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>
// Device helper: define an inline exponential function for float and double types
template <typename scalar_t>
__device__ inline scalar_t my_exp(scalar_t x);
template <>
__device__ inline float my_exp<float>(float x) {
return expf(x);
}
template <>
__device__ inline double my_exp<double>(double x) {
return exp(x);
}
// CUDA kernel that applies the SELU activation ensuring memory coalescing.
// Each thread in a warp accesses consecutive memory locations, which maximizes
// memory bandwidth utilization on the NVIDIA H100 GPU.
template <typename scalar_t>
__global__ void selu_kernel_coalesced(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
// Compute global index for each thread
size_t idx = blockIdx.x * blockDim.x + threadIdx.x;
// Use grid-stride loop so that each thread processes a contiguous chunk
size_t stride = blockDim.x * gridDim.x;
// Loop over elements in steps of 'stride'. When entering the loop,
// threads in the same warp will access consecutive elements, ensuring
// coalesced global memory accesses.
for (size_t i = idx; i < numel; i += stride) {
scalar_t x = input[i];
scalar_t res = (x > static_cast<scalar_t>(0))
? x
: static_cast<scalar_t>(1.67326324235437728481) * (my_exp(x) - static_cast<scalar_t>(1));
output[i] = static_cast<scalar_t>(1.05070098735548049342) * res;
}
}
// Host function that sets up the kernel launch parameters and calls the kernel.
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);
size_t numel = input.numel();
// Using 1024 threads per block for high occupancy
const int threads = 1024;
const int blocks = (numel + threads - 1) / threads;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "selu_forward_coalesced_cuda", ([&] {
const scalar_t* input_ptr = input.data_ptr<scalar_t>();
scalar_t* output_ptr = output.data_ptr<scalar_t>();
selu_kernel_coalesced<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 Coalesced Memory Access (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.988 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.386 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 27.158 | % | 0.090 | 5 |
| Issued Ipc Active | 1.086 | inst/cycle | 0.000 | 5 |
| SM Busy | 27.158 | % | 0.090 | 5 |
| Memory Throughput | 277177497966.442 | byte/second | 40019172472106008576.000 | 5 |
| Mem Busy | 13.214 | % | 0.085 | 5 |
| Max Bandwidth | 12.190 | % | 0.073 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 66.974 | % | 0.126 | 5 |
| Mem Pipes Busy | 5.236 | % | 0.016 | 5 |
| Warp Cycles Per Issued Instruction | 47.438 | cycle | 0.377 | 5 |
| Warp Cycles Per Executed Instruction | 52.140 | cycle | 0.456 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.540 | 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 | 4.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 8.000 | block | 0.000 | 5 |
| Block Limit Warps | 2.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 80.660 | % | 0.315 | 5 |
| Achieved Active Warps Per SM | 51.622 | warp | 0.128 | 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 (80.4%) 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 | 627128.97 | μs |
| Device Time | 40.06 | μs |
| Self CPU Time | 33.77 | 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::_to_copy | ||
| CPU Time | 627095.20 | μs |
| Device Time | 40.06 | μs |
| Self CPU Time | 84.49 | 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 | 647421.96 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 20746.38 | 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 | 625780.29 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 625780.29 | 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 | 489464.00 | μs |
| Device Time | 22723.32 | μs |
| Self CPU Time | 489464.00 | μs |
| Self Device Time | 22723.32 | 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_coalesced<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 33047.36 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 33047.36 | 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 | 34455.78 | μs |
| Device Time | 43783.00 | μs |
| Self CPU Time | 34455.78 | μs |
| Self Device Time | 43783.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::zero_ | ||
| CPU Time | 64956.49 | μs |
| Device Time | 648403.08 | μs |
| Self CPU Time | 14047.75 | 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 | 50909.92 | μs |
| Device Time | 648403.08 | μs |
| Self CPU Time | 16285.57 | μs |
| Self Device Time | 648403.08 | 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 | 648403.08 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 648403.08 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45279 warnings generated when compiling for host. Suppressed 45321 warnings (45274 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.