41_Max_Pooling_1D • optimized_workload_distributed_pool1d_baseimport torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(
x: torch.Tensor,
kernel_size: int,
stride: int,
padding: int,
dilation: int,
return_indices: bool,
) -> torch.Tensor:
"""
Functional implementation of Max Pooling 1D.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, num_features, sequence_length).
kernel_size (int): Size of the window to take a max over.
stride (int): Stride of the window.
padding (int): Implicit zero padding to be added on both sides.
dilation (int): Spacing between kernel elements.
return_indices (bool): Whether to return the indices of the maximum values.
Returns:
torch.Tensor: Output tensor with Max Pooling 1D applied.
"""
return F.max_pool1d(
x,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
return_indices=return_indices,
)
class Model(nn.Module):
"""
Simple model that performs Max Pooling 1D.
"""
def __init__(
self,
kernel_size: int,
stride: int,
padding: int,
dilation: int,
return_indices: bool,
):
"""
Initializes the Max Pooling 1D layer.
Args:
kernel_size (int): Size of the window to take a max over.
stride (int): Stride of the window.
padding (int): Implicit zero padding to be added on both sides.
dilation (int): Spacing between kernel elements.
return_indices (bool): Whether to return the indices of the maximum values.
"""
super(Model, self).__init__()
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.return_indices = return_indices
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
"""
Applies Max Pooling 1D to the input tensor.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, num_features, sequence_length).
fn: Function to apply (defaults to module_fn)
Returns:
torch.Tensor: Output tensor with Max Pooling 1D applied.
"""
return fn(
x,
self.kernel_size,
self.stride,
self.padding,
self.dilation,
self.return_indices,
)
batch_size = 16
features = 64
sequence_length = 128
kernel_size = 4
stride = 2
padding = 2
dilation = 3
return_indices = False
def get_inputs():
x = torch.randn(batch_size, features, sequence_length)
return [x]
def get_init_inputs():
return [kernel_size, stride, padding, dilation, return_indices]
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs Max Pooling 1D.
"""
def __init__(self, kernel_size: int, stride: int = None, padding: int = 0, dilation: int = 1, return_indices: bool = False):
"""
Initializes the Max Pooling 1D layer.
Args:
kernel_size (int): Size of the window to take a max over.
stride (int, optional): Stride of the window. Defaults to None (same as kernel_size).
padding (int, optional): Implicit zero padding to be added on both sides. Defaults to 0.
dilation (int, optional): Spacing between kernel elements. Defaults to 1.
return_indices (bool, optional): Whether to return the indices of the maximum values. Defaults to False.
"""
super(Model, self).__init__()
self.maxpool = nn.MaxPool1d(kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, return_indices=return_indices)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies Max Pooling 1D to the input tensor.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, num_features, sequence_length).
Returns:
torch.Tensor: Output tensor with Max Pooling 1D applied, shape (batch_size, num_features, output_sequence_length).
"""
return self.maxpool(x)
batch_size = 16
features = 64
sequence_length = 128
kernel_size = 4
stride = 2
padding = 2
dilation = 3
return_indices = False
def get_inputs():
x = torch.randn(batch_size, features, sequence_length)
return [x]
def get_init_inputs():
return [kernel_size, stride, padding, dilation, return_indices]#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
__global__ void optimized_workload_distributed_pool1d_kernel(
const float* __restrict__ input,
float* __restrict__ output,
int64_t* __restrict__ indices,
const int batch_size,
const int num_channels,
const int input_length,
const int kernel_size,
const int stride,
const int padding,
const int dilation,
const int output_length,
bool return_indices)
{
const int tid = threadIdx.x + blockIdx.x * blockDim.x;
const int total_outputs = output_length * num_channels * batch_size;
if (tid >= total_outputs) return;
const int i = tid % output_length;
const int c = (tid / output_length) % num_channels;
const int b = tid / (output_length * num_channels);
const int input_start = i * stride - padding;
const int base_idx = b * num_channels * input_length + c * input_length;
float max_val = -INFINITY;
int max_idx = -1;
// Loop unrolling to improve performance
#pragma unroll
for (int k = 0; k < kernel_size; ++k) {
const int pos = input_start + k * dilation;
if (pos >= 0 && pos < input_length) {
