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 #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
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__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)])23 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const uint8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 uint8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % 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 const uint8_t* a0 = a;
46 uint8_t* c0 = c;
47
48 do {
49 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
50 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
51 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
52 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t k = 0;
56 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
57 const __m128i vzero = _mm_setzero_si128();
58 while (k < kc) {
59 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
60 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
61 a0 += 8;
62
63 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
64 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
65 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
66
67 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
68 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
69 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
70 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
71 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
72
73 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
74 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
75
76 w = (const void*) ((const uint8_t*) w + 32);
77 k += 8 * sizeof(uint8_t);
78 }
79
80 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
81 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
82
83 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
84
85 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
86
87 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
88 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
89
90 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
91 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
92
93 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
94
95 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
96 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
97
98 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
99
100 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
101
102 if (nc >= 4) {
103 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
104
105 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
106
107 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
108
109 nc -= 4;
110 } else {
111 if (nc & 2) {
112 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
113 c0 += 2;
114 vout = _mm_srli_epi32(vout, 16);
115 }
116 if (nc & 1) {
117 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
118 }
119
120 nc = 0;
121 }
122 } while (nc != 0);
123 }
124