1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_qc8_igemm_minmax_fp32_ukernel_2x4c2s4__sse41_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2s4__sse41_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
50
51 do {
52 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53 __m128i vacc1x0123 = vacc0x0123;
54 w = (const void*) ((const int32_t*) w + 4);
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const int8_t* restrict a1 = a[1];
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 do {
70 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
71 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
72 a0 += 8;
73 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
74 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
75 a1 += 8;
76
77 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
78 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
81 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
82 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
83 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
84 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
85 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
86
87 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
88 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
89 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
90 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
91 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
92 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
95 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
97 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
98 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
99 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
102 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
103
104 w = (const void*) ((const int8_t*) w + 32);
105 k -= 8 * sizeof(int8_t);
106 } while (k != 0);
107 p -= 2 * sizeof(void*);
108 } while (p != 0);
109
110 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
111 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
112
113 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
114 w = (const void*) ((const float*) w + 4);
115 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
116 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
117
118 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
119 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
120 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
121
122 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
123 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
124
125 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
126 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
127
128
129 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
130
131 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
132
133 if (nc >= 4) {
134 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
135 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
136 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
137 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
138
139 a = (const int8_t**restrict) ((uintptr_t) a - ks);
140
141 nc -= 4;
142 } else {
143 if (nc & 2) {
144 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
145 c1 += 2;
146 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
147 c0 += 2;
148 vout = _mm_srli_epi32(vout, 16);
149 }
150 if (nc & 1) {
151 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
152 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
153 }
154
155 nc = 0;
156 }
157 } while (nc != 0);
158 }
159