Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-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_qb4w_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_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 uint16_t* c0 = (uint16_t*) 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 8 columns.
50 do {
51 // Initialize accumulators with bias. 8 bias values are loaded from the
52 // weight matrix, at the start of the group of 8 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
59 for (size_t kb=0; kb < kc; kb += bl) {
60 int32x4_t vacc0x0123 = vdupq_n_s32(0);
61 int32x4_t vacc0x4567 = vdupq_n_s32(0);
62 // Inner accumulation loop along the 8 columns.
63 size_t k = bl;
64 // 2x partial unrolled loop to load 8 bytes at a time.
65 while (k >= 8 * sizeof(int8_t)) {
66 // Load a 1x8 block of activations.
67 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
68
69 // Load a 8x8 block of weights.
70 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
71 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
72 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
73 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
74 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
75 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
76
77 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
78 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
79 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
80 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
81 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
82
83 k -= 8 * sizeof(int8_t);
84 }
85 // Handle up to 4 final positions of `k`
86 if XNN_UNLIKELY(k != 0) {
87 // Load a 1x4 block of activations.
88 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
89
90 // Load a 4x8 block of weights.
91 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
93
94 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
95 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
96 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
97 }
98 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
99 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
100
101 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
102 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
103 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
104 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
105 }
106
107 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
108 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
109 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
110
111
112 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
113 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
114 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
115 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
116
117 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
118
119 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
120 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
121 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
122 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
123 if XNN_LIKELY(nc >= 8) {
124 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
125
126 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
127
128 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
129
130 nc -= 8;
131 } else {
132 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
133 if (nc & 4) {
134 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
135 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
136 }
137 if (nc & 2) {
138 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
139 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
140 }
141 if (nc & 1) {
142 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
143 }
144 nc = 0;
145 }
146 } while (nc != 0);
147 }