xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/MRx4c8-sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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 <emmintrin.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 1);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 8);
45   int8_t* c0 = c;
46 
47   do {
48     __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
49     __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
50     __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
51     __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
52     w = (const int32_t*) w + 4;
53 
54     size_t p = ks;
55     do {
56       const int8_t* restrict a0 = a[0];
57       if XNN_UNPREDICTABLE(a0 != zero) {
58         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
59       }
60       a += 1;
61 
62       size_t k = 0;
63       while (k < kc) {
64         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
65         const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
66         a0 += 8;
67 
68         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
69         const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
70 
71         vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
72         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
73         const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
74 
75         vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
76         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
77         const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
78 
79         vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
80         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
81         const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
82 
83         vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
84 
85         w = (const void*) ((const int8_t*) w + 32);
86         k += 8 * sizeof(int8_t);
87       }
88       p -= 1 * sizeof(void*);
89     } while (p != 0);
90 
91     const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
92     const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
93 
94     __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
95 
96     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
97 
98     const __m128 vscale0123 = _mm_load_ps((const float*) w);
99     w = (const void*) ((const float*) w + 4);
100     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
101 
102     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
103     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
104 
105     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
106 
107     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
108     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
109 
110     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
111     vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min);
112 
113     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
114 
115 
116     if (nc >= 4) {
117       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
118       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
119 
120       a = (const int8_t**restrict) ((uintptr_t) a - ks);
121 
122       nc -= 4;
123     } else {
124       if (nc & 2) {
125         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
126         c0 += 2;
127         vout = _mm_srli_epi32(vout, 16);
128       }
129       if (nc & 1) {
130         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
131       }
132 
133       nc = 0;
134     }
135   } while (nc != 0);
136 }
137