Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-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_qc4w_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 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 <= 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 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
45 // Loop over groups of 16 columns.
46 do {
47 // Initialize accumulators with bias. 16 bias values are loaded from the
48 // weight matrix, at the start of the group of 16 columns.
49 const int32x4_t vinput_zero_point0 = vld1q_dup_s32(&quantization_params[0].zero_point);
50 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point0);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point0);
54 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point0);
56 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point0);
58
59 // Inner accumulation loop along the 16 columns.
60 size_t k = kc;
61 // 2x partial unrolled loop to load 8 bytes at a time.
62 while (k >= 8 * sizeof(int8_t)) {
63 // Load a 1x8 block of activations.
64 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
65
66 // Load a 8x16 block of weights.
67 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
68 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
69 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
70 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
71 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
72 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
73 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
74 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
75 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
76 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
77 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
78 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
79
80 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
81 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
82 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
83 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
84 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
85 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
86 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
87 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
88 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
89
90 k -= 8 * sizeof(int8_t);
91 }
92 // Handle up to 4 final positions of `k`
93 if XNN_UNLIKELY(k != 0) {
94 // Load a 1x4 block of activations.
95 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
96
97 // Load a 4x16 block of weights.
98 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
99 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
100 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
103 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
104 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
105 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
106
107 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
108 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
109 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
110 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
111 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
112 }
113
114 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
115 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
116 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
117 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
118 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
119 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
120 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
121 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
122 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
123
124 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
125 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
126 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
127 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
128
129 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
130 #if XNN_ARCH_ARM64
131 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
132 #else
133 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
134 #endif
135 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
136 #if XNN_ARCH_ARM64
137 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
138 #else
139 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
140 #endif
141 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
142 #if XNN_ARCH_ARM64
143 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
144 #else
145 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
146 #endif
147 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
148 #if XNN_ARCH_ARM64
149 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
150 #else
151 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
152 #endif
153
154 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
155 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
156
157 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
158 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
159 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
160 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
161 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
162 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
163 if XNN_LIKELY(nc >= 16) {
164 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
165 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
166
167 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
168
169 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
170
171 nc -= 16;
172 } else {
173 if (nc & 8) {
174 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
175 vfp16out0x01234567 = vfp16out0x89ABCDEF;
176 }
177 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
178 if (nc & 4) {
179 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
180 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
181 }
182 if (nc & 2) {
183 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
184 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
185 }
186 if (nc & 1) {
187 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
188 }
189 nc = 0;
190 }
191 } while (nc != 0);
192 }