1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params 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(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43
44 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->fp32_neonv8.kernel_zero_point[0]);
45
46 // Loop over groups of 16 columns.
47 do {
48 // Initialize accumulators with bias. 16 bias values are loaded from the
49 // weight matrix, at the start of the group of 16 columns.
50 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
51 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
52 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
54 uint32x2_t vnacc0 = vmov_n_u32(0);
55
56 // Inner accumulation loop along the 16 columns.
57 size_t k = kc;
58 // 2x partial unrolled loop to load 8 bytes at a time.
59 while (k >= 8 * sizeof(uint8_t)) {
60 // Load a 1x8 block of activations.
61 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
62
63 // Load a 8x16 block of weights.
64 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
65 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
66 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
67 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
68 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
69 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
70 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
71 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
72
73 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
74 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
75 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
76 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
77 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
78 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
79 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
80 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
81 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
82 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
83
84 k -= 8 * sizeof(uint8_t);
85 }
86 // Handle up to 4 final positions of `k`
87 if XNN_UNLIKELY(k != 0) {
88 // Load a 1x4 block of activations.
89 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
90
91 // Load a 4x16 block of weights.
92 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
96
97 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
98 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
99 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
100 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
101 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
102 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
103 }
104
105 // Subtract zero point from accumulators.
106 vnacc0 = vpadd_u32(vnacc0, vnacc0);
107 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
108 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
109 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
110 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
111 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
112
113 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
114 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
115 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
116 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
117
118 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
119 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
120 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
121 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
122 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
123
124 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
125 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
126 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
127 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
128
129 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
130 #if XNN_ARCH_ARM64
131 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
132 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
133
134 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
135 #else
136 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
137 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
138
139 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
140 #endif
141 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
142 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
143
144 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
145
146 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
147
148 if (nc >= 16) {
149 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
150
151 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
152
153 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
154
155 nc -= 16;
156 } else {
157 uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
158 if (nc & 8) {
159 vst1_u8(c0, vout0x01234567); c0 += 8;
160 vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
161 }
162 if (nc & 4) {
163 vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
164 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
165 }
166 if (nc & 2) {
167 vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
168 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
169 }
170 if (nc & 1) {
171 vst1_lane_u8(c0, vout0x01234567, 0);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }
178