Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x8c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_1x8c8__neoni8mm(
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, 8 * sizeof(int8_t));
41 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
48
49 // Loop over groups of 8 columns.
50 do {
51 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
52 // weight matrix, at the start of the group of 8 columns.
53 const float32x4_t vinput_zero_point01 = 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_point01);
56 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
57 float32x4_t vout0x4567 = vmulq_f32(vksum4567, vinput_zero_point01);
58
59
60 for (size_t kb=0; kb < kc; kb += bl) {
61 int32x4_t vacc01x01 = vdupq_n_s32(0);
62 int32x4_t vacc01x23 = vdupq_n_s32(0);
63 int32x4_t vacc01x45 = vdupq_n_s32(0);
64 int32x4_t vacc01x67 = vdupq_n_s32(0);
65
66 size_t k = bl;
67 // 2x partial unrolled loop to load 8 bytes at a time.
68
69 uint64x2x2_t va01x0123456789ABCDEF;
70 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
71 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
72
73 while (k >= 16 * sizeof(int8_t)) {
74 // Load a 1x16 block of activations.
75 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
76
77 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
78 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
79
80 // Load a 16x8 block of weights.
81 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
86 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
87 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
88 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
89 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
90 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
91 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
92 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
93
94 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
95 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
96 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
97 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
98 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
99 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
100 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
101 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
102 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
103
104 k -= 16 * sizeof(int8_t);
105 }
106 // Handle up to 8 final positions of `k`
107 if XNN_UNLIKELY(k != 0) {
108 // Load a 1x8 block of activations.
109 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
110
111 // Load a 16x8 block of weights.
112 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
117 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
118 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
119 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
120
121 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
122 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
123 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
124 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
125 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
126 }
127
128 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
129 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
130 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
131 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
132 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
133 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
134 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
135 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
136 }
137
138 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
139 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
140 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
141
142
143 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
144 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
145 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
146 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
147
148 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
149
150 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
151 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
152 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
153 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
154 if XNN_LIKELY(nc >= 8) {
155 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
156
157 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
158
159 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
160
161 nc -= 8;
162 } else {
163 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
164 if (nc & 4) {
165 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
166 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
167 }
168 if (nc & 2) {
169 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
170 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
171 }
172 if (nc & 1) {
173 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
174 }
175 nc = 0;
176 }
177 } while (nc != 0);
178 }