104 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
105
106 // Load a 4x16 block of weights.
107 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
109 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
111
112 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
113 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
114 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
115 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
116 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
117 }
118 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
119 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
120 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
121 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
122
123 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
124 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
125 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
126 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
127 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
128 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
129 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
130 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
131 }
132
133 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
134 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
135 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
136 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
137 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
138
139
140 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
141 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
142 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
143 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
144 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
145 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
146 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
147 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
148
149 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
150 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
151 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
152 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
153 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
154
155 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
156 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
157 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
158 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
159 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
160
161 if XNN_LIKELY(nc >= 16) {
162 vst1q_f32(c0, vout0x0123);
163 vst1q_f32(c0 + 4, vout0x4567);
164 vst1q_f32(c0 + 8, vout0x89AB);
165 vst1q_f32(c0 + 12, vout0xCDEF);
166
167 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
168
169 c0 = (float*) ((uintptr_t) c0 + cn_stride);
170
171 nc -= 16;
172 } else {
173 if (nc & 8) {
174 vst1q_f32(c0, vout0x0123); c0 += 4;
175 vout0x0123 = vout0x89AB;
176 vst1q_f32(c0, vout0x4567); c0 += 4;
177 vout0x4567 = vout0xCDEF;
178 }
179 if (nc & 4) {
180 vst1q_f32(c0, vout0x0123); c0 += 4;
181 vout0x0123 = vout0x4567;
182 }
183 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
184 if (nc & 2) {
185 vst1_f32(c0, vout0x01); c0 += 2;
186 vout0x01 = vget_high_f32(vout0x0123);
187 }
188 if (nc & 1) {
189 vst1_lane_f32(c0, vout0x01, 0);
190 }
191 nc = 0;
192 }
193 } while (nc != 0);
194 }