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_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_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(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
50 while (k >= 8 * sizeof(uint8_t)) {
51 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
52 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
53 a0 += 8;
54
55 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
56 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
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 uint8_t*) w + 8));
61 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
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 uint8_t*) w + 16));
66 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
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 uint8_t*) w + 24));
71 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
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 uint8_t*) w + 32);
77 k -= 8 * sizeof(uint8_t);
78 }
79 if (k != 0) {
80 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
81 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
82 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
83
84 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
85 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
86 w = (const void*) ((const uint8_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(uint8_t)) {
92 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
93 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
94 w = (const void*) ((const uint8_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(uint8_t)) {
100 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
101 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
102 w = (const void*) ((const uint8_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 vscale = _mm_load_ps(params->fp32_sse2.scale);
113 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
114
115 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
116 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
117
118 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
119
120 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
121 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
122
123 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
124
125 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
126
127 if (nc >= 4) {
128 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
129
130 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
131
132 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
133
134 nc -= 4;
135 } else {
136 if (nc & 2) {
137 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
138 c0 += 2;
139 vout = _mm_srli_epi32(vout, 16);
140 }
141 if (nc & 1) {
142 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
143 }
144
145 nc = 0;
146 }
147 } while (nc != 0);
148 }
149