Welcome to mirror list, hosted at ThFree Co, Russian Federation.

types.h « intgemm - github.com/marian-nmt/intgemm/intgemm.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
blob: a4b35b4087cac3267c7a7e6e5df569c8fcc04f76 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
#pragma once
#include "intgemm/intgemm_config.h"

#include <exception>
#ifdef INTGEMM_COMPILER_SUPPORTS_AVX2
#include <immintrin.h>
#endif
#include <emmintrin.h>

#if defined(_MSC_VER) || defined(__INTEL_COMPILER)
/* MSVC does not appear to have target attributes but is also fine with just
 * using intrinsics anywhere.
 *
 * The Intel compiler has a bug whereby constructors with target attributes do
 * not link.  Like this program doesn't compile with icpc:
 * class Foo {
 *   public:
 *     __attribute__ ((target ("avx2"))) Foo() {}
 * };
 * int main() { Foo a; }
 *
 * It appears to be erroneously activating function multiversioning when only
 * one version of a constructor with target attributes is defined.  Normal
 * methods with one target attribute work fine.  The Intel compiler also allows
 * intrinsics without any target attributes so we just leave them blank.
 */
  #define INTGEMM_SSE2
  #define INTGEMM_SSSE3
  #define INTGEMM_AVX2
  #define INTGEMM_AVX512F
  #define INTGEMM_AVX512BW
  #define INTGEMM_AVX512DQ
  #define INTGEMM_AVX512VNNI
#else
  /* gcc and clang take lists of all the flavors */
  #define INTGEMM_SSE2 __attribute__ ((target ("sse2")))
  #define INTGEMM_SSSE3 __attribute__ ((target ("ssse3")))
  #define INTGEMM_AVX2 __attribute__ ((target ("avx2")))
  #define INTGEMM_AVX512F __attribute__ ((target ("avx512f")))
  #define INTGEMM_AVX512BW __attribute__ ((target ("avx512f,avx512bw,avx512dq")))
  #define INTGEMM_AVX512DQ __attribute__ ((target ("avx512f,avx512bw,avx512dq")))
  #define INTGEMM_AVX512VNNI __attribute__ ((target ("avx512f,avx512bw,avx512dq,avx512vnni")))
#endif
namespace intgemm {

// This will be thrown if a CPU isn't supported by the routines (16-bit without SSE2 or 8-bit without SSSE3).
class UnsupportedCPU : public std::exception {
  public:
    UnsupportedCPU() {}

    ~UnsupportedCPU() throw() {}

    const char *what() const throw() override {
      return "Integer matrix multiplication has not been efficiently implemented for your CPU.";
    }
};

typedef unsigned int Index;

// If you want to detect the CPU and dispatch yourself, here's what to use:
enum class CPUType {
  UNSUPPORTED = 0,
  SSE2 = 1,
  SSSE3 = 2,
  AVX2 = 3,
  AVX512BW = 4,
  AVX512VNNI = 5
};

// Running CPU type.  This is defined in intgemm.cc (as the dispatcher).
extern const CPUType kCPU;

struct MeanStd {
  float mean;
  float stddev;
};

#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
namespace AVX512VNNI {
typedef __m512i Register;
typedef __m512 FRegister;
} // namespace AVX512VNNI
#endif
#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
namespace AVX512BW {
typedef __m512i Register;
typedef __m512 FRegister;
} // namespace AVX512BW
#endif
#ifdef INTGEMM_COMPILER_SUPPORTS_AVX2
namespace AVX2 {
typedef __m256i Register;
typedef __m256 FRegister;
} // namespace AVX2
#endif
namespace SSSE3 {
typedef __m128i Register;
typedef __m128 FRegister;
} // namespace SSSE3
namespace SSE2 {
typedef __m128i Register;
typedef __m128 FRegister;
} // namespace SSE2

} // namespace intgemm