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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21
22
23
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128(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)])24 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const uint8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 1);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(uint8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2);
46 const uint8_t* a0 = a;
47 uint8_t* c0 = c;
48
49 do {
50 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
51 w = (const void*) ((const int32_t*) w + 4);
52
53 size_t k = kc;
54 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
55 const __m128i vzero = _mm_setzero_si128();
56 while (k >= 8 * sizeof(uint8_t)) {
57 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
58 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
59 a0 += 8;
60
61 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
62 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
63 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
64
65 vacc0x0123 = _mm_maddd_epi16(
66 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
67
68 vacc0x0123 = _mm_maddd_epi16(
69 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
70 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
71 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
72 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
73
74 vacc0x0123 = _mm_maddd_epi16(
75 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
76
77 vacc0x0123 = _mm_maddd_epi16(
78 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
79
80 w = (const void*) ((const uint8_t*) w + 32);
81 k -= 8 * sizeof(uint8_t);
82 }
83 if (k != 0) {
84 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
85 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
86 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
87
88 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
89 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
90 w = (const void*) ((const uint8_t*) w + 8);
91
92 vacc0x0123 = _mm_maddd_epi16(
93 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
94
95 if (k > 2 * sizeof(uint8_t)) {
96 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
97 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
98 w = (const void*) ((const uint8_t*) w + 8);
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
102
103 if (k > 4 * sizeof(uint8_t)) {
104 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
105 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
106 w = (const void*) ((const uint8_t*) w + 8);
107
108 vacc0x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
110 }
111 }
112 }
113
114 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
115
116 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
117 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
118
119 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
120 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
121
122 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
123
124 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
125 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
126
127 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
128
129 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
130
131 if (nc >= 4) {
132 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
133
134 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
135
136 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
137
138 nc -= 4;
139 } else {
140 if (nc & 2) {
141 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
142 c0 += 2;
143 vout = _mm_srli_epi32(vout, 16);
144 }
145 if (nc & 1) {
146 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
147 }
148
149 nc = 0;
150 }
151 } while (nc != 0);
152 }
153