forked from ccf-ai-infra/GPUCodeForces
117 lines
3.7 KiB
Python
117 lines
3.7 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 double_glu_cuda(torch::Tensor input);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
|
|
__device__ __forceinline__ float sigmoid_f(float x) {
|
|
return 1.0f / (1.0f + expf(-x));
|
|
}
|
|
|
|
__global__ void double_glu_vec4_kernel(
|
|
const float4* __restrict__ x,
|
|
float4* __restrict__ y,
|
|
int chunk_vec_dim,
|
|
int total_chunk_vecs)
|
|
{
|
|
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
int stride = gridDim.x * blockDim.x;
|
|
|
|
for (int i = idx; i < total_chunk_vecs; i += stride) {
|
|
int row = i / chunk_vec_dim;
|
|
int col = i % chunk_vec_dim;
|
|
|
|
// Input width = 4 * chunk
|
|
// Output width = 2 * chunk
|
|
|
|
int row_offset_in = row * 4 * chunk_vec_dim;
|
|
int row_offset_out = row * 2 * chunk_vec_dim;
|
|
|
|
// Process Pair 1 (G1, X1)
|
|
int g1_idx = row_offset_in + col;
|
|
int x1_idx = row_offset_in + chunk_vec_dim + col;
|
|
|
|
float4 g1 = x[g1_idx];
|
|
float4 x1 = x[x1_idx];
|
|
float4 out1;
|
|
|
|
out1.x = sigmoid_f(g1.x) * x1.x;
|
|
out1.y = sigmoid_f(g1.y) * x1.y;
|
|
out1.z = sigmoid_f(g1.z) * x1.z;
|
|
out1.w = sigmoid_f(g1.w) * x1.w;
|
|
|
|
y[row_offset_out + col] = out1;
|
|
|
|
// Process Pair 2 (G2, X2)
|
|
int g2_idx = row_offset_in + 2 * chunk_vec_dim + col;
|
|
int x2_idx = row_offset_in + 3 * chunk_vec_dim + col;
|
|
|
|
float4 g2 = x[g2_idx];
|
|
float4 x2 = x[x2_idx];
|
|
float4 out2;
|
|
|
|
out2.x = sigmoid_f(g2.x) * x2.x;
|
|
out2.y = sigmoid_f(g2.y) * x2.y;
|
|
out2.z = sigmoid_f(g2.z) * x2.z;
|
|
out2.w = sigmoid_f(g2.w) * x2.w;
|
|
|
|
y[row_offset_out + chunk_vec_dim + col] = out2;
|
|
}
|
|
}
|
|
|
|
torch::Tensor double_glu_cuda(torch::Tensor input) {
|
|
auto x_c = input.contiguous();
|
|
|
|
int last_dim = x_c.size(-1);
|
|
TORCH_CHECK(last_dim % 16 == 0, "Feature dim must be divisible by 16 (4 chunks * float4) for optimization");
|
|
|
|
auto out_sizes = x_c.sizes().vec();
|
|
out_sizes.back() /= 2;
|
|
auto output = torch::empty(out_sizes, x_c.options());
|
|
|
|
int chunk_dim = last_dim / 4;
|
|
int chunk_vec_dim = chunk_dim / 4;
|
|
|
|
int batch_size = x_c.numel() / last_dim;
|
|
int total_chunk_vecs = batch_size * chunk_vec_dim;
|
|
|
|
int threads = 256;
|
|
int blocks = (total_chunk_vecs + threads - 1) / threads;
|
|
if (blocks > 65535) blocks = 65535;
|
|
if (blocks == 0) blocks = 1;
|
|
|
|
double_glu_vec4_kernel<<<blocks, threads>>>(
|
|
reinterpret_cast<const float4*>(x_c.data_ptr<float>()),
|
|
reinterpret_cast<float4*>(output.data_ptr<float>()),
|
|
chunk_vec_dim,
|
|
total_chunk_vecs
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="double_glu_opt_vec4",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["double_glu_cuda"],
|
|
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.op.double_glu_cuda(x) |