GPUCodeForces/S1/uucoco_#115/ValueLoss_cuda.py

86 lines
2.5 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 = """
torch::Tensor value_loss_cuda(torch::Tensor values, torch::Tensor returns);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
__global__ void value_loss_kernel(
const float* __restrict__ values,
const float* __restrict__ returns,
float* __restrict__ output,
int n)
{
extern __shared__ float sdata[];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int gridSize = blockDim.x * gridDim.x;
float local_sum = 0.0f;
while (i < n) {
float diff = values[i] - returns[i];
local_sum += diff * diff;
i += gridSize;
}
sdata[tid] = local_sum;
__syncthreads();
for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {
if (tid < s) {
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
if (tid == 0) {
atomicAdd(output, sdata[0] / n);
}
}
torch::Tensor value_loss_cuda(torch::Tensor values, torch::Tensor returns) {
auto values_c = values.contiguous();
auto returns_c = returns.contiguous();
int n = values_c.numel();
auto output = torch::zeros({1}, values.options());
const int threads = 256;
const int blocks = min((n + threads - 1) / threads, 1024);
const int shared_mem = threads * sizeof(float);
value_loss_kernel<<<blocks, threads, shared_mem>>>(
values_c.data_ptr<float>(),
returns_c.data_ptr<float>(),
output.data_ptr<float>(),
n
);
return output[0];
}
"""
self.op = load_inline(
name="value_loss_op",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["value_loss_cuda"],
extra_cuda_cflags=["-O3"],
verbose=False
)
def forward(self, values, returns):
return self.op.value_loss_cuda(values, returns)