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 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64(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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43
44 do {
45 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
46 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
47 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
48 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
49 w = (const int32_t*) w + 4;
50
51 size_t k = 0;
52 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
53 while (k < kc) {
54 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
55 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
56 a0 += 8;
57
58 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
59 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
60
61 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
62 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
63 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
64
65 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
66 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
67 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
68
69 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
70 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
71 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
72
73 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
74
75 w = (const void*) ((const uint8_t*) w + 32);
76 k += 8 * sizeof(uint8_t);
77 }
78
79 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
80 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
81
82 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
83
84 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
85
86 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
87 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
88
89 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
90 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
91
92 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
93
94 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
95 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
96
97 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
98
99 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
100
101 if (nc >= 4) {
102 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
103
104 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
105
106 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
107
108 nc -= 4;
109 } else {
110 if (nc & 2) {
111 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
112 c0 += 2;
113 vout = _mm_srli_epi32(vout, 16);
114 }
115 if (nc & 1) {
116 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
117 }
118
119 nc = 0;
120 }
121 } while (nc != 0);
122 }
123