xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x8c2s4-minmax-fp32-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-shuffle.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_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__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_qs8_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   kc = round_up_po2(kc, 8 * sizeof(int8_t));
43   do {
44     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
45     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
46 
47     size_t k = kc;
48     while (k >= 16 * sizeof(int8_t)) {
49       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
50       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
51 
52       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
53       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
54       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 
61       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
62       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
63       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
64       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
65       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
66       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
67       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
68       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
69       va0x0 = vext_s8(va0x0, va0x0, 2);
70       va0x1 = vext_s8(va0x1, va0x1, 2);
71       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
72       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
73       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
74       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
75       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
76       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
77       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
78       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
79       va0x0 = vext_s8(va0x0, va0x0, 2);
80       va0x1 = vext_s8(va0x1, va0x1, 2);
81       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
82       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
83       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
84       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
85       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
86       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
87       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
88       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
89       va0x0 = vext_s8(va0x0, va0x0, 2);
90       va0x1 = vext_s8(va0x1, va0x1, 2);
91       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
92       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
93       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
94       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
95       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
96       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
97       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
98       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
99 
100       k -= 16 * sizeof(int8_t);
101     }
102     if (k != 0) {
103       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
104 
105       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113 
114       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
115       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
116       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
117       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
118       va0x0 = vext_s8(va0x0, va0x0, 2);
119       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
120       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
121       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
122       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
123       va0x0 = vext_s8(va0x0, va0x0, 2);
124       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
125       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
126       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
127       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
128       va0x0 = vext_s8(va0x0, va0x0, 2);
129       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
130       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
131       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
132       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
133 
134     }
135 
136     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
137     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
138 
139     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
140     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
141     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
142 
143     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
144     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
145     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
146 
147     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
148     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
149     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
150 
151 #if XNN_ARCH_ARM64
152     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
153 
154 
155     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
156 #else
157     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
158 
159 
160     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
161 #endif
162 
163     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
164     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
165 
166     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
167     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
168 
169     if (nc >= 8) {
170       vst1_s8(c0 + 0, vout0x01234567);
171 
172       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
173 
174       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
175 
176       nc -= 8;
177     } else {
178       // Final case where not all of the 8 columns fit in the destination.
179       if (nc & 4) {
180         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
181         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
182       }
183       if (nc & 2) {
184         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
185         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
186       }
187       if (nc & 1) {
188         vst1_lane_s8(c0, vout0x01234567, 0);
189       }
190 
191       nc = 0;
192     }
193   } while (nc != 0);
194 }
195