xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8-minmax-fp32-neonv8-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/neon-mlal-lane.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/common.h>
15 #include <xnnpack/igemm.h>
16 #include <xnnpack/intrinsics-polyfill.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 1);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   int8_t* c0 = c;
45 
46   do {
47     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49 
50     size_t p = ks;
51     do {
52       const int8_t* restrict a0 = a[0];
53       if XNN_UNPREDICTABLE(a0 != zero) {
54         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55       }
56       a += 1;
57 
58       size_t k = kc;
59       while (k >= 8 * sizeof(int8_t)) {
60         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
61         const int16x8_t vxa0 = vmovl_s8(va0);
62 
63         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
64         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
65 
66         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
67         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
68         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
69         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
70 
71         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
72         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
73         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74         const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
75 
76         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
77         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
78         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
79         const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
80 
81         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
82         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
83 
84 
85         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
86         const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
87 
88         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
89         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
90         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
91         const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
92 
93         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
94         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
95         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
96         const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
97 
98         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
99         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
100         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
101         const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
102 
103         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
104         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
105 
106         k -= 8 * sizeof(int8_t);
107       }
108       if XNN_UNLIKELY(k != 0) {
109         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
110         const int16x8_t vxa0 = vmovl_s8(va0);
111 
112         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
113         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
114 
115         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
116         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
117 
118         if (k >= 2 * sizeof(int8_t)) {
119           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
120           const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
121 
122           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
123           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
124 
125           if (k > 2 * sizeof(int8_t)) {
126             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
127             const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
128 
129             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
130             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
131 
132             if (k >= 4 * sizeof(int8_t)) {
133               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
134               const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
135 
136               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
137               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
138 
139               if (k > 4 * sizeof(int8_t)) {
140                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
141                 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
142 
143                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
144                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
145 
146                 if (k >= 6 * sizeof(int8_t)) {
147                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
148                   const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
149 
150                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
151                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
152 
153                   if (k > 6 * sizeof(int8_t)) {
154                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
155                     const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
156 
157                     vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
158                     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
159                   }
160                 }
161               }
162             }
163           }
164         }
165       }
166       p -= 1 * sizeof(void*);
167     } while (p != 0);
168 
169     // Post-accumulation work
170     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
171     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
172 
173     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
174     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
175     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
176     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
177 
178     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
179     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
180 
181     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
182 #if XNN_ARCH_ARM64
183     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
184 
185     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
186 
187     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
188 #else
189     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
190 
191     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
192 
193     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
194 #endif
195 
196     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
197     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
198 
199     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
200     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
201 
202     if (nc >= 8) {
203       vst1_s8(c0 + 0, vout0x01234567);
204 
205       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
206 
207       a = (const int8_t**restrict) ((uintptr_t) a - ks);
208 
209       nc -= 8;
210     } else {
211       if (nc & 4) {
212         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
213         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
214       }
215       if (nc & 2) {
216         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
217         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
218       }
219       if (nc & 1) {
220         vst1_lane_s8(c0, vout0x01234567, 0);
221       }
222 
223       nc = 0;
224     }
225   } while (nc != 0);
226 }
227