import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
return F.hardsigmoid(x)
class Model(nn.Module):
"""
Simple model that performs a HardSigmoid activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
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 HardSigmoid activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
return torch.nn.functional.hardsigmoid(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>
// CUDA kernel that distributes workloads evenly by assigning each thread a contiguous chunk of data
template <typename scalar_t>
__global__ void even_chunk_hardsigmoid_kernel(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
// Compute global thread id and total number of threads
int tid = blockIdx.x * blockDim.x + threadIdx.x;
int total_threads = gridDim.x * blockDim.x;
// Calculate the number of elements each thread should process (ceiling division)
size_t items_per_thread = (numel + total_threads - 1) / total_threads;
// Determine the contiguous block of indices this thread will handle
size_t start = tid * items_per_thread;
size_t end = start + items_per_thread;
if (end > numel) end = numel;
// Process each element in the assigned contiguous chunk
for (size_t i = start; i < end; i++) {
scalar_t x = input[i];
scalar_t y = (x + static_cast<scalar_t>(3)) / static_cast<scalar_t>(6);
// Clamp y to the range [0, 1]
if (y < static_cast<scalar_t>(0))
y = static_cast<scalar_t>(0);
else if (y > static_cast<scalar_t>(1))
y = static_cast<scalar_t>(1);
output[i] = y;
}
}
// Host function that dispatches the kernel
torch::Tensor forward(torch::Tensor input) {
TORCH_CHECK(input.is_cuda(), "Input tensor must be on CUDA");
auto output = torch::empty_like(input);
size_t numel = input.numel();
// Configure kernel launch parameters
// Using 1024 threads per block; blocks is computed to cover all elements
const int threads = 1024;
const int blocks = (numel + threads - 1) / threads;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "even_chunk_hardsigmoid_cuda", ([&] {
even_chunk_hardsigmoid_kernel<scalar_t><<<blocks, threads>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
numel);
}));
cudaError_t err = cudaGetLastError();
TORCH_CHECK(err == cudaSuccess, "CUDA kernel failed: ", cudaGetErrorString(err));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "HardSigmoid activation forward (CUDA) with even workload chunk distribution");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.678 | inst/cycle | 0.004 | 5 |
| Executed Ipc Elapsed | 0.814 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 42.702 | % | 2.845 | 5 |
| Issued Ipc Active | 1.706 | inst/cycle | 0.005 | 5 |
| SM Busy | 42.702 | % | 2.845 | 5 |
| Memory Throughput | 242270239803.094 | byte/second | 10870226859294586880.000 | 5 |
| Mem Busy | 11.440 | % | 0.033 | 5 |
| Max Bandwidth | 10.696 | % | 0.027 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 68.230 | % | 0.090 | 5 |
| Mem Pipes Busy | 9.106 | % | 0.018 | 5 |
| Warp Cycles Per Issued Instruction | 29.932 | cycle | 0.005 | 5 |
| Warp Cycles Per Executed Instruction | 30.474 | cycle | 0.005 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.510 | 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 | 2.000 | 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 | 82.204 | % | 0.223 | 5 |
| Achieved Active Warps Per SM | 52.608 | warp | 0.090 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | ALU is the highest-utilized pipeline (41.8%) based on active cycles, taking into account the rates of its different instructions. It executes integer and logic operations. It is well-utilized, but should not be a bottleneck. |
| 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 (81.9%) 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 | 600157.56 | μs |
| Device Time | 40.38 | μs |
| Self CPU Time | 35.12 | 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 | 600122.45 | μs |
| Device Time | 40.38 | μs |
| Self CPU Time | 85.08 | 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 | 619962.98 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 20286.43 | 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 | 592692.57 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 592692.57 | 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 |
| cudaLaunchKernel | ||
| CPU Time | 514815.51 | μs |
| Device Time | 627.71 | μs |
| Self CPU Time | 514815.51 | μs |
| Self Device Time | 627.71 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void even_chunk_hardsigmoid_kernel<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 30274.26 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 30274.26 | 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 | 23761.33 | μs |
| Device Time | 43162.96 | μs |
| Self CPU Time | 23761.33 | μs |
| Self Device Time | 43162.96 | 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 | 68836.84 | μs |
| Device Time | 661182.26 | μs |
| Self CPU Time | 12354.60 | 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 | 56484.04 | μs |
| Device Time | 661182.26 | μs |
| Self CPU Time | 15853.88 | μs |
| Self Device Time | 661182.26 | 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 | 661261.78 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 661261.78 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45280 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.