• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c8-neon-mull.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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_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)])19 void xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 8 * sizeof(int8_t));
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43 
44   do {
45     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
46     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
47     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 
54     size_t k = kc;
55     // 2x partial unrolled loop to load 16 bytes at a time using MLA.
56     while (k >= 16 * sizeof(int8_t)) {
57       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
58       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
59 
60       const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
61       const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
62       const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
63       const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
64       const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
65       const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
66       const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
67       const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
68 
69       const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
70       int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
71       vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
72       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
73       const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
74       int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
75       vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
76       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
77       const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
78       int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
79       vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
80       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
81       const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
82       int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
83       vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
84       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
85       const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
86       int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
87       vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
88       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
89       const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
90       int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
91       vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
92       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
93       const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
94       int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
95       vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
96       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
97       const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
98       int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
99       vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
100       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
101 
102       k -= 16 * sizeof(int8_t);
103     }
104 
105     // Handle 8 bytes at a time using MUL.
106     if (k != 0) {
107       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
108 
109       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
111       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
112       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
114       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
115       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
117       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
118       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
120       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
121       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
123       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
124       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
126       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
127       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
129       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
130       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
132       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
133 
134       k -= 8 * sizeof(int8_t);
135     }
136 
137 #if XNN_ARCH_ARM64
138     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
139     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
140     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
141     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
142 
143     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
144     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
145 #else
146     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
147     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
148     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
149     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
150     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
151     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
152     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
153     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
154     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
155     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
156     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
157     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
158     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
159     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
160 #endif
161 
162     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
163     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
164 
165     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
166     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
167     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
168 
169     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
170     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
171 
172     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
173 #if XNN_ARCH_ARM64
174     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
175 
176     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
177 
178     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
179 #else
180     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
181 
182     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
183 
184     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
185 #endif
186 
187     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
188     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
189 
190     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
191     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
192 
193     if (nc >= 8) {
194       vst1_s8(c0 + 0, vout0x01234567);
195 
196       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
197 
198       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
199 
200       nc -= 8;
201     } else {
202       // Final case where not all of the 8 columns fit in the destination.
203       if (nc & 4) {
204         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
205         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
206       }
207       if (nc & 2) {
208         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
209         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
210       }
211       if (nc & 1) {
212         vst1_lane_s8(c0, vout0x01234567, 0);
213       }
214 
215       nc = 0;
216     }
217   } while (nc != 0);
218 }
219