Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-sse41-madd.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-ssevnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x4c8__sse41_madd(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45
46 const __m128i vsign_mask = _mm_set1_epi8(0x80);
47 XNN_FORCE_REALIZATION(vsign_mask);
48 const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128);
49 const __m128 voutput_min = _mm_set1_ps(params->scalar.min);
50 const __m128 voutput_max = _mm_set1_ps(params->scalar.max);
51 const __m128i vmask = _mm_set1_epi8(0x0F);
52 XNN_FORCE_REALIZATION(vmask);
53 do {
54 const __m128i vksum0123 = _mm_load_si128(w);
55 const __m128i vsum0x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point0);
56 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x0123, _mm_setzero_si128());
57 __m128i vacc0x23 = _mm_unpackhi_epi32(vsum0x0123, _mm_setzero_si128());
58 __m128i vacc1x0x01 = _mm_setzero_si128();
59 __m128i vacc1x0x23 = _mm_setzero_si128();
60 w = (const int32_t*) w + 4;
61
62 size_t k = kc;
63 while (k >= 16 * sizeof(int8_t)) {
64 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
65 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
66 a0 += 16;
67
68 const __m128i vbb01234567x0123 = _mm_load_si128(w);
69 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
70 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
71 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
72 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
73 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
74 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
75 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
76
77 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
78 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
79 vacc1x0x01 = _mm_dpbusd_epi32_madd(vacc1x0x01, va0x89ABCDEF, vb01234567x23);
80 vacc1x0x23 = _mm_dpbusd_epi32_madd(vacc1x0x23, va0x89ABCDEF, vb89ABCDEFx23);
81
82 w = (const int8_t*) w + 32;
83 k -= 16 * sizeof(int8_t);
84 }
85
86 if (k != 0) {
87 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
88 a0 += 8;
89
90 const __m128i vbb01234567x0123 = _mm_load_si128(w);
91 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
92 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
93 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
94
95 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
96 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
97
98 w = (const int8_t*) w + 32;
99 k -= 8 * sizeof(int8_t);
100 }
101 vacc0x01 = _mm_add_epi32(vacc0x01, vacc1x0x01);
102 vacc0x23 = _mm_add_epi32(vacc0x23, vacc1x0x23);
103
104 // Add adjacent pairs
105 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
106
107 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
108
109 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
110
111 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
112 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
113 w = (const float*) w + 8;
114
115 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
116
117 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
118
119 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
120
121 if XNN_LIKELY(nc >= 4) {
122 _mm_storeu_ps(c0, vout0x0123);
123 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
124 c0 = (float*) ((uintptr_t) c0 + cn_stride);
125 nc -= 4;
126 } else {
127 if (nc & 2) {
128 _mm_storel_pi((__m64*) c0, vout0x0123);
129 c0 += 2;
130 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
131 }
132 if (nc & 1) {
133 _mm_store_ss(c0, vout0x0123);
134 }
135 nc = 0;
136 }
137 } while (nc != 0);
138 }