xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8-minmax-fp32-neon-mlal-lane-prfm.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 
17 
xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm(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)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm(
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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   int8_t* c0 = c;
44 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
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       while (k >= 8 * sizeof(int8_t)) {
59         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
60         const int16x8_t vxa0 = vmovl_s8(va0);
61 
62         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
63         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
64 
65         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
66         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
67         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
68         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
69 
70         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
71         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
72         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
73         const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
74 
75         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
76         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
77         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
78         const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
79 
80         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
81         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
82 
83         __builtin_prefetch((const int8_t*) w + 448);
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     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
179     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
180     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
181 
182     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
183     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
184     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
185 
186 #if XNN_ARCH_ARM64
187     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
188 
189 
190     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
191 #else
192     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
193 
194 
195     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
196 #endif
197 
198     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
199     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
200 
201     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
202     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
203 
204     if (nc >= 8) {
205       vst1_s8(c0 + 0, vout0x01234567);
206 
207       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
208 
209       a = (const int8_t**restrict) ((uintptr_t) a - ks);
210 
211       nc -= 8;
212     } else {
213       if (nc & 4) {
214         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
215         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
216       }
217       if (nc & 2) {
218         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
219         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
220       }
221       if (nc & 1) {
222         vst1_lane_s8(c0, vout0x01234567, 0);
223       }
224 
225       nc = 0;
226     }
227   } while (nc != 0);
228 }
229