1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-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_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__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_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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 do {
44 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
45 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
46 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
47 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
48
49 size_t k = kc;
50 while (k >= 16 * sizeof(int8_t)) {
51 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
52 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
53
54 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
62
63 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
64 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
65 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
66 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
67 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
68 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
69 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
70 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
71 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
72 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
73 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
74 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
75 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
76 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
77 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
78 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
79 va0x0 = vext_s8(va0x0, va0x0, 4);
80 va0x1 = vext_s8(va0x1, va0x1, 4);
81 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
82 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
83 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
84 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
85 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
86 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
87 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
88 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
89 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
90 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
91 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
92 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
93 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
94 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
95 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
96 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
97
98 k -= 16 * sizeof(int8_t);
99 }
100 if (k != 0) {
101 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
102
103 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111
112 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
113 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
114 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
115 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
116 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
117 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
118 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
119 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
120 va0x0 = vext_s8(va0x0, va0x0, 4);
121 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
122 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
123 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
124 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
125 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
126 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
127 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
128 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
129
130 }
131
132 #if XNN_ARCH_ARM64
133 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
134 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
135 #else
136 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
137 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
138 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
139 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
140 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
141 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
142 #endif
143
144 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
145 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
146
147 const float32x4_t vscale0123 = vld1q_f32(w); w = (const float*) w + 4;
148 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
149 const float32x4_t vscale4567 = vld1q_f32(w); w = (const float*) w + 4;
150 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
151
152 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
153 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
154 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
155
156 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
157 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
158 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
159
160 #if XNN_ARCH_ARM64
161 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
162
163
164 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
165 #else
166 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
167
168
169 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
170 #endif
171
172 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
173 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
174
175 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
176 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
177
178 if (nc >= 8) {
179 vst1_s8(c0 + 0, vout0x01234567);
180
181 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
184
185 nc -= 8;
186 } else {
187 // Final case where not all of the 8 columns fit in the destination.
188 if (nc & 4) {
189 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
190 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
191 }
192 if (nc & 2) {
193 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
194 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
195 }
196 if (nc & 1) {
197 vst1_lane_s8(c0, vout0x01234567, 0);
198 }
199
200 nc = 0;
201 }
202 } while (nc != 0);
203 }
204