xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-igemm/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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const uint8_t** restrict a,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const uint8_t* zero,
30     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(uint8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44   uint8_t* c0 = c;
45 
46   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
47 
48   do {
49     // Initialize accumulators with bias. 8 bias values are loaded from the
50     // weight matrix, at the start of the group of 8 columns.
51     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
52     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53     uint32x2_t vnacc0 = vmov_n_u32(0);
54 
55     size_t p = ks;
56     do {
57       const uint8_t* restrict a0 = a[0];
58       if XNN_UNPREDICTABLE(a0 != zero) {
59         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
60       }
61       a += 1;
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(uint8_t)) {
67         // Load a 1x8 block of activations.
68         const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
69 
70         // Load a 8x8 block of weights.
71         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
72         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
73         const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
74         const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
75 
76         // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
77         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
78         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
79         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
80         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
81         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
82 
83         k -= 8 * sizeof(uint8_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 uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
89 
90         // Load a 4x8 block of weights.
91         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 
94         // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
95         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
96         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
97         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
98       }
99       p -= 1 * sizeof(void*);
100     } while (p != 0);
101 
102     // Subtract zero point from accumulators.
103     vnacc0 = vpadd_u32(vnacc0, vnacc0);
104     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
105     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
106     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
107 
108     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
109     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
110     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
111 
112     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
113     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
114 
115     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
116     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
117 
118     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
119     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
120 
121     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
122 #if XNN_ARCH_ARM64
123     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
124 
125     uint8x8_t vout0x01234567 = vqmovun_s16(vacc0x01234567);
126 #else
127     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
128 
129     uint8x8_t vout0x01234567 = vqmovun_s16(vacc0x01234567);
130 #endif
131     const uint8x8_t voutput_min = vld1_dup_u8(&params->rndnu_neon.output_min);
132     const uint8x8_t voutput_max = vld1_dup_u8(&params->rndnu_neon.output_max);
133 
134     vout0x01234567 = vmax_u8(vout0x01234567, voutput_min);
135 
136     vout0x01234567 = vmin_u8(vout0x01234567, voutput_max);
137 
138     if (nc >= 8) {
139       vst1_u8(c0 + 0, vout0x01234567);
140 
141       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
142 
143       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
144 
145       nc -= 8;
146     } else {
147       if (nc & 4) {
148         vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
149         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
150       }
151       if (nc & 2) {
152         vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
153         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
154       }
155       if (nc & 1) {
156         vst1_lane_u8(c0, vout0x01234567, 0);
157       }
158 
159       nc = 0;
160     }
161   } while (nc != 0);
162 }
163