Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-1x16c4-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_qc8w_gemm_minmax_ukernel_1x16c4__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 union xnn_f16_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 uint16_t* c0 = (uint16_t*) c;
43
44 // Loop over groups of 16 columns.
45 do {
46 // Initialize accumulators with bias. 16 bias values are loaded from the
47 // weight matrix, at the start of the group of 16 columns.
48 const int32x4_t vinput_zero_point0 = vld1q_dup_s32(&quantization_params[0].zero_point);
49 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
50 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point0);
51 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point0);
53 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point0);
55 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point0);
57
58 // Inner accumulation loop along the 16 columns.
59 size_t k = kc;
60 // 2x partial unrolled loop to load 8 bytes at a time.
61 while (k >= 8 * sizeof(int8_t)) {
62 // Load a 1x8 block of activations.
63 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
64
65 // Load a 8x16 block of weights.
66 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
67 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
68 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
69 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
70 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
71 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
72 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
73 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
74
75 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
76 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
77 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
78 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
79 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
80 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
81 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
82 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
83 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
84
85 k -= 8 * sizeof(int8_t);
86 }
87 // Handle up to 4 final positions of `k`
88 if XNN_UNLIKELY(k != 0) {
89 // Load a 1x4 block of activations.
90 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
91
92 // Load a 4x16 block of weights.
93 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
97
98 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
99 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
100 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
101 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
102 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
103 }
104
105 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
106 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
107 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
108 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
109 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
110 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
111 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
112 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
113 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
114
115 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
116 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
117 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
118 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
119
120 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
121 #if XNN_ARCH_ARM64
122 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
123 #else
124 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
125 #endif
126 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
127 #if XNN_ARCH_ARM64
128 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
129 #else
130 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
131 #endif
132 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
133 #if XNN_ARCH_ARM64
134 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
135 #else
136 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
137 #endif
138 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
139 #if XNN_ARCH_ARM64
140 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
141 #else
142 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
143 #endif
144
145 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
146 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
147
148 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
149 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
150 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
151 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
152 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
153 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
154 if XNN_LIKELY(nc >= 16) {
155 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
156 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
157
158 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
159
160 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
161
162 nc -= 16;
163 } else {
164 if (nc & 8) {
165 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
166 vfp16out0x01234567 = vfp16out0x89ABCDEF;
167 }
168 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
169 if (nc & 4) {
170 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
171 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
172 }
173 if (nc & 2) {
174 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
175 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
176 }
177 if (nc & 1) {
178 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
179 }
180 nc = 0;
181 }
182 } while (nc != 0);
183 }