1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-armsimd32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_acle.h>
13
14 #include <xnnpack/intrinsics-polyfill.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/gemm.h>
17 #include <xnnpack/unaligned.h>
18
19
xnn_qu8_gemm_minmax_fp32_ukernel_2x1c4__armsimd32(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)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_2x1c4__armsimd32(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36
37 kc = round_up_po2(kc, 4 * sizeof(int8_t));
38 const uint8_t* a0 = a;
39 uint8_t* c0 = c;
40 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
41 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr != 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46
47 const int16x2_t vb_minus_zero_point = (int16x2_t) params->fp32_armsimd32.minus_kernel_zero_point;
48 const float vscale = params->fp32_armsimd32.scale;
49 const float vmagic_bias = params->fp32_armsimd32.magic_bias;
50 do {
51 int32_t vacc0x0 = ((const int32_t*) w)[0];
52 int32_t vacc1x0 = vacc0x0;
53 w = (const void*) ((const int32_t*) w + 1);
54
55 size_t k = kc;
56 do {
57 const int8x4_t va0 = (int8x4_t) unaligned_load_s32(a0); a0 += 4;
58 const int8x4_t va1 = (int8x4_t) unaligned_load_s32(a1); a1 += 4;
59
60 const int16x2_t va0c02 = __uxtb16(va0);
61 const int16x2_t va0c13 = __uxtb16(__ror(va0, 8));
62 const int16x2_t va1c02 = __uxtb16(va1);
63 const int16x2_t va1c13 = __uxtb16(__ror(va1, 8));
64
65 const int8x4_t vb0 = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
66 const int16x2_t vb0c02 = __uxtab16(vb_minus_zero_point, vb0);
67
68 vacc0x0 = __smlad(va0c02, vb0c02, vacc0x0);
69 vacc1x0 = __smlad(va1c02, vb0c02, vacc1x0);
70
71 const int16x2_t vb0c13 = __uxtab16(vb_minus_zero_point, __ror(vb0, 8));
72 vacc0x0 = __smlad(va0c13, vb0c13, vacc0x0);
73 vacc1x0 = __smlad(va1c13, vb0c13, vacc1x0);
74
75 k -= 4 * sizeof(uint8_t);
76 } while (k != 0);
77
78 float vfpacc0x0 = (float) vacc0x0;
79 float vfpacc1x0 = (float) vacc1x0;
80
81 vfpacc0x0 *= vscale;
82 vfpacc1x0 *= vscale;
83
84 vfpacc0x0 += vmagic_bias;
85 vfpacc1x0 += vmagic_bias;
86
87 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
88 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
89
90 const int32_t vmagic_bias_less_zero_point = params->fp32_armsimd32.magic_bias_less_zero_point;
91 vout0x0 = __qsub(vout0x0, vmagic_bias_less_zero_point);
92 vout1x0 = __qsub(vout1x0, vmagic_bias_less_zero_point);
93
94 vout0x0 = __usat(vout0x0, 8);
95 vout1x0 = __usat(vout1x0, 8);
96
97 const uint32_t vout0 = (uint32_t) vout0x0;
98 const uint32_t vout1 = (uint32_t) vout1x0;
99
100 uint32_t vout = (uint32_t) (uint16_t) vout0 | (vout1 << 16);
101
102 const int8x4_t voutput_min = (int8x4_t) params->fp32_armsimd32.output_min;
103 __usub8((int8x4_t) vout, voutput_min);
104 vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min);
105
106 const int8x4_t voutput_max = (int8x4_t) params->fp32_armsimd32.output_max;
107 __usub8((int8x4_t) vout, voutput_max);
108 vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout);
109
110 *c0 = (uint8_t) vout;
111 vout >>= 16;
112 *c1 = (uint8_t) vout;
113
114 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
115 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
116
117 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
118 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
119
120 nc -= 1;
121 } while (nc != 0);
122 }
123