xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x8-minmax-fp32-neon-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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/gemm.h>
16 
17 
xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane(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_1x8__neon_mlal_lane(
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   const int8_t* a0 = a;
40   int8_t* c0 = c;
41 
42   do {
43     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
44     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
45 
46     size_t k = kc;
47     while (k >= 8 * sizeof(int8_t)) {
48       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
49       const int16x8_t vxa0 = vmovl_s8(va0);
50 
51       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
52       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
53 
54       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
55       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
56       const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
57       const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
58 
59       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
60       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
61       const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
62       const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
63 
64       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
65       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
66       const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
67       const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
68 
69       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
70       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
71 
72 
73       const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74       const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
75 
76       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
77       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
78       const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
79       const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
80 
81       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
82       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
83       const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
84       const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
85 
86       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
87       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
88       const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
89       const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
90 
91       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
92       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
93 
94       k -= 8 * sizeof(int8_t);
95     }
96     if XNN_UNLIKELY(k != 0) {
97       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
98       const int16x8_t vxa0 = vmovl_s8(va0);
99 
100       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
101       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
102 
103       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
104       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
105 
106       if (k >= 2 * sizeof(int8_t)) {
107         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
108         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
109 
110         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
111         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
112 
113         if (k > 2 * sizeof(int8_t)) {
114           const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
115           const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
116 
117           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
118           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
119 
120           if (k >= 4 * sizeof(int8_t)) {
121             const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
122             const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
123 
124             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
125             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
126 
127             if (k > 4 * sizeof(int8_t)) {
128               const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
129               const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
130 
131               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
132               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
133 
134               if (k >= 6 * sizeof(int8_t)) {
135                 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
136                 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
137 
138                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
139                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
140 
141                 if (k > 6 * sizeof(int8_t)) {
142                   const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
143                   const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
144 
145                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
146                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
147                 }
148               }
149             }
150           }
151         }
152       }
153     }
154 
155     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
156     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
157 
158     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
159     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
160     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
161     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
162 
163     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
164     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
165     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
166 
167     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
168     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
169     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
170 
171 #if XNN_ARCH_ARM64
172     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
173 
174 
175     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
176 #else
177     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
178 
179 
180     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
181 #endif
182 
183     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
184     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
185 
186     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
187     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
188 
189     if (nc >= 8) {
190       vst1_s8(c0 + 0, vout0x01234567);
191 
192       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
193 
194       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
195 
196       nc -= 8;
197     } else {
198       if (nc & 4) {
199         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
200         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
201       }
202       if (nc & 2) {
203         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
204         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
205       }
206       if (nc & 1) {
207         vst1_lane_s8(c0, vout0x01234567, 0);
208       }
209 
210       nc = 0;
211     }
212   } while (nc != 0);
213 }
214