1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neon-mull.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
17
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_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(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
45 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
46 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
47 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52
53 size_t k = kc;
54 // 2x partial unrolled loop to load 16 bytes at a time using MLA.
55 while (k >= 16 * sizeof(int8_t)) {
56 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
57 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
58
59 const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
60 const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
61 const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
62 const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
63 const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
64 const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
65 const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
66 const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
67
68 const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
69 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
70 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
71 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
72 const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
73 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
74 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
75 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
76 const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
77 int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
78 vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
79 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
80 const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
81 int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
82 vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
83 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
84 const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
85 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
86 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
87 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
88 const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
89 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
90 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
91 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
92 const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
93 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
94 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
95 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
96 const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
97 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
98 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
99 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
100
101 k -= 16 * sizeof(int8_t);
102 }
103
104 // Handle 8 bytes at a time using MUL.
105 if (k != 0) {
106 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
107
108 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
110 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
111 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
113 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
114 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
116 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
117 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
119 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
120 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
122 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
123 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
125 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
126 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
128 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
129 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
131 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
132
133 k -= 8 * sizeof(int8_t);
134 }
135
136 #if XNN_ARCH_ARM64
137 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
138 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
139 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
140 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
141
142 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
143 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
144 #else
145 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
146 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
147 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
148 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
149 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
150 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
151 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
152 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
153 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
154 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
155 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
157 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
158 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
159 #endif
160
161 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
162 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
163
164 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
165 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
166 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
167 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
168
169 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
170 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
171 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
172
173 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
174 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
175 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
176
177 #if XNN_ARCH_ARM64
178 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
179
180
181 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
182 #else
183 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
184
185
186 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
187 #endif
188
189 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
190 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
191
192 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
193 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
194
195 if (nc >= 8) {
196 vst1_s8(c0 + 0, vout0x01234567);
197
198 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
199
200 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
201
202 nc -= 8;
203 } else {
204 // Final case where not all of the 8 columns fit in the destination.
205 if (nc & 4) {
206 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
207 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
208 }
209 if (nc & 2) {
210 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
211 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
212 }
213 if (nc & 1) {
214 vst1_lane_s8(c0, vout0x01234567, 0);
215 }
216
217 nc = 0;
218 }
219 } while (nc != 0);
220 }
221