forked from ccf-ai-infra/GPUCodeForces
96 lines
3.3 KiB
Python
96 lines
3.3 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
torch::Tensor geglu_dynamic_parallel(torch::Tensor input);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
|
|
__device__ float gelu_exact(float x) {
|
|
return 0.5f * x * (1.0f + erff(x * 0.7071067811865475f));
|
|
}
|
|
|
|
__global__ void geglu_dynamic_kernel(
|
|
const float* __restrict__ input,
|
|
float* __restrict__ output,
|
|
int feature_dim, int total_elements) {
|
|
|
|
extern __shared__ float shared_data[];
|
|
|
|
int tid = threadIdx.x;
|
|
int bid = blockIdx.x;
|
|
int bdim = blockDim.x;
|
|
|
|
// 动态确定每个block处理的元素数量
|
|
int elements_per_block = min(bdim * 4, total_elements - bid * bdim * 4);
|
|
elements_per_block = max(elements_per_block, 0);
|
|
|
|
float* gate_shared = shared_data;
|
|
float* act_shared = shared_data + elements_per_block;
|
|
|
|
// 协作加载
|
|
for (int i = tid; i < elements_per_block; i += bdim) {
|
|
int global_idx = bid * bdim * 4 + i;
|
|
if (global_idx < total_elements) {
|
|
int row = global_idx / (feature_dim / 2);
|
|
int col = global_idx % (feature_dim / 2);
|
|
|
|
gate_shared[i] = input[row * feature_dim + col];
|
|
act_shared[i] = input[row * feature_dim + col + (feature_dim / 2)];
|
|
}
|
|
}
|
|
__syncthreads();
|
|
|
|
// 处理
|
|
for (int i = tid; i < elements_per_block; i += bdim) {
|
|
int global_idx = bid * bdim * 4 + i;
|
|
if (global_idx < total_elements) {
|
|
float gate_val = gate_shared[i];
|
|
float act_val = act_shared[i];
|
|
output[global_idx] = gelu_exact(gate_val) * act_val;
|
|
}
|
|
}
|
|
}
|
|
|
|
torch::Tensor geglu_dynamic_parallel(torch::Tensor input) {
|
|
input = input.contiguous();
|
|
auto sizes = input.sizes().vec();
|
|
int feature_dim = sizes.back();
|
|
sizes.back() /= 2;
|
|
auto output = torch::empty(sizes, input.options());
|
|
|
|
int total_elements = output.numel();
|
|
int threads = 128;
|
|
int blocks = (total_elements + threads * 4 - 1) / (threads * 4);
|
|
int shared_mem = threads * 4 * 2 * sizeof(float);
|
|
|
|
geglu_dynamic_kernel<<<blocks, threads, shared_mem>>>(
|
|
input.data_ptr<float>(), output.data_ptr<float>(),
|
|
feature_dim, total_elements);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="geglu_dynamic",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["geglu_dynamic_parallel"],
|
|
extra_cuda_cflags=["-O3"],
|
|
verbose=True
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.op.geglu_dynamic_parallel(x) |