xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c2-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/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17 
18 
19 
xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64(
21     size_t mr,
22     size_t nc,
23     size_t kc,
24     const int8_t* restrict a,
25     size_t a_stride,
26     const void* restrict w,
27     int8_t* restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(int8_t) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   kc = round_up_po2(kc, 2 * sizeof(int8_t));
42   const int8_t* a0 = a;
43   int8_t* c0 = c;
44 
45   do {
46     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
47     w = (const void*) ((const int32_t*) w + 4);
48 
49     size_t k = kc;
50     while (k >= 8 * sizeof(int8_t)) {
51       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
52       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
53       a0 += 8;
54 
55       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
56       const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
57 
58       vacc0x0123 = _mm_add_epi32(vacc0x0123,
59         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
60       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
61       const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
62 
63       vacc0x0123 = _mm_add_epi32(vacc0x0123,
64         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
65       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
66       const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
67 
68       vacc0x0123 = _mm_add_epi32(vacc0x0123,
69         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
70       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
71       const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
72 
73       vacc0x0123 = _mm_add_epi32(vacc0x0123,
74         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
75 
76       w = (const void*) ((const int8_t*) w + 32);
77       k -= 8 * sizeof(int8_t);
78     }
79     if (k != 0) {
80       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
81       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
82       a0 = (const int8_t*) ((uintptr_t) a0 + k);
83 
84       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
85       const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
86       w = (const void*) ((const int8_t*) w + 8);
87 
88       vacc0x0123 = _mm_add_epi32(vacc0x0123,
89         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
90 
91       if (k > 2 * sizeof(int8_t)) {
92         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
93         const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
94         w = (const void*) ((const int8_t*) w + 8);
95 
96         vacc0x0123 = _mm_add_epi32(vacc0x0123,
97           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
98 
99         if (k > 4 * sizeof(int8_t)) {
100           const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
101           const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
102           w = (const void*) ((const int8_t*) w + 8);
103 
104           vacc0x0123 = _mm_add_epi32(vacc0x0123,
105             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
106         }
107       }
108     }
109 
110     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
111 
112     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
113     w = (const void*) ((const float*) w + 4);
114     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
115 
116     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
117     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
118 
119     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
120 
121     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
122     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
123 
124     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
125     vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min);
126 
127     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
128 
129 
130     if (nc >= 4) {
131       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
132 
133       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
134 
135       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
136 
137       nc -= 4;
138     } else {
139       if (nc & 2) {
140         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
141         c0 += 2;
142         vout = _mm_srli_epi32(vout, 16);
143       }
144       if (nc & 1) {
145         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
146       }
147 
148       nc = 0;
149     }
150   } while (nc != 0);
151 }
152