-
Notifications
You must be signed in to change notification settings - Fork 134
/
gelu.cu
224 lines (194 loc) · 10 KB
/
gelu.cu
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
#include <stdio.h>
#include <stdlib.h>
#include <float.h>
#include <vector>
#include <algorithm>
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <cuda_bf16.h>
#include <cuda_fp8.h>
#include <torch/types.h>
#include <torch/extension.h>
#define WARP_SIZE 32
#define INT4(value) (reinterpret_cast<int4*>(&(value))[0])
#define FLOAT4(value) (reinterpret_cast<float4*>(&(value))[0])
#define HALF2(value) (reinterpret_cast<half2*>(&(value))[0])
#define BFLOAT2(value) (reinterpret_cast<__nv_bfloat162*>(&(value))[0])
#define LDST128BITS(value) (reinterpret_cast<float4*>(&(value))[0])
#define MAX_EXP_F32 88.3762626647949f
#define MIN_EXP_F32 -88.3762626647949f
#define MAX_EXP_F16 __float2half(11.089866488461016f)
#define MIN_EXP_F16 __float2half(-9.704060527839234f)
#define SQRT_2_PI M_SQRT2 * M_2_SQRTPI * 0.5f
#define HALF_1 __float2half(1.0f)
#define HALF_2 __float2half(2.0f)
#define HALF_DIV2 __float2half(0.5f)
// to clear the error among self defined gelu and pytorch gelu. Calculate $\sqrt{\frac{\pi}{2}}$ by $\sqrt{2 * \pi} / 2$
#define HALF_SQRT_2_PI __float2half(M_SQRT2) * __float2half(M_2_SQRTPI) * HALF_DIV2
#define HALF_V_APP __float2half(0.044715f)
#define HALF_GELU_OPS gelu_tanh_approximate
#define GELU_OPS gelu_tanh_approximate
// There is no half presicion operation like sinh, cosh, tanh. [Half Math Functions](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF__FUNCTIONS.html#group__CUDA__MATH____HALF__FUNCTIONS)
// $$ tanh(x) = \frac{exp^{2x} - 1}{exp^{2x} + 1}$$
// But ops above will introduce error.
// pytorch transform type while do tanh operator which include in the [pytorch/c10/util/BFloat16-math.h](https://github.com/pytorch/pytorch/blob/main/c10/util/BFloat16-math.h)
__inline__ __device__ half gelu_tanh_approximate(half x){
half x_cube = x * x * x;
// compute mid value : inner = 0.7978845608 * (x + 0.044715 * x * x * x)
half inner = HALF_SQRT_2_PI * (x + HALF_V_APP * x_cube);
// compute tanh
return HALF_DIV2 * x * (HALF_1 + ((hexp(inner * HALF_2) - HALF_1) / (hexp(inner * HALF_2) + HALF_1)));
}
__inline__ __device__ float gelu_tanh_approximate(float x){
return 0.5f * x * (1.0f + tanhf(SQRT_2_PI * (x + 0.044715f * x * x * x)));
}
__inline__ __device__ float gelu_none_approximate(float x){
return x * 0.5 * (1 + erff(x * M_SQRT1_2));
}
// -------------------------------------- FP32 --------------------------------------
// GELU tanh approximate: x, y:x 0.5 * x * (1.0 + tanh(0.7978845608 * x * (1.0 + 0.044715 * x * x)))
// grid(N/256), block(K=256)
__global__ void gelu_f32_kernel(float* x, float* y, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) {
float v = fminf(fmaxf(x[idx], MIN_EXP_F32), MAX_EXP_F32);
y[idx] = GELU_OPS(v);
}
}
// GELU tanh approximate; Vec4
