xref: /aosp_15_r20/external/XNNPACK/src/qu8-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_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_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(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const uint8_t* a0 = a;
40   uint8_t* c0 = c;
41 
42   const uint8x8_t vb_zero_point = vld1_dup_u8(&params->fp32_neon.kernel_zero_point[0]);
43   do {
44     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
45     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
46 
47     size_t k = kc;
48     while (k >= 8 * sizeof(uint8_t)) {
49       const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
50       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
51 
52       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
53       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
54 
55       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
56       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
57       const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
58       const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
59 
60       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
61       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
62       const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
63       const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
64 
65       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
66       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
67       const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
68       const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
69 
70       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
71       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
72 
73 
74       const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
75       const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
76 
77       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
78       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
79       const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
80       const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
81 
82       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
83       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
84       const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
85       const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
86 
87       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
88       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
89       const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
90       const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
91 
92       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
93       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
94 
95       k -= 8 * sizeof(uint8_t);
96     }
97     if XNN_UNLIKELY(k != 0) {
98       const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
99       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
100 
101       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
102       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
103 
104       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
105       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
106 
107       if (k >= 2 * sizeof(uint8_t)) {
108         const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
109         const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
110 
111         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
112         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
113 
114         if (k > 2 * sizeof(uint8_t)) {
115           const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
116           const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
117 
118           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
119           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
120 
121           if (k >= 4 * sizeof(uint8_t)) {
122             const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
123             const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
124 
125             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
126             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
127 
128             if (k > 4 * sizeof(uint8_t)) {
129               const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
130               const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
131 
132               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
133               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
134 
135               if (k >= 6 * sizeof(uint8_t)) {
136                 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
137                 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
138 
139                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
140                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
141 
142                 if (k > 6 * sizeof(uint8_t)) {
143                   const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
144                   const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
145 
146                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
147                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
148                 }
149               }
150             }
151           }
152         }
153       }
154     }
155 
156     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
157     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
158 
159     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
160     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
161     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
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     uint8x8_t vout0x01234567 = vqmovun_s16(vacc0x01234567);
176 #else
177     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
178 
179 
180     uint8x8_t vout0x01234567 = vqmovun_s16(vacc0x01234567);
181 #endif
182 
183     const uint8x8_t voutput_min = vld1_dup_u8(&params->fp32_neon.output_min);
184     vout0x01234567 = vmax_u8(vout0x01234567, voutput_min);
185 
186     const uint8x8_t voutput_max = vld1_dup_u8(&params->fp32_neon.output_max);
187     vout0x01234567 = vmin_u8(vout0x01234567, voutput_max);
188 
189     if (nc >= 8) {
190       vst1_u8(c0 + 0, vout0x01234567);
191 
192       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
193 
194       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
195 
196       nc -= 8;
197     } else {
198       if (nc & 4) {
199         vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
200         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
201       }
202       if (nc & 2) {
203         vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
204         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
205       }
206       if (nc & 1) {
207         vst1_lane_u8(c0, vout0x01234567, 0);
208       }
209 
210       nc = 0;
211     }
212   } while (nc != 0);
213 }
214