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 <smmintrin.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_1x4c8__sse41_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_1x4c8__sse41_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, 8);
45 uint8_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 uint8_t* restrict a0 = a[0];
57 if XNN_UNPREDICTABLE(a0 != zero) {
58 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
59 }
60 a += 1;
61
62 size_t k = 0;
63 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
64 while (k < kc) {
65 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
66 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
67 a0 += 8;
68
69 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
70 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
71
72 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
73 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
74 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
75
76 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
77 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
78 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
79
80 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
81 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
82 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
83
84 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
85
86 w = (const void*) ((const uint8_t*) w + 32);
87 k += 8 * sizeof(uint8_t);
88 }
89 p -= 1 * sizeof(void*);
90 } while (p != 0);
91
92 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
93 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
94
95 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
96
97 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
98
99 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
100 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
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 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
111
112 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
113
114 if (nc >= 4) {
115 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
116 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
117
118 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
119
120 nc -= 4;
121 } else {
122 if (nc & 2) {
123 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
124 c0 += 2;
125 vout = _mm_srli_epi32(vout, 16);
126 }
127 if (nc & 1) {
128 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
129 }
130
131 nc = 0;
132 }
133 } while (nc != 0);
134 }
135