// grid(N/256), block(256/4)
__global__ void gelu_f32x4_kernel(float* x, float* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4;
float4 reg_x = FLOAT4(x[idx]);
float4 reg_y;
reg_x.x = fminf(fmaxf(reg_x.x, MIN_EXP_F32), MAX_EXP_F32);
reg_x.y = fminf(fmaxf(reg_x.y, MIN_EXP_F32), MAX_EXP_F32);
reg_x.z = fminf(fmaxf(reg_x.z, MIN_EXP_F32), MAX_EXP_F32);
reg_x.w = fminf(fmaxf(reg_x.w, MIN_EXP_F32), MAX_EXP_F32);
reg_y.x = GELU_OPS(reg_x.x);
reg_y.y = GELU_OPS(reg_x.y);
reg_y.z = GELU_OPS(reg_x.z);
reg_y.w = GELU_OPS(reg_x.w);
if ((idx + 0) < N) { FLOAT4(y[idx]) = reg_y; }
}
// -------------------------------------- FP16 --------------------------------------
// GELU approximate: x, y:x 0.5 * x * (1.0 + tanh(0.7978845608 (x + 0.044715 * x * x * x))) Vec4
__global__ void gelu_f16_kernel(half* x, half* y, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) {
half v = x[idx];
v = __hmin(__hmax(v, MIN_EXP_F16), MAX_EXP_F16);
y[idx] = HALF_GELU_OPS(v);
}
}
__global__ void gelu_f16x2_kernel(half* x, half* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
half2 reg_x = HALF2(x[idx]);
half2 reg_y;
reg_x.x = __hmin(__hmax(reg_x.x, MIN_EXP_F16), MAX_EXP_F16);
reg_x.y = __hmin(__hmax(reg_x.y, MIN_EXP_F16), MAX_EXP_F16);
reg_y.x = HALF_GELU_OPS(reg_x.x);
reg_y.y = HALF_GELU_OPS(reg_x.y);
if ((idx + 0) < N) { HALF2(y[idx]) = reg_y; }
}
// unpack f16x8
__global__ void gelu_f16x8_kernel(half* x, half* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 8;
half2 reg_x_0 = HALF2(x[idx + 0]);
half2 reg_x_1 = HALF2(x[idx + 2]);
half2 reg_x_2 = HALF2(x[idx + 4]);
half2 reg_x_3 = HALF2(x[idx + 6]);
reg_x_0.x = __hmin(__hmax(reg_x_0.x, MIN_EXP_F16), MAX_EXP_F16);
reg_x_0.y = __hmin(__hmax(reg_x_0.y, MIN_EXP_F16), MAX_EXP_F16);
reg_x_1.x = __hmin(__hmax(reg_x_1.x, MIN_EXP_F16), MAX_EXP_F16);
reg_x_1.y = __hmin(__hmax(reg_x_1.y, MIN_EXP_F16), MAX_EXP_F16);
reg_x_2.x = __hmin(__hmax(reg_x_2.x, MIN_EXP_F16), MAX_EXP_F16);
reg_x_2.y = __hmin(__hmax(reg_x_2.y, MIN_EXP_F16), MAX_EXP_F16);
reg_x_3.x = __hmin(__hmax(reg_x_3.x, MIN_EXP_F16), MAX_EXP_F16);
reg_x_3.y = __hmin(__hmax(reg_x_3.y, MIN_EXP_F16), MAX_EXP_F16);
half2 reg_y_0, reg_y_1, reg_y_2, reg_y_3;
reg_x_0.x = HALF_GELU_OPS(reg_x_0.x);
reg_x_0.y = HALF_GELU_OPS(reg_x_0.y);
reg_x_1.x = HALF_GELU_OPS(reg_x_1.x);
reg_x_1.y = HALF_GELU_OPS(reg_x_1.y);
reg_x_2.x = HALF_GELU_OPS(reg_x_2.x);
reg_x_2.y = HALF_GELU_OPS(reg_x_2.y);
reg_x_3.x = HALF_GELU_OPS(reg_x_3.x);
reg_x_3.y = HALF_GELU_OPS(reg_x_3.y);
if ((idx + 0) < N) { HALF2(y[idx + 0]) = reg_x_0; }
if ((idx + 2) < N) { HALF2(y[idx + 2]) = reg_x_1; }
if ((idx + 4) < N) { HALF2(y[idx + 4]) = reg_x_2; }
if ((idx + 6) < N) { HALF2(y[idx + 6]) = reg_x_3; }
}
// pack f16x8
__global__ void gelu_f16x8_pack_kernel(half* x, half* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 8;
// temporary register(memory), .local space in ptx, addressable
half pack_x[8], pack_y[8]; // 8x16 bits=128 bits.