const float val = input[base_idx + pos];
if (val > max_val) {
max_val = val;
max_idx = pos;
}
}
}
const int output_index = b * num_channels * output_length + c * output_length + i;
output[output_index] = max_val;
if (return_indices) indices[output_index] = max_idx;
}
torch::Tensor forward(
torch::Tensor x,
int64_t kernel_size,
int64_t stride,
int64_t padding,
int64_t dilation,
bool return_indices)
{
TORCH_CHECK(x.dim() == 3, "Input must be 3D");
TORCH_CHECK(x.is_cuda(), "Input must be on CUDA");
TORCH_CHECK(x.is_contiguous(), "Input must be contiguous");
const int batch_size = x.size(0);
const int num_channels = x.size(1);
const int input_length = x.size(2);
const int output_length = ((input_length + 2 * padding - dilation * (kernel_size - 1) - 1) / stride) + 1;
TORCH_CHECK(output_length > 0, "Output length must be positive");
auto options = torch::TensorOptions().dtype(x.dtype()).device(x.device());
auto output = torch::empty({batch_size, num_channels, output_length}, options);
torch::Tensor indices;
if (return_indices) {
indices = torch::empty({batch_size, num_channels, output_length},
options.dtype(torch::kInt64));
}
const int total_elements = batch_size * num_channels * output_length;
const int threads_per_block = 512; // Optimized to better utilize GPU capabilities
const int num_blocks = (total_elements + threads_per_block - 1) / threads_per_block;
optimized_workload_distributed_pool1d_kernel<<<num_blocks, threads_per_block>>>(
x.data_ptr<float>(),
output.data_ptr<float>(),
return_indices ? indices.data_ptr<int64_t>() : nullptr,
batch_size,
num_channels,
input_length,
kernel_size,
stride,
padding,
dilation,
output_length,
return_indices
);
return return_indices ? torch::cat({output, indices}, -1) : output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "MaxPool1D forward (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.034 | inst/cycle | 0.001 | 5 |
| Executed Ipc Elapsed | 0.402 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 26.880 | % | 0.773 | 5 |
| Issued Ipc Active | 1.074 | inst/cycle | 0.001 | 5 |
| SM Busy | 26.880 | % | 0.773 | 5 |
| Memory Throughput | 139982328045.906 | byte/second | 24696796611886718976.000 | 5 |
| Mem Busy | 10.594 | % | 0.139 | 5 |
| Max Bandwidth | 6.932 | % | 0.056 | 5 |
| L1/TEX Hit Rate | 69.000 | % | 0.000 | 5 |
| L2 Hit Rate | 72.530 | % | 0.007 | 5 |
| Mem Pipes Busy | 7.070 | % | 0.056 | 5 |
| Warp Cycles Per Issued Instruction | 14.566 | cycle | 0.927 | 5 |
| Warp Cycles Per Executed Instruction | 15.164 | cycle | 1.009 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.340 | 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 | 16.000 | block | 0.000 | 5 |
| Block Limit Warps | 4.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 23.602 | % | 0.003 | 5 |
| Achieved Active Warps Per SM | 15.104 | 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 (23.5%) 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 | 254756.23 | μs |
| Device Time | 21.34 | μs |
| Self CPU Time | 37.39 | 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 | 254718.84 | μs |
| Device Time | 21.34 | μs |
| Self CPU Time | 86.56 | 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::empty_strided | ||
| CPU Time | 254434.40 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 89.63 | 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 | 254046.17 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 254046.17 | 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 | 462228.28 | μs |
| Device Time | 21155.12 | μs |
| Self CPU Time | 462228.28 | μs |
| Self Device Time | 21155.12 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| optimized_workload_distributed_pool1d_kernel(float const*, float*, long*, int, int, int, int, int, int, int, int, bool) | ||
| CPU Time | 0.00 | μs |
| Device Time | 30294.25 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 30294.25 | 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 | 24421.44 | μs |
| Device Time | 41993.32 | μs |
| Self CPU Time | 24421.44 | μs |
| Self Device Time | 41993.32 | 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 | 73032.97 | μs |
| Device Time | 628118.07 | μs |
| Self CPU Time | 15796.42 | 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 | 57238.42 | μs |
| Device Time | 628118.07 | μs |
| Self CPU Time | 19580.82 | μs |
| Self Device Time | 628118.07 | 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 | 628118.07 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 628118.07 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45291 warnings generated when compiling for host. Suppressed 45326 warnings (45279 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.