Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c4-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_2x8c4__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 <= 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 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
51 // Loop over groups of 8 columns.
52 do {
53 // Initialize accumulators with bias. 8 bias values are loaded from the
54 // weight matrix, at the start of the group of 8 columns.
55 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
56 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
58 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
59 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
61 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
62
63 // Inner accumulation loop along the 8 columns.
64 size_t k = kc;
65 // 2x partial unrolled loop to load 8 bytes at a time.
66 while (k >= 8 * sizeof(int8_t)) {
67 // Load a 2x8 block of activations.
68 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
69 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
70
71 // Load a 8x8 block of weights.
72 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
73 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
74 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
75 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
76 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
77 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
78
79 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
80 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
81 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
82 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
83 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
84 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
85 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
86 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
87 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
88
89 k -= 8 * sizeof(int8_t);
90 }
91 // Handle up to 4 final positions of `k`
92 if XNN_UNLIKELY(k != 0) {
93 // Load a 2x4 block of activations.
94 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
95 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
96
97 // Load a 4x8 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 vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
101 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
102
103 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
104 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
105 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
106 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
107 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
108 }
109
110 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
111 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
112 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
113 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
114 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
115 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
116 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
117 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
118 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
119
120 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
121 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
122
123 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
124 #if XNN_ARCH_ARM64
125 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
126 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
127 #else
128 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
129 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
130 #endif
131 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
132 #if XNN_ARCH_ARM64
133 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
134 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
135 #else
136 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
137 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
138 #endif
139
140 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
141 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
142
143 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
144 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
145 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
146 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
147 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
148 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
149 if XNN_LIKELY(nc >= 8) {
150 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
151 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
152
153 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
154 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
155
156 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
157 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
158
159 nc -= 8;
160 } else {
161 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
162 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
163 if (nc & 4) {
164 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
165 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
166 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
167 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
168 }
169 if (nc & 2) {
170 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
171 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
172 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
173 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
174 }
175 if (nc & 1) {
176 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
177 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
178 }
179 nc = 0;
180 }
181 } while (nc != 0);
182 }