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 <smmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43
44 do {
45 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
46 w = (const void*) ((const int32_t*) w + 4);
47
48 size_t k = kc;
49 while (k >= 8 * sizeof(int8_t)) {
50 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
51 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
52 a0 += 8;
53
54 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
55 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
56 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
57
58 vacc0x0123 = _mm_add_epi32(vacc0x0123,
59 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
60
61 vacc0x0123 = _mm_add_epi32(vacc0x0123,
62 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
63 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
64 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
65 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
66
67 vacc0x0123 = _mm_add_epi32(vacc0x0123,
68 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
69
70 vacc0x0123 = _mm_add_epi32(vacc0x0123,
71 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
72
73 w = (const void*) ((const int8_t*) w + 32);
74 k -= 8 * sizeof(int8_t);
75 }
76 if (k != 0) {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
79 a0 = (const int8_t*) ((uintptr_t) a0 + k);
80
81 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
82 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
83 w = (const void*) ((const int8_t*) w + 8);
84
85 vacc0x0123 = _mm_add_epi32(vacc0x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
87
88 if (k > 2 * sizeof(int8_t)) {
89 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
91 w = (const void*) ((const int8_t*) w + 8);
92
93 vacc0x0123 = _mm_add_epi32(vacc0x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
95
96 if (k > 4 * sizeof(int8_t)) {
97 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
98 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
99 w = (const void*) ((const int8_t*) w + 8);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
103 }
104 }
105 }
106
107 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
108
109 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
110 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
111
112 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
113 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
114
115 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
116
117 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
118 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
119
120
121 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
122
123 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
124
125 if (nc >= 4) {
126 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
127
128 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
129
130 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
131
132 nc -= 4;
133 } else {
134 if (nc & 2) {
135 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
136 c0 += 2;
137 vout = _mm_srli_epi32(vout, 16);
138 }
139 if (nc & 1) {
140 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
141 }
142
143 nc = 0;
144 }
145 } while (nc != 0);
146 }
147