1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
45 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
46 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
47 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
48 w = (const void*) ((const int32_t*) w + 4);
49
50 size_t k = 0;
51 while (k < kc) {
52 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
53 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
54 a0 += 8;
55
56 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
57 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
58 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
59
60 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
61 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
62 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
63 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
64 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
65
66 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
67 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
68
69 w = (const void*) ((const int8_t*) w + 32);
70 k += 8 * sizeof(int8_t);
71 }
72
73 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
74 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
75
76 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
77
78 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
79
80 const __m128 vscale0123 = _mm_load_ps((const float*) w);
81 w = (const void*) ((const float*) w + 4);
82 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
83
84 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
85 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
86
87 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
88
89 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
90 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
91
92
93 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
94
95 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
96
97 if (nc >= 4) {
98 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
99
100 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
101
102 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
103
104 nc -= 4;
105 } else {
106 if (nc & 2) {
107 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
108 c0 += 2;
109 vout = _mm_srli_epi32(vout, 16);
110 }
111 if (nc & 1) {
112 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
113 }
114
115 nc = 0;
116 }
117 } while (nc != 0);
118 }
119