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_2x1c4__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_2x1c4__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 <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * 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 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
50
51 const int16x2_t vb_minus_zero_point = (int16x2_t) params->fp32_armsimd32.minus_kernel_zero_point;
52 const float vscale = params->fp32_armsimd32.scale;
53 const float vmagic_bias = params->fp32_armsimd32.magic_bias;
54 do {
55 int32_t vacc0x0 = ((const int32_t*) w)[0];
56 int32_t vacc1x0 = vacc0x0;
57 w = (const void*) ((const int32_t*) w + 1);
58
59 size_t p = ks;
60 do {
61 const uint8_t* restrict a0 = a[0];
62 assert(a0 != NULL);
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const uint8_t* restrict a1 = a[1];
67 assert(a1 != NULL);
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = kc;
74 do {
75 const int8x4_t va0 = (int8x4_t) unaligned_load_s32(a0); a0 += 4;
76 const int8x4_t va1 = (int8x4_t) unaligned_load_s32(a1); a1 += 4;
77
78 const int16x2_t va0c02 = __uxtb16(va0);
79 const int16x2_t va0c13 = __uxtb16(__ror(va0, 8));
80 const int16x2_t va1c02 = __uxtb16(va1);
81 const int16x2_t va1c13 = __uxtb16(__ror(va1, 8));
82
83 const int8x4_t vb0 = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
84 const int16x2_t vb0c02 = __uxtab16(vb_minus_zero_point, vb0);
85
86 vacc0x0 = __smlad(va0c02, vb0c02, vacc0x0);
87 vacc1x0 = __smlad(va1c02, vb0c02, vacc1x0);
88
89 const int16x2_t vb0c13 = __uxtab16(vb_minus_zero_point, __ror(vb0, 8));
90 vacc0x0 = __smlad(va0c13, vb0c13, vacc0x0);
91 vacc1x0 = __smlad(va1c13, vb0c13, vacc1x0);
92
93 k -= 4 * sizeof(uint8_t);
94 } while (k != 0);
95 p -= 2 * sizeof(void*);
96 } while (p != 0);
97
98 float vfpacc0x0 = (float) vacc0x0;
99 float vfpacc1x0 = (float) vacc1x0;
100
101 vfpacc0x0 *= vscale;
102 vfpacc1x0 *= vscale;
103
104 vfpacc0x0 += vmagic_bias;
105 vfpacc1x0 += vmagic_bias;
106
107 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
108 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
109
110 const int32_t vmagic_bias_less_zero_point = params->fp32_armsimd32.magic_bias_less_zero_point;
111 vout0x0 = __qsub(vout0x0, vmagic_bias_less_zero_point);
112 vout1x0 = __qsub(vout1x0, vmagic_bias_less_zero_point);
113
114 vout0x0 = __usat(vout0x0, 8);
115 vout1x0 = __usat(vout1x0, 8);
116
117 const uint32_t vout0 = (uint32_t) vout0x0;
118 const uint32_t vout1 = (uint32_t) vout1x0;
119
120 uint32_t vout = (uint32_t) (uint16_t) vout1 | (vout0 << 16);
121
122 const int8x4_t voutput_min = (int8x4_t) params->fp32_armsimd32.output_min;
123 __usub8((int8x4_t) vout, voutput_min);
124 vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min);
125
126 const int8x4_t voutput_max = (int8x4_t) params->fp32_armsimd32.output_max;
127 __usub8((int8x4_t) vout, voutput_max);
128 vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout);
129
130 *c1 = (uint8_t) vout;
131 vout >>= 16;
132 *c0 = (uint8_t) vout;
133
134 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
135 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
136
137 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
138 nc -= 1;
139 } while (nc != 0);
140 }
141