// reinterpret as float4 and load 128 bits in 1 memory issue.
LDST128BITS(pack_x[0]) = LDST128BITS(x[idx]); // load 128 bits
#pragma unroll
for (int i = 0; i < 8; ++i) {
half v = __hmin(__hmax(pack_x[i], MIN_EXP_F16), MAX_EXP_F16);
pack_y[i] = HALF_GELU_OPS(v);
}
// reinterpret as float4 and store 128 bits in 1 memory issue.
if ((idx + 7) < N) { LDST128BITS(y[idx]) = LDST128BITS(pack_y[0]); }
}
// --------------------- PyTorch bindings for custom kernel -----------------------
#define STRINGFY(str) #str
#define TORCH_BINDING_COMMON_EXTENSION(func) \
m.def(STRINGFY(func), &func, STRINGFY(func));
#define CHECK_TORCH_TENSOR_DTYPE(T, th_type) \
if(((T).options().dtype() != (th_type))) { \
std::cout << "Tensor Info:" << (T).options() << std::endl; \
throw std::runtime_error("values must be "#th_type); \
}
#define TORCH_BINDING_GELU(packed_type, th_type, element_type, n_elements) \
void gelu_##packed_type(torch::Tensor x, torch::Tensor y) { \
CHECK_TORCH_TENSOR_DTYPE(x, (th_type)) \
CHECK_TORCH_TENSOR_DTYPE(y, (th_type)) \
const int ndim = x.dim(); \
if (ndim != 2) { \
int N = 1; \
for (int i = 0; i < ndim; ++i) { N *= x.size(i); } \
dim3 block(256 / (n_elements)); \
dim3 grid((N + 256 - 1) / 256); \
gelu_##packed_type##_kernel<<<grid, block>>>( \
reinterpret_cast<element_type*>(x.data_ptr()), \
reinterpret_cast<element_type*>(y.data_ptr()), N); \
} else { \
const int S = x.size(0); \
const int K = x.size(1); \
const int N = S * K; \
if ((K/(n_elements)) <= 1024) { \
dim3 block(K/(n_elements)); \
dim3 grid(S); \
gelu_##packed_type##_kernel<<<grid, block>>>( \
reinterpret_cast<element_type*>(x.data_ptr()), \
reinterpret_cast<element_type*>(y.data_ptr()), N); \
} else { \
int N = 1; \
for (int i = 0; i < ndim; ++i) { N *= x.size(i); } \
dim3 block(256 / (n_elements)); \
dim3 grid((N + 256 - 1) / 256); \
gelu_##packed_type##_kernel<<<grid, block>>>( \
reinterpret_cast<element_type*>(x.data_ptr()), \
reinterpret_cast<element_type*>(y.data_ptr()), N); \
} \
} \
}
TORCH_BINDING_GELU(f32, torch::kFloat32, float, 1)
TORCH_BINDING_GELU(f32x4, torch::kFloat32, float, 4)
TORCH_BINDING_GELU(f16, torch::kHalf, half, 1)
TORCH_BINDING_GELU(f16x2, torch::kHalf, half, 2)
TORCH_BINDING_GELU(f16x8, torch::kHalf, half, 8)
TORCH_BINDING_GELU(f16x8_pack, torch::kHalf, half, 8)
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
TORCH_BINDING_COMMON_EXTENSION(gelu_f32)
TORCH_BINDING_COMMON_EXTENSION(gelu_f32x4)
TORCH_BINDING_COMMON_EXTENSION(gelu_f16)
TORCH_BINDING_COMMON_EXTENSION(gelu_f16x2)
TORCH_BINDING_COMMON_EXTENSION(gelu_f16x8)
TORCH_BINDING_COMMON_EXTENSION(gelu_f16x8_pack)
}