xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x1c4-minmax-fp32-armsimd32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_qs8_igemm_minmax_fp32_ukernel_2x1c4__armsimd32(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_igemm_minmax_fp32_ukernel_2x1c4__armsimd32(
21     size_t mr,
22     size_t nc,
23     size_t kc,
24     size_t ks,
25     const int8_t**restrict a,
26     const void*restrict w,
27     int8_t*restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     size_t a_offset,
31     const int8_t* zero,
32     const union xnn_qs8_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   int8_t* c0 = c;
46   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr != 2) {
48     c1 = c0;
49   }
50 
51   const float vscale = params->fp32_armsimd32.scale;
52   const float vmagic_bias = params->fp32_armsimd32.magic_bias;
53   do {
54     int32_t vacc0x0 = ((const int32_t*) w)[0];
55     int32_t vacc1x0 = vacc0x0;
56     w = (const void*) ((const int32_t*) w + 1);
57 
58     size_t p = ks;
59     do {
60       const int8_t* restrict a0 = a[0];
61       assert(a0 != NULL);
62       if XNN_UNPREDICTABLE(a0 != zero) {
63         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
64       }
65       const int8_t* restrict a1 = a[1];
66       assert(a1 != NULL);
67       if XNN_UNPREDICTABLE(a1 != zero) {
68         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69       }
70       a += 2;
71 
72       size_t k = kc;
73       do {
74         const int8x4_t va0 = (int8x4_t) unaligned_load_s32(a0); a0 += 4;
75         const int8x4_t va1 = (int8x4_t) unaligned_load_s32(a1); a1 += 4;
76 
77         const int16x2_t va0c02 = __sxtb16(va0);
78         const int16x2_t va0c13 = __sxtb16(__ror(va0, 8));
79         const int16x2_t va1c02 = __sxtb16(va1);
80         const int16x2_t va1c13 = __sxtb16(__ror(va1, 8));
81 
82         const int8x4_t vb0 = *((const int8x4_t*) w); w = (const int8_t*) w + 4;
83         const int16x2_t vb0c02 = __sxtb16(vb0);
84 
85         vacc0x0 = __smlad(va0c02, vb0c02, vacc0x0);
86         vacc1x0 = __smlad(va1c02, vb0c02, vacc1x0);
87 
88         const int16x2_t vb0c13 = __sxtb16(__ror(vb0, 8));
89         vacc0x0 = __smlad(va0c13, vb0c13, vacc0x0);
90         vacc1x0 = __smlad(va1c13, vb0c13, vacc1x0);
91 
92         k -= 4 * sizeof(int8_t);
93       } while (k != 0);
94       p -= 2 * sizeof(void*);
95     } while (p != 0);
96 
97     float vfpacc0x0 = (float) vacc0x0;
98     float vfpacc1x0 = (float) vacc1x0;
99 
100     vfpacc0x0 *= vscale;
101     vfpacc1x0 *= vscale;
102 
103     vfpacc0x0 += vmagic_bias;
104     vfpacc1x0 += vmagic_bias;
105 
106     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
107     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
108 
109     const int32_t vmagic_bias_less_zero_point = params->fp32_armsimd32.magic_bias_less_zero_point;
110     vout0x0 = __qsub(vout0x0, vmagic_bias_less_zero_point);
111     vout1x0 = __qsub(vout1x0, vmagic_bias_less_zero_point);
112 
113     vout0x0 = __ssat(vout0x0, 8);
114     vout1x0 = __ssat(vout1x0, 8);
115 
116     const uint32_t vout0 = (uint32_t) vout0x0;
117     const uint32_t vout1 = (uint32_t) vout1x0;
118 
119     uint32_t vout = (uint32_t) (uint16_t) vout1 | (vout0 << 16);
120 
121     const int8x4_t voutput_min = (int8x4_t) params->fp32_armsimd32.output_min;
122     __ssub8((int8x4_t) vout, voutput_min);
123     vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min);
124 
125     const int8x4_t voutput_max = (int8x4_t) params->fp32_armsimd32.output_max;
126     __ssub8((int8x4_t) vout, voutput_max);
127     vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout);
128 
129     *c1 = (int8_t) vout;
130     vout >>= 16;
131     *c0 = (int8_t) vout;
132 
133     c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
134     c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
135 
136     a = (const int8_t**restrict) ((uintptr_t) a - ks);
137     nc -= 1;
138   } while (nc != 0);
139 }
140