• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-neon-mull-dup.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_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld1r(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_1x8c4__neon_mlal_ld1r(
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   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42 
43   do {
44     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 
49     size_t k = kc;
50 
51     while (k >= 16 * sizeof(int8_t)) {
52       const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
53       const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
54       const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
55       const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
56 
57       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 
66       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
67       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
68 
69       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
70       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
72       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
73       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
74       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
76       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
77       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
78       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
80       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
81       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
82       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
84       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
85       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
86       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
87 
88       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
89       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
91       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
92       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
93       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
95       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
96       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
97       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
99       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
100       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
101       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
103       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
104 
105       k -= 16 * sizeof(int8_t);
106     }
107 
108     if (k >= 8 * sizeof(int8_t)) {
109       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
110       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
111 
112       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 
121       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
122 
123       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
124       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
125       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
126       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
127       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
128       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
129       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
130       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
131       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
132 
133       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
134       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
135       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
136       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
137       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
138       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
139       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
140       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
141 
142       k -= 8 * sizeof(int8_t);
143     }
144 
145     if XNN_UNLIKELY(k != 0) {
146       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
147 
148       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 
153       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
154       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
155       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
156       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
157       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
158       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
159       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
160       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
161       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
162     }
163 
164 #if XNN_ARCH_ARM64
165     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
166     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
167 #else
168     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
169     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
170     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
171     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
172     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
173     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
174 #endif
175 
176     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
177     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
178 
179     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
180     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
181     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
182     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
183 
184     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
185     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
186     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
187 
188     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
189     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
190     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
191 
192 #if XNN_ARCH_ARM64
193     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
194 
195 
196     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
197 #else
198     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
199 
200 
201     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
202 #endif
203 
204     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
205     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
206 
207     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
208     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
209 
210     if (nc >= 8) {
211       vst1_s8(c0 + 0, vout0x01234567);
212 
213       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
214 
215       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
216 
217       nc -= 8;
218     } else {
219       // Final case where not all of the 8 columns fit in the destination.
220       if (nc & 4) {
221         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
222         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
223       }
224       if (nc & 2) {
225         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
226         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
227       }
228       if (nc & 1) {
229         vst1_lane_s8(c0, vout0x01234567, 0);
230       }
231 
232       nc = 0;
233     }
234   } while (nc != 0);
235 }
236