1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neon-mull-padal.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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal(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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63
64 size_t p = ks;
65 do {
66 const int8_t* restrict a0 = a[0];
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
69 }
70 a += 1;
71
72 size_t k = kc;
73
74 // Handle 8 bytes at a time using MUL.
75 while (k > 0) {
76 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
77
78 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
80 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
81 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
83 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
84 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
86 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
87 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
89 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
90 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
92 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
93 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
95 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
96 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
98 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
99 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
101 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
102 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
104 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
105 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
107 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
108 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
110 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
111 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
113 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
114 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
116 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
117 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
119 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
120 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
122 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
123 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
125 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
126
127 k -= 8 * sizeof(int8_t);
128 }
129
130 p -= 1 * sizeof(void*);
131 } while (p != 0);
132
133 #if XNN_ARCH_ARM64
134 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
135 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
136 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
137 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
138 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
139 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
140 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
141 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
142 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
143 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
144 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
145 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
146 #else
147 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
148 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
149 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
150 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
151 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
152 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
153 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
154 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
155 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
156 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
157 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
158 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
159 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
160 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
161 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
162 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
163 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
164 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
165 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
166 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
167 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
168 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
169 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
170 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
171 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
172 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
173 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
174 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
175 #endif
176
177 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier);
178 vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
179 vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
180 vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
181 vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
182
183 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
184 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
185 vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
186 vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
187 vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
188 vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
189
190 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
191 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
192 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
193 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
194
195 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
196 #if XNN_ARCH_ARM64
197 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
198 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
199 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
200 #else
201 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
202 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
203
204 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
205 #endif
206 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
207 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
208
209 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
210
211 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
212
213 if (nc >= 16) {
214 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
215
216 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
217
218 a = (const int8_t**restrict) ((uintptr_t) a - ks);
219
220 nc -= 16;
221 } else {
222 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
223 if (nc & 8) {
224 vst1_s8(c0, vout0x01234567); c0 += 8;
225 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
226 }
227 if (nc & 4) {
228 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
229 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
230 }
231 if (nc & 2) {
232 vst1_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
233 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
234 }
235 if (nc & 1) {
236 vst1_lane_s8(c0, vout0x01234567, 0);
237 }
238
239 nc = 0;
240 }
241 } while (nc != 0);
242 }
243