xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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/math.h>
16 
17 
xnn_qu8_gemm_minmax_rndnu_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)])18 void xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
40   const uint8_t* a0 = a;
41   uint8_t* c0 = c;
42 
43   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
44 
45   // Loop over groups of 16 columns.
46   do {
47     // Initialize accumulators with bias. 16 bias values are loaded from the
48     // weight matrix, at the start of the group of 16 columns.
49     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
50     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
51     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
52     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53     uint32x2_t vnacc0 = vmov_n_u32(0);
54 
55     // Inner accumulation loop along the 16 columns.
56     size_t k = kc;
57     // 2x partial unrolled loop to load 8 bytes at a time.
58     while (k >= 8 * sizeof(uint8_t)) {
59       // Load a 1x8 block of activations.
60       const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
61 
62       // Load a 8x16 block of weights.
63       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
64       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
65       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
66       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
67       const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
68       const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
69       const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
70       const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
71 
72       // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
73       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
74       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
75       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
76       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
77       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
78       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
79       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
80       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
81       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, 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 4x16 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       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95 
96       // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
97       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
98       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
99       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
100       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
101       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
102     }
103 
104     // Subtract zero point from accumulators.
105     vnacc0 = vpadd_u32(vnacc0, vnacc0);
106     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
107     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
108     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
109     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
110     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
111 
112     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
113     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
114     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
115 
116     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
117     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
118     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
119     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
120 
121     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
122     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
123     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
124     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
125 
126     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
127     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
128     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
129     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
130 
131     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
132 #if XNN_ARCH_ARM64
133     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
134     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
135 
136     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
137 #else
138     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
139     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
140 
141     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
142 #endif
143     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
144     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
145 
146     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
147 
148     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
149 
150     if (nc >= 16) {
151       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
152 
153       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
154 
155       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
156 
157       nc -= 16;
158     } else {
159       uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
160       if (nc & 8) {
161         vst1_u8(c0, vout0x01234567); c0 += 8;
162         vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
163       }
164       if (nc & 4) {
165         vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
166         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
167       }
168       if (nc & 2) {
169         vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
170         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
171       }
172       if (nc & 1) {
173         vst1_lane_u8(c0, vout0x01234567, 0);
174       }
175 
176       nc = 0;
177     }
178   } while (nc != 0);
179 }
180