import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies Softplus activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Softplus applied, same shape as input.
"""
return F.softplus(x)
class Model(nn.Module):
"""
Simple model that performs a Softplus 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 Softplus activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies Softplus activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Softplus applied, same shape as input.
"""
return torch.nn.functional.softplus(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 <type_traits>
// Threshold constants in constant memory
__constant__ float c_upper_threshold_float = 20.0f;
__constant__ float c_lower_threshold_float = -20.0f;
__constant__ double c_upper_threshold_double = 20.0;
__constant__ double c_lower_threshold_double = -20.0;
template <typename scalar_t>
__device__ __forceinline__ scalar_t compute_softplus(const scalar_t x) {
if constexpr (std::is_same<scalar_t, float>::value) {
if (x > c_upper_threshold_float) return x;
if (x < c_lower_threshold_float) return __expf(x);
return log1pf(__expf(x));
} else {
if (x > c_upper_threshold_double) return x;
if (x < c_lower_threshold_double) return exp(x);
return log1p(exp(x));
}
}
template <typename scalar_t>
__global__ void softplus_kernel_optimized(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int size) {
const int tid = threadIdx.x;
int idx = blockIdx.x * blockDim.x * 4 + tid;
const scalar_t* input_block = input + blockIdx.x * blockDim.x * 4;
scalar_t* output_block = output + blockIdx.x * blockDim.x * 4;
// Use shared memory for coalesced memory access
__shared__ scalar_t shared_input[1024];
// Load data into shared memory
#pragma unroll
for (int i = 0; i < 4; i++) {
if (idx + i * blockDim.x < size) {
shared_input[tid + i * blockDim.x] = input_block[tid + i * blockDim.x];
}
}
__syncthreads();
// Process data from shared memory
#pragma unroll
for (int i = 0; i < 4; i++) {
if (idx + i * blockDim.x < size) {
output_block[tid + i * blockDim.x] = compute_softplus(shared_input[tid + i * blockDim.x]);
}
}
}
torch::Tensor softplus_cuda_forward(torch::Tensor input) {
auto output = torch::empty_like(input);
const int size = input.numel();
const int threads = 256;
const int blocks = (size + threads * 4 - 1) / (threads * 4);
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "softplus_forward_cuda", ([&] {
softplus_kernel_optimized<scalar_t><<<blocks, threads>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
size);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &softplus_cuda_forward, "Softplus forward (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.206 | inst/cycle | 0.001 | 5 |
| Executed Ipc Elapsed | 0.564 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 30.648 | % | 1.052 | 5 |
| Issued Ipc Active | 1.224 | inst/cycle | 0.002 | 5 |
| SM Busy | 30.648 | % | 1.052 | 5 |
| Memory Throughput | 257268295374.282 | byte/second | 17749528052329951232.000 | 5 |
| Mem Busy | 12.188 | % | 0.039 | 5 |
| Max Bandwidth | 11.306 | % | 0.040 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.380 | % | 0.095 | 5 |
| Mem Pipes Busy | 5.988 | % | 0.012 | 5 |
| Warp Cycles Per Issued Instruction | 12.076 | cycle | 0.005 | 5 |
| Warp Cycles Per Executed Instruction | 12.280 | cycle | 0.005 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 30.360 | 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 | 23.586 | % | 0.001 | 5 |
| Achieved Active Warps Per SM | 15.096 | 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. |
| 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 (23.6%) 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 | 2190813.28 | μs |
| Device Time | 40.38 | μs |
| Self CPU Time | 40.58 | 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 | 2190772.69 | μs |
| Device Time | 40.38 | μs |
| Self CPU Time | 103.96 | 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 | 2190950.78 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 670.99 | 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 | 2166599.69 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 2166599.69 | 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 |
| cudaLaunchKernel | ||
| CPU Time | 234556.11 | μs |
| Device Time | 314.49 | μs |
| Self CPU Time | 234556.11 | μs |
| Self Device Time | 314.49 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void softplus_kernel_optimized<float>(float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 751.75 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 751.75 | 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 | 631.52 | μs |
| Device Time | 472.57 | μs |
| Self CPU Time | 631.52 | μs |
| Self Device Time | 472.57 | 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 | 222054.44 | μs |
| Device Time | 18154.37 | μs |
| Self CPU Time | 373.60 | 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 | 221682.56 | μs |
| Device Time | 18154.37 | μs |
| Self CPU Time | 503.31 | μs |
| Self Device Time | 18154.37 | 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 | 18154.37 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 18154.37 | 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.