import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies Sigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Sigmoid applied, same shape as input.
"""
return torch.sigmoid(x)
class Model(nn.Module):
"""
Simple model that performs a Sigmoid 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 Sigmoid activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies Sigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Sigmoid applied, same shape as input.
"""
return torch.sigmoid(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>
const int THREADS = 256;
const int ELEMENTS_PER_THREAD = 4;
const int SHARED_MEM_SIZE = THREADS * ELEMENTS_PER_THREAD;
template <typename scalar_t>
__forceinline__ __device__ float sigmoid_compute(float x) {
return 1.0f / (1.0f + expf(-x));
}
template <typename scalar_t>
__global__ void sigmoid_kernel(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int64_t size) {
__shared__ float shared_data[SHARED_MEM_SIZE];
const int tid = threadIdx.x;
const int block_offset = blockIdx.x * SHARED_MEM_SIZE;
// Vectorized loading into shared memory
using float4_t = float4;
if (tid * ELEMENTS_PER_THREAD + ELEMENTS_PER_THREAD <= size) {
float4_t* in_ptr = (float4_t*)&input[block_offset + tid * ELEMENTS_PER_THREAD];
float4_t* shared_ptr = (float4_t*)&shared_data[tid * ELEMENTS_PER_THREAD];
*shared_ptr = *in_ptr;
} else {
#pragma unroll
for (int i = 0; i < ELEMENTS_PER_THREAD; i++) {
const int idx = block_offset + tid + i * THREADS;
if (idx < size) {
shared_data[tid + i * THREADS] = static_cast<float>(input[idx]);
}
}
}
__syncthreads();
// Process elements using optimized sigmoid computation
#pragma unroll
for (int i = 0; i < ELEMENTS_PER_THREAD; i++) {
const int idx = block_offset + tid + i * THREADS;
if (idx < size) {
float val = shared_data[tid + i * THREADS];
output[idx] = static_cast<scalar_t>(sigmoid_compute<scalar_t>(val));
}
}
}
torch::Tensor forward(torch::Tensor input) {
auto output = torch::empty_like(input);
const int64_t size = input.numel();
const int blocks = (size + SHARED_MEM_SIZE - 1) / SHARED_MEM_SIZE;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "sigmoid_kernel", [&] {
const auto* input_data = input.data_ptr<scalar_t>();
auto* output_data = output.data_ptr<scalar_t>();
sigmoid_kernel<scalar_t><<<blocks, THREADS>>>(input_data, output_data, size);
});
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "Sigmoid forward (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.494 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.186 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 12.844 | % | 0.011 | 5 |
| Issued Ipc Active | 0.514 | inst/cycle | 0.000 | 5 |
| SM Busy | 12.844 | % | 0.011 | 5 |
| Memory Throughput | 295787805916.480 | byte/second | 9551212054040031232.000 | 5 |
| Mem Busy | 14.066 | % | 0.063 | 5 |
| Max Bandwidth | 12.932 | % | 0.032 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.060 | % | 0.096 | 5 |
| Mem Pipes Busy | 4.396 | % | 0.003 | 5 |
| Warp Cycles Per Issued Instruction | 28.344 | cycle | 0.191 | 5 |
| Warp Cycles Per Executed Instruction | 29.612 | cycle | 0.207 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 31.520 | 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 | 20.000 | 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 | 22.834 | % | 0.002 | 5 |
| Achieved Active Warps Per SM | 14.612 | warp | 0.001 | 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 (22.8%) 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 | 605607.89 | μs |
| Device Time | 40.26 | μs |
| Self CPU Time | 39.18 | 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 | 605568.71 | μs |
| Device Time | 40.26 | μs |
| Self CPU Time | 96.14 | 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 | 622196.18 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 17081.69 | 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 | 603111.21 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 603111.21 | 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 | 397946.86 | μs |
| Device Time | 19184.66 | μs |
| Self CPU Time | 397946.86 | μs |
| Self Device Time | 19184.66 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void sigmoid_kernel<float>(float const*, float*, long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 19742.39 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 19742.39 | 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 | 19914.20 | μs |
| Device Time | 36949.88 | μs |
| Self CPU Time | 19914.20 | μs |
| Self Device Time | 36949.88 | 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 | 62766.87 | μs |
| Device Time | 548345.29 | μs |
| Self CPU Time | 12187.25 | 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 | 50581.76 | μs |
| Device Time | 548345.29 | μs |
| Self CPU Time | 15302.85 | μs |
| Self Device Time | 548345.29 | 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 | 548345.29 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 548345.29 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45281 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.