import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies SELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with SELU applied, same shape as input.
"""
return F.selu(x)
class Model(nn.Module):
"""
Simple model that performs a SELU activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
return fn(x)
batch_size = 16
dim = 16384
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [] # No special initialization inputs needed
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs a SELU activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies SELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with SELU applied, same shape as input.
"""
return torch.selu(x)
batch_size = 16
dim = 16384
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [] # No special initialization inputs needed#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>
template <typename scalar_t>
__device__ __forceinline__ scalar_t my_exp(scalar_t x);
template <>
__device__ __forceinline__ float my_exp<float>(float x) {
return expf(x);
}
template <>
__device__ __forceinline__ double my_exp<double>(double x) {
return exp(x);
}
template <typename scalar_t>
__global__ void selu_kernel_shared(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const size_t numel) {
__shared__ scalar_t shared_data[256];
const int tid = threadIdx.x;
const int gid = blockIdx.x * blockDim.x + threadIdx.x;
const int stride = blockDim.x * gridDim.x;
const scalar_t alpha = 1.67326324235437728481;
const scalar_t lambda = 1.05070098735548049342;
for (int idx = gid; idx < numel; idx += stride) {
// Load data into shared memory
shared_data[tid] = __ldg(&input[idx]);
__syncthreads();
// Process data from shared memory
const scalar_t x = shared_data[tid];
const scalar_t result = (x > static_cast<scalar_t>(0))
? x
: alpha * (my_exp(x) - static_cast<scalar_t>(1));
output[idx] = lambda * result;
__syncthreads();
}
}
torch::Tensor selu_forward(torch::Tensor input) {
TORCH_CHECK(input.is_cuda(), "Input tensor must be a CUDA tensor");
auto output = torch::empty_like(input);
const size_t numel = input.numel();
const int threads = 256;
const int blocks = std::min(65535, (int)((numel + threads - 1) / threads));
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "selu_forward_cuda", ([&] {
const scalar_t* input_ptr = input.data_ptr<scalar_t>();
scalar_t* output_ptr = output.data_ptr<scalar_t>();
selu_kernel_shared<scalar_t><<<blocks, threads>>>(input_ptr, output_ptr, numel);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &selu_forward, "SELU Activation Forward with Shared Memory (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.212 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.498 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 32.636 | % | 0.007 | 5 |
| Issued Ipc Active | 1.304 | inst/cycle | 0.000 | 5 |
| SM Busy | 32.636 | % | 0.007 | 5 |
| Memory Throughput | 277601673709.032 | byte/second | 5716683369323156480.000 | 5 |
| Mem Busy | 13.150 | % | 0.015 | 5 |
| Max Bandwidth | 12.212 | % | 0.012 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 66.016 | % | 0.150 | 5 |
| Mem Pipes Busy | 11.490 | % | 0.011 | 5 |
| Warp Cycles Per Issued Instruction | 39.326 | cycle | 0.153 | 5 |
| Warp Cycles Per Executed Instruction | 42.374 | cycle | 0.178 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.330 | 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 | 80.604 | % | 0.068 | 5 |
| Achieved Active Warps Per SM | 51.588 | warp | 0.029 | 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 (80.3%) 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 | 342396.22 | μs |
| Device Time | 39.97 | μs |
| Self CPU Time | 33.57 | 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 | 342362.65 | μs |
| Device Time | 39.97 | μs |
| Self CPU Time | 70.77 | 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 | 357213.80 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 15253.31 | 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 | 341781.29 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 341781.29 | 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 | 412287.67 | μs |
| Device Time | 19112.85 | μs |
| Self CPU Time | 412287.67 | μs |
| Self Device Time | 19112.85 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void selu_kernel_shared<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 27491.45 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 27491.45 | 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 | 19622.33 | μs |
| Device Time | 36719.20 | μs |
| Self CPU Time | 19622.33 | μs |
| Self Device Time | 36719.20 | 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 | 57851.92 | μs |
| Device Time | 545273.72 | μs |
| Self CPU Time | 10413.30 | 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 | 47440.21 | μs |
| Device Time | 545273.72 | μs |
| Self CPU Time | 13258.03 | μs |
| Self Device Time | 545273.72 | 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 | 545273.72 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 545273.72 | 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 45323 warnings (45276 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.