import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor, alpha: float) -> torch.Tensor:
"""
Applies ELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
alpha (float): The alpha parameter for the ELU function.
Returns:
torch.Tensor: Output tensor with ELU applied, same shape as input.
"""
return F.elu(x, alpha=alpha)
class Model(nn.Module):
"""
Simple model that performs an ELU activation.
"""
def __init__(self, alpha):
"""
Initializes the ELU model.
Args:
alpha (float): The alpha parameter for the ELU function.
"""
super(Model, self).__init__()
self.alpha = alpha
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
"""
Applies ELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with ELU applied, same shape as input.
"""
return fn(x, self.alpha)
batch_size = 16
dim = 16384
alpha = 1.0
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [alpha]
import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(nn.Module):
"""
Simple model that performs an ELU activation.
"""
def __init__(self, alpha: float = 1.0):
"""
Initializes the ELU model.
Args:
alpha (float, optional): The alpha parameter for the ELU function. Defaults to 1.0.
"""
super(Model, self).__init__()
self.alpha = alpha
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies ELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with ELU applied, same shape as input.
"""
return F.elu(x, alpha=self.alpha)
batch_size = 16
dim = 16384
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [1.0] # Provide alpha value for initialization#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>
#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
// CUDA kernel leveraging shared memory to preload a tile of input data
__global__ void elu_kernel_shared(const float* x, float* out, float alpha, int n) {
extern __shared__ float tile[]; // dynamically allocated shared memory
int tid = threadIdx.x;
int globalIdx = blockIdx.x * blockDim.x + tid;
// Load input data from global memory to shared memory if within bounds
if (globalIdx < n) {
tile[tid] = x[globalIdx];
}
__syncthreads(); // Ensure all threads have loaded data to shared memory before processing
// Compute the ELU activation using data from shared memory
if (globalIdx < n) {
float val = tile[tid];
out[globalIdx] = (val > 0.0f) ? val : alpha * (expf(val) - 1.0f);
}
}
// Interface function called from Python
torch::Tensor elu_cuda_shared(torch::Tensor x, float alpha) {
CHECK_INPUT(x);
auto out = torch::empty_like(x);
int n = x.numel();
// Use block size of 256 threads
const int threads = 256;
const int blocks = (n + threads - 1) / threads;
// Allocate shared memory per block
size_t sharedMemSize = threads * sizeof(float);
elu_kernel_shared<<<blocks, threads, sharedMemSize>>>(x.data_ptr<float>(), out.data_ptr<float>(), alpha, n);
return out;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &elu_cuda_shared, "ELU activation with shared memory (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.742 | inst/cycle | 0.001 | 5 |
| Executed Ipc Elapsed | 0.296 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 20.962 | % | 0.937 | 5 |
| Issued Ipc Active | 0.838 | inst/cycle | 0.002 | 5 |
| SM Busy | 20.962 | % | 0.937 | 5 |
| Memory Throughput | 272954988986.316 | byte/second | 17600545736449437696.000 | 5 |
| Mem Busy | 13.008 | % | 0.033 | 5 |
| Max Bandwidth | 12.060 | % | 0.034 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 66.082 | % | 0.219 | 5 |
| Mem Pipes Busy | 16.202 | % | 0.062 | 5 |
| Warp Cycles Per Issued Instruction | 60.728 | cycle | 0.023 | 5 |
| Warp Cycles Per Executed Instruction | 68.514 | cycle | 0.029 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.140 | 0.000 | 5 | |
| Max Active Clusters | 0.000 | cluster | 0.000 | 5 |
| Max Cluster Size | 8.000 | block | 0.000 | 5 |
| Overall GPU Occupancy | 0.000 | % | 0.000 | 5 |
| Cluster Occupancy | 0.000 | % | 0.000 | 5 |
| Block Limit SM | 32.000 | block | 0.000 | 5 |
| Block Limit Registers | 16.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 16.000 | block | 0.000 | 5 |
| Block Limit Warps | 8.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 82.532 | % | 0.686 | 5 |
| Achieved Active Warps Per SM | 52.820 | warp | 0.280 | 5 |
| Rule | Description |
|---|---|
| WRN HighPipeUtilization | All compute pipelines are under-utilized. Either this kernel is very small or it doesn't issue enough warps per scheduler. Check the Launch Statistics and Scheduler Statistics sections for further details. |
| INF CPIStall | Check the Warp Stall Sampling (All Cycles) table for the top stall locations in your source based on sampling data. The Kernel Profiling Guide (https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-reference) provides more details on each stall reason. |
| WRN Occupancy | This kernel's theoretical occupancy is not impacted by any block limit. The difference between calculated theoretical (100.0%) and measured achieved occupancy (84.2%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide (https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on optimizing occupancy. |
| Operation / Metric | Value | Unit |
|---|---|---|
| aten::to | ||
| CPU Time | 642168.55 | μs |
| Device Time | 40.16 | μs |
| Self CPU Time | 31.07 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::_to_copy | ||
| CPU Time | 642137.48 | μs |
| Device Time | 40.16 | μs |
| Self CPU Time | 81.48 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::empty_strided | ||
| CPU Time | 658447.52 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 16744.01 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaDeviceGetStreamPriorityRange | ||
| CPU Time | 641464.79 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 641464.79 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaLaunchKernel | ||
| CPU Time | 406572.36 | μs |
| Device Time | 18483.25 | μs |
| Self CPU Time | 406572.36 | μs |
| Self Device Time | 18483.25 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| elu_kernel_shared(float const*, float*, float, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 26032.53 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 26032.53 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaEventRecord | ||
| CPU Time | 19293.87 | μs |
| Device Time | 35635.29 | μs |
| Self CPU Time | 19293.87 | μs |
| Self Device Time | 35635.29 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::zero_ | ||
| CPU Time | 57360.72 | μs |
| Device Time | 529649.91 | μs |
| Self CPU Time | 10151.61 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::fill_ | ||
| CPU Time | 47211.03 | μs |
| Device Time | 529649.91 | μs |
| Self CPU Time | 12744.75 | μs |
| Self Device Time | 529649.91 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, at::detail::Array<char*, 1> >(int, at::native::FillFunctor<int>, at::detail::Array<char*, 1>) | ||
| CPU Time | 0.00 | μs |
| Device Time | 529728.66 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 529728.66 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45282 warnings generated when compiling for host. Suppressed 45322 warnings (45275 in non-user code, 47 NOLINT). Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.