1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_1x2c4__armsimd32(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_igemm_minmax_fp32_ukernel_1x2c4__armsimd32(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const uint8_t**restrict a,
26 const void*restrict w,
27 uint8_t*restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const uint8_t* zero,
32 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 uint8_t* c0 = c;
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 vacc0x1 = ((const int32_t*) w)[1];
53 w = (const void*) ((const int32_t*) w + 2);
54
55 size_t p = ks;
56 do {
57 const uint8_t* restrict a0 = a[0];
58 assert(a0 != NULL);
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 a += 1;
63
64 size_t k = kc;
65 do {
66 const int8x4_t va0 = (int8x4_t) unaligned_load_s32(a0); a0 += 4;
67
68 const int16x2_t va0c02 = __uxtb16(va0);
69 const int16x2_t va0c13 = __uxtb16(__ror(va0, 8));
70
71 const int8x4_t vb0 = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
72 const int16x2_t vb0c02 = __uxtab16(vb_minus_zero_point, vb0);
73
74 vacc0x0 = __smlad(va0c02, vb0c02, vacc0x0);
75
76 const int16x2_t vb0c13 = __uxtab16(vb_minus_zero_point, __ror(vb0, 8));
77 vacc0x0 = __smlad(va0c13, vb0c13, vacc0x0);
78 const int8x4_t vb1 = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
79 const int16x2_t vb1c02 = __uxtab16(vb_minus_zero_point, vb1);
80
81 vacc0x1 = __smlad(va0c02, vb1c02, vacc0x1);
82
83 const int16x2_t vb1c13 = __uxtab16(vb_minus_zero_point, __ror(vb1, 8));
84 vacc0x1 = __smlad(va0c13, vb1c13, vacc0x1);
85
86 k -= 4 * sizeof(uint8_t);
87 } while (k != 0);
88 p -= 1 * sizeof(void*);
89 } while (p != 0);
90
91 float vfpacc0x0 = (float) vacc0x0;
92 float vfpacc0x1 = (float) vacc0x1;
93
94 vfpacc0x0 *= vscale;
95 vfpacc0x1 *= vscale;
96
97 vfpacc0x0 += vmagic_bias;
98 vfpacc0x1 += vmagic_bias;
99
100 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
101 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
102
103 const int32_t vmagic_bias_less_zero_point = params->fp32_armsimd32.magic_bias_less_zero_point;
104 vout0x0 = __qsub(vout0x0, vmagic_bias_less_zero_point);
105 vout0x1 = __qsub(vout0x1, vmagic_bias_less_zero_point);
106
107 vout0x0 = __usat(vout0x0, 8);
108 vout0x1 = __usat(vout0x1, 8);
109
110 const uint32_t vout0 = (uint32_t) (uint8_t) vout0x0 | ((uint32_t) vout0x1 << 8);
111
112 uint32_t vout = vout0;
113
114 const int8x4_t voutput_min = (int8x4_t) params->fp32_armsimd32.output_min;
115 __usub8((int8x4_t) vout, voutput_min);
116 vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min);
117
118 const int8x4_t voutput_max = (int8x4_t) params->fp32_armsimd32.output_max;
119 __usub8((int8x4_t) vout, voutput_max);
120 vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout);
121
122 if XNN_LIKELY(nc >= 2) {
123 unaligned_store_u16(c0, (uint16_t) vout);
124
125 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
126
127 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
128 nc -= 2;
129 } else {
130 *c0 = (uint8_t) vout;
131
132 nc = 0;
133 }
134 } while (nc != 0);
135 }
136