hip_vec_bf16_impl.h 4.52 KB
Newer Older
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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
#pragma once

#if USE_ROCM

#include <hip/hip_bf16.h>
#include <hip/hip_common.h>

// Adapted from flashinfer-rocm [PR#491](https://github.com/flashinfer-ai/flashinfer/pull/491)

using nv_bfloat16 = __hip_bfloat16;
using nv_bfloat162 = __hip_bfloat162;

__BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 make_bfloat162(const __hip_bfloat16 x, const __hip_bfloat16 y) {
  __hip_bfloat162 t;
  t.x = x;
  t.y = y;
  return t;
}

namespace sgl_hip {

// nv_bfloat16 x 1
template <>
struct vec_t<nv_bfloat16, 1> {
  nv_bfloat16 data;
  SGL_HIP_INLINE nv_bfloat16& operator[](size_t i) {
    return ((nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE const nv_bfloat16& operator[](size_t i) const {
    return ((const nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE nv_bfloat16* ptr() {
    return reinterpret_cast<nv_bfloat16*>(&data);
  }
  SGL_HIP_INLINE void load(const nv_bfloat16* ptr);
  SGL_HIP_INLINE void store(nv_bfloat16* ptr) const;
  template <typename T>
  SGL_HIP_INLINE void cast_from(const vec_t<T, 1>& src) {
    cast_from_impl(*this, src);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_load(const T* ptr) {
    cast_load_impl(*this, ptr);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_store(T* ptr) const {
    cast_store_impl(ptr, *this);
  }
};

SGL_HIP_INLINE void vec_t<nv_bfloat16, 1>::load(const nv_bfloat16* ptr) {
  data = *ptr;
}

SGL_HIP_INLINE void vec_t<nv_bfloat16, 1>::store(nv_bfloat16* ptr) const {
  *ptr = data;
}

// nv_bfloat16 x 2
template <>
struct vec_t<nv_bfloat16, 2> {
  nv_bfloat162 data;

  SGL_HIP_INLINE nv_bfloat16& operator[](size_t i) {
    return ((nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE const nv_bfloat16& operator[](size_t i) const {
    return ((const nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE nv_bfloat16* ptr() {
    return reinterpret_cast<nv_bfloat16*>(&data);
  }
  SGL_HIP_INLINE void load(const nv_bfloat16* ptr);
  SGL_HIP_INLINE void store(nv_bfloat16* ptr) const;
  template <typename T>
  SGL_HIP_INLINE void cast_from(const vec_t<T, 2>& src) {
    cast_from_impl(*this, src);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_load(const T* ptr) {
    cast_load_impl(*this, ptr);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_store(T* ptr) const {
    cast_store_impl(ptr, *this);
  }
};

SGL_HIP_INLINE void vec_t<nv_bfloat16, 2>::load(const nv_bfloat16* ptr) {
  data = *((nv_bfloat162*)ptr);
}

SGL_HIP_INLINE void vec_t<nv_bfloat16, 2>::store(nv_bfloat16* ptr) const {
  *((nv_bfloat162*)ptr) = data;
}

template <>
struct vec_t<nv_bfloat16, 4> {
  uint2 data;

  SGL_HIP_INLINE nv_bfloat16& operator[](size_t i) {
    return ((nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE const nv_bfloat16& operator[](size_t i) const {
    return ((const nv_bfloat16*)(&data))[i];
  }
  SGL_HIP_INLINE nv_bfloat16* ptr() {
    return reinterpret_cast<nv_bfloat16*>(&data);
  }
  SGL_HIP_INLINE void load(const nv_bfloat16* ptr);
  SGL_HIP_INLINE void store(nv_bfloat16* ptr) const;
  template <typename T>
  SGL_HIP_INLINE void cast_from(const vec_t<T, 4>& src) {
    cast_from_impl(*this, src);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_load(const T* ptr) {
    cast_load_impl(*this, ptr);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_store(T* ptr) const {
    cast_store_impl(ptr, *this);
  }
};

SGL_HIP_INLINE void vec_t<nv_bfloat16, 4>::load(const nv_bfloat16* ptr) {
  data = *((uint2*)ptr);
}

SGL_HIP_INLINE void vec_t<nv_bfloat16, 4>::store(nv_bfloat16* ptr) const {
  *((uint2*)ptr) = data;
}

// nv_bfloat16 x 8 or more

template <size_t vec_size>
struct vec_t<nv_bfloat16, vec_size> {
  uint4 data[vec_size / 8];

  SGL_HIP_INLINE nv_bfloat16& operator[](size_t i) {
    return ((nv_bfloat16*)data)[i];
  }
  SGL_HIP_INLINE const nv_bfloat16& operator[](size_t i) const {
    return ((const nv_bfloat16*)data)[i];
  }
  SGL_HIP_INLINE nv_bfloat16* ptr() {
    return reinterpret_cast<nv_bfloat16*>(&data);
  }
  SGL_HIP_INLINE void load(const nv_bfloat16* ptr) {
#pragma unoll
    for (size_t i = 0; i < vec_size / 8; ++i) {
      data[i] = ((uint4*)ptr)[i];
    }
  }
  SGL_HIP_INLINE void store(nv_bfloat16* ptr) const {
#pragma unoll
    for (size_t i = 0; i < vec_size / 8; ++i) {
      ((uint4*)ptr)[i] = data[i];
    }
  }
  template <typename T>
  SGL_HIP_INLINE void cast_from(const vec_t<T, vec_size>& src) {
    cast_from_impl(*this, src);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_load(const T* ptr) {
    cast_load_impl(*this, ptr);
  }
  template <typename T>
  SGL_HIP_INLINE void cast_store(T* ptr) const {
    cast_store_impl(ptr, *this);
  }
};

}  // namespace sgl_hip

#endif