|
| 1 | +#include <cuda_runtime.h> |
| 2 | +#include <stdio.h> |
| 3 | + |
| 4 | +#define SQRT_2_PI 0.7978845608028654 |
| 5 | + |
| 6 | +template <typename T> |
| 7 | +__global__ void gelu(const T * in, T * out, int const size) { |
| 8 | + int const idx = blockIdx.x * blockDim.x + threadIdx.x; |
| 9 | + if (idx >= size) {return;} |
| 10 | + |
| 11 | + out[idx] = in[idx] * 0.5 * (1.0 + tanh(SQRT_2_PI * (in[idx] + 0.044715 * in[idx] * in[idx] *in[idx]))); |
| 12 | +} |
| 13 | + |
| 14 | +template <typename T> |
| 15 | +__global__ void gelu_grad(const T * dy, const T * in, T * out, int const size) { |
| 16 | + int const idx = blockIdx.x * blockDim.x + threadIdx.x; |
| 17 | + if (idx >= size) {return;} |
| 18 | + |
| 19 | + // out[idx] = in[idx] * 0.5 * (1.0 + tanh(SQRT_2_PI * (in[idx] + 0.044715 * in[idx] * in[idx] *in[idx]))); |
| 20 | + T const var1 = tanh(SQRT_2_PI * (in[idx] + 0.044715 * in[idx] * in[idx] *in[idx])); |
| 21 | + out[idx] = dy[idx] * (0.5 * SQRT_2_PI * in[idx] * (1 - var1 * var1) * (0.134145 * in[idx] * in[idx] + 1) + 0.5 * var1 + 0.5); |
| 22 | +} |
| 23 | + |
| 24 | +template <typename T> |
| 25 | +__global__ void gelu_grad_grad(const T * dy, const T * dy_, const T * in, T * out, int const size) { |
| 26 | + int const idx = blockIdx.x * blockDim.x + threadIdx.x; |
| 27 | + if (idx >= size) {return;} |
| 28 | + |
| 29 | + // out[idx] = in[idx] * 0.5 * (1.0 + tanh(SQRT_2_PI * (in[idx] + 0.044715 * in[idx] * in[idx] *in[idx]))); |
| 30 | + T const var1 = tanh(SQRT_2_PI * (in[idx] + 0.044715 * in[idx] * in[idx] *in[idx])); |
| 31 | + T const var2 = SQRT_2_PI * (1 - var1 * var1) * (0.134145 * in[idx] * in[idx] + 1); |
| 32 | + |
| 33 | + out[idx] = dy[idx] * dy_[idx] * (0.134145 * SQRT_2_PI * in[idx] * in[idx] * (1 - var1 * var1) - SQRT_2_PI * in[idx] * var2 * (0.134145 * in[idx] * in[idx] + 1) * var1 + var2); |
| 34 | +} |
| 35 | + |
| 36 | + |
| 37 | +void GeluLauncher(const float * in, float * out, int const size) { |
| 38 | + int const THREAD_ITEMS = 1024; |
| 39 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 40 | + |
| 41 | + gelu<<<BLOCK_NUMS, THREAD_ITEMS>>>(in, out, size); |
| 42 | +} |
| 43 | + |
| 44 | +void GeluLauncher(const double * in, double * out, int const size) { |
| 45 | + int const THREAD_ITEMS = 1024; |
| 46 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 47 | + |
| 48 | + gelu<<<BLOCK_NUMS, THREAD_ITEMS>>>(in, out, size); |
| 49 | +} |
| 50 | + |
| 51 | +void GeluGradLauncher(const float * dy, const float * in, float * out, int const size) { |
| 52 | + int const THREAD_ITEMS = 1024; |
| 53 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 54 | + |
| 55 | + gelu_grad<<<BLOCK_NUMS, THREAD_ITEMS>>>(dy, in, out, size); |
| 56 | +} |
| 57 | + |
| 58 | +void GeluGradLauncher(const double * dy, const double * in, double * out, int const size) { |
| 59 | + int const THREAD_ITEMS = 1024; |
| 60 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 61 | + |
| 62 | + gelu_grad<<<BLOCK_NUMS, THREAD_ITEMS>>>(dy, in, out, size); |
| 63 | +} |
| 64 | + |
| 65 | +void GeluGradGradLauncher(const float * dy, const float * dy_, const float * in, float * out, int const size) { |
| 66 | + int const THREAD_ITEMS = 1024; |
| 67 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 68 | + |
| 69 | + gelu_grad_grad<<<BLOCK_NUMS, THREAD_ITEMS>>>(dy, dy_, in, out, size); |
| 70 | +} |
| 71 | + |
| 72 | +void GeluGradGradLauncher(const double * dy, const double * dy_, const double * in, double * out, int const size) { |
| 73 | + int const THREAD_ITEMS = 1024; |
| 74 | + int const BLOCK_NUMS = (size + THREAD_ITEMS - 1) / THREAD_ITEMS; |
| 75 | + |
| 76 | + gelu_grad_grad<<<BLOCK_NUMS, THREAD_ITEMS>>>(dy, dy_, in, out, size); |
| 77 | +} |
0 commit comments