Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x8c4-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_1x8c4__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 8 columns.
46 do {
47 // Initialize accumulators with bias. 8 bias values are loaded from the
48 // weight matrix, at the start of the group of 8 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
55 // Inner accumulation loop along the 8 columns.
56 size_t k = kc;
57 // 2x partial unrolled loop to load 8 bytes at a time.
58 while (k >= 8 * sizeof(int8_t)) {
59 // Load a 1x8 block of activations.
60 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
61
62 // Load a 8x8 block of weights.
63 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
64 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
65 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
66 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
67 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
68 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
69
70 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
71 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
72 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
73 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
74 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
75
76 k -= 8 * sizeof(int8_t);
77 }
78 // Handle up to 4 final positions of `k`
79 if XNN_UNLIKELY(k != 0) {
80 // Load a 1x4 block of activations.
81 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
82
83 // Load a 4x8 block of weights.
84 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
87 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
88
89 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
90 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
91 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
92 }
93
94 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
95 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
96 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
97 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
98 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
99
100 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
101 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
102
103 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
104 #if XNN_ARCH_ARM64
105 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
106 #else
107 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
108 #endif
109 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
110 #if XNN_ARCH_ARM64
111 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
112 #else
113 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
114 #endif
115
116 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
117
118 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
119 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
120 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
121 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
122 if XNN_LIKELY(nc >= 8) {
123 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
124
125 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
126
127 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
128
129 nc -= 8;
130 } else {
131 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
132 if (nc & 4) {
133 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
134 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
135 }
136 if (nc & 2) {
137 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
138 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
139 }
140 if (nc & 1) {
141 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
142 }
143 nc = 0;
144 }
145 } while (nc != 0);
146 }