• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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 
xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld2r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_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(ks != 0);
36   assert(ks % (1 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48 
49     size_t p = ks;
50     do {
51       const int8_t* restrict a0 = a[0];
52       if XNN_UNPREDICTABLE(a0 != zero) {
53         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54       }
55       a += 1;
56 
57       size_t k = kc;
58 
59 
60       while (k >= 8 * sizeof(int8_t)) {
61         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
62         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
63 
64         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 
73         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
74 
75         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
76         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
77         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
78         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
79         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
80 
81         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
82         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
83         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
84         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
85         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
86 
87         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
88         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
89         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
90         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
91         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
92 
93         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
94         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
95         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
96         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
97 
98         k -= 8 * sizeof(int8_t);
99       }
100 
101       if XNN_UNLIKELY(k != 0) {
102         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
103 
104         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 
107         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
108         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
109         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
110         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
111         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
112 
113         if (k > 2 * sizeof(int8_t)) {
114           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 
117           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
118           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
119           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
120           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
121           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
122 
123           if (k > 4 * sizeof(int8_t)) {
124             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 
127             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
128             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
129             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
130             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
131             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
132           }
133         }
134       }
135       p -= 1 * sizeof(void*);
136     } while (p != 0);
137 
138     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
139     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
140     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
141 
142     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
143     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
144 
145     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
146     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
147 
148     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
149     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
150 
151     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
152 #if XNN_ARCH_ARM64
153     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
154 
155     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
156 
157     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
158 #else
159     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
160 
161     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
162 
163     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
164 #endif
165 
166     const int8x8_t voutput_min = vld1_dup_s8(&params->rndnu_neon.output_min);
167     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
168 
169     const int8x8_t voutput_max = vld1_dup_s8(&params->rndnu_neon.output_max);
170     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
171 
172     if (nc >= 8) {
173       vst1_s8(c0 + 0, vout0x01234567);
174 
175       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
176 
177       a = (const int8_t**restrict) ((uintptr_t) a - ks);
178 
179       nc -= 8;
180     } else {
181       if (nc & 4) {
182         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
183         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
184       }
185       if (nc & 2) {
186         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
187         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
188       }
189       if (nc & 1) {
190         vst1_lane_s8(c0, vout0x01234567, 0);
191       }
192 
193       nc = 0;
194     }
195   } while (nc != 0);
196 }
197