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;
// Kernel optimized by minimizing __syncthreads() usage
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;
using Vec4T = float4;
const Vec4T* input_vec = reinterpret_cast<const Vec4T*>(input + block_offset);
Vec4T* output_vec = reinterpret_cast<Vec4T*>(output + block_offset);
if (block_offset + tid * 4 + 3 < size) {
Vec4T in_vec = input_vec[tid];
shared_data[tid * 4] = in_vec.x;
shared_data[tid * 4 + 1] = in_vec.y;
shared_data[tid * 4 + 2] = in_vec.z;
shared_data[tid * 4 + 3] = in_vec.w;
} else {
#pragma unroll
for (int i = 0; i < 4; i++) {
int idx = block_offset + tid * 4 + i;
if (idx < size) {
shared_data[tid * 4 + i] = static_cast<float>(input[idx]);
}
}
}
// Synchronize only after loading data into shared memory
__syncthreads();
#pragma unroll
for (int i = 0; i < 4; i++) {
const int idx = block_offset + tid * 4 + i;
if (idx < size) {
float val = -shared_data[tid * 4 + i];
float exp_val = __expf(val);
float r = __fdividef(1.0f, (1.0f + exp_val));
shared_data[tid * 4 + i] = r;
}
}
// Synchronize only if data is needed by other threads
__syncthreads();
if (block_offset + tid * 4 + 3 < size) {
Vec4T out_vec;
out_vec.x = shared_data[tid * 4];
out_vec.y = shared_data[tid * 4 + 1];
out_vec.z = shared_data[tid * 4 + 2];
out_vec.w = shared_data[tid * 4 + 3];
output_vec[tid] = out_vec;
} else {
#pragma unroll
for (int i = 0; i < 4; i++) {
int idx = block_offset + tid * 4 + i;
if (idx < size) {
output[idx] = static_cast<scalar_t>(shared_data[tid * 4 + i]);
}
}
}
}
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, "Optimized Sigmoid forward (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.482 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.200 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 12.694 | % | 0.075 | 5 |
| Issued Ipc Active | 0.508 | inst/cycle | 0.000 | 5 |
| SM Busy | 12.694 | % | 0.075 | 5 |
| Memory Throughput | 281184284693.810 | byte/second | 2310912917160634880.000 | 5 |
| Mem Busy | 13.300 | % | 0.011 | 5 |
| Max Bandwidth | 12.294 | % | 0.003 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.106 | % | 0.056 | 5 |
| Mem Pipes Busy | 4.686 | % | 0.001 | 5 |
| Warp Cycles Per Issued Instruction | 28.618 | cycle | 0.138 | 5 |
| Warp Cycles Per Executed Instruction | 30.124 | cycle | 0.150 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 32.000 | 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 | 10.000 | 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 | 23.104 | % | 0.001 | 5 |
| Achieved Active Warps Per SM | 14.790 | warp | 0.000 | 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 (23.1%) 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 | 234113.69 | μs |
| Device Time | 39.94 | μs |
| Self CPU Time | 37.56 | 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 | 234076.13 | μs |
| Device Time | 39.94 | μs |
| Self CPU Time | 82.83 | 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 | 253023.89 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 19388.69 | 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 | 233450.54 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 233450.54 | 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 | 496471.05 | μs |
| Device Time | 22182.02 | μs |
| Self CPU Time | 496471.05 | μs |
| Self Device Time | 22182.02 | 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 | 31239.57 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 31239.57 | 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 | 19938.90 | μs |
| Device Time | 42801.13 | μs |
| Self CPU Time | 19938.90 | μs |
| Self Device Time | 42801.13 | 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 | 63797.97 | μs |
| Device Time | 634067.95 | μs |
| Self CPU Time | 14060.89 | 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 | 49738.76 | μs |
| Device Time | 634067.95 | μs |
| Self CPU Time | 15607.02 | μs |
| Self Device Time | 634067.95 | 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 | 634067.95 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 634067.95 | 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 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.