1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_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(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46
47 do {
48 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
49 w = (const void*) ((const int32_t*) w + 4);
50
51 size_t p = ks;
52 do {
53 const uint8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
61 const __m128i vzero = _mm_setzero_si128();
62 while (k >= 8 * sizeof(uint8_t)) {
63 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
64 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
65 a0 += 8;
66
67 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
68 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
69
70 vacc0x0123 = _mm_add_epi32(vacc0x0123,
71 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
72 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
73 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
74
75 vacc0x0123 = _mm_add_epi32(vacc0x0123,
76 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
77 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
78 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
82 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
83 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
84
85 vacc0x0123 = _mm_add_epi32(vacc0x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
87
88 w = (const void*) ((const uint8_t*) w + 32);
89 k -= 8 * sizeof(uint8_t);
90 }
91 if (k != 0) {
92 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
93 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
94 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
95
96 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
97 w = (const void*) ((const uint8_t*) w + 8);
98 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
99
100 vacc0x0123 = _mm_add_epi32(vacc0x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
102
103 if (k > 2 * sizeof(uint8_t)) {
104 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
105 w = (const void*) ((const uint8_t*) w + 8);
106 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
107
108 vacc0x0123 = _mm_add_epi32(vacc0x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
110
111 if (k > 4 * sizeof(uint8_t)) {
112 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
113 w = (const void*) ((const uint8_t*) w + 8);
114 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
115
116 vacc0x0123 = _mm_add_epi32(vacc0x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
118 }
119 }
120 }
121 p -= 1 * sizeof(void*);
122 } while (p != 0);
123
124 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
125
126 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
127 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
128
129 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
130 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
131
132 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
133
134 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
135 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
136
137 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
138
139 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
140
141 if (nc >= 4) {
142 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
143 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
144
145 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
146
147 nc -= 4;
148 } else {
149 if (nc & 2) {
150 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
151 c0 += 2;
152 vout = _mm_srli_epi32(vout, 16);
153 }
154 if (nc & 1) {
155 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
156 }
157
158 nc = 0;
159 }
160 } while (nc != 0);
161 }
162