← Back to Leaderboard

The AI CUDA Engineer 👷

45_Average_Pooling_2D • warp_divergence_avg_pool2d_base

Level 1 • Task 45
import 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
) -> torch.Tensor:
    """
    Applies 2D Average Pooling using functional interface.

    Args:
        x (torch.Tensor): Input tensor
        kernel_size (int): Size of pooling window
        stride (int): Stride of pooling operation
        padding (int): Input padding

    Returns:
        torch.Tensor: Output tensor with 2D Average Pooling applied
    """
    return F.avg_pool2d(x, kernel_size=kernel_size, stride=stride, padding=padding)


class Model(nn.Module):
    """
    Simple model that performs 2D Average Pooling.
    """

    def __init__(self, kernel_size: int, stride: int, padding: int):
        """
        Initializes the Average Pooling layer.

        Args:
            kernel_size (int): Size of the pooling window
            stride (int): Stride of the pooling operation
            padding (int): Padding applied to input tensor
        """
        super(Model, self).__init__()
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding

    def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
        """
        Applies 2D Average Pooling to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, channels, height, width)
            fn: Function to apply pooling operation, defaults to module_fn

        Returns:
            torch.Tensor: Output tensor with Average Pooling applied
        """
        return fn(
            x,
            self.kernel_size,
            self.stride,
            self.padding,
        )


batch_size = 16
channels = 64
height = 256
width = 256
kernel_size = 3
stride = None  # Defaults to kernel_size
padding = 0


def get_inputs():
    x = torch.randn(batch_size, channels, height, width)
    return [x]


def get_init_inputs():
    return [kernel_size, stride if stride is not None else kernel_size, padding]

import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Simple model that performs 2D Average Pooling.
    """

    def __init__(self, kernel_size: int, stride: int = None, padding: int = 0):
        """
        Initializes the Average Pooling layer.

        Args:
            kernel_size (int): Size of the pooling window.
            stride (int, optional): Stride of the pooling operation. Defaults to None (same as kernel_size).
            padding (int, optional): Padding applied to the input tensor. Defaults to 0.
        """
        super(Model, self).__init__()
        self.avg_pool = nn.AvgPool2d(
            kernel_size=kernel_size, stride=stride, padding=padding
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies 2D Average Pooling to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, channels, height, width).

        Returns:
            torch.Tensor: Output tensor with Average Pooling applied.
        """
        return self.avg_pool(x)


batch_size = 16
channels = 64
height = 256
width = 256
kernel_size = 3
stride = None  # Defaults to kernel_size
padding = 0


def get_inputs():
    x = torch.randn(batch_size, channels, height, width)
    return [x]


def get_init_inputs():
    return [kernel_size, stride if stride is not None else kernel_size, padding]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 45_Average_Pooling_2D
Level ID 1
Task ID 45
Kernel Name warp_divergence_avg_pool2d_base
CUDA Speedup (Native) 1.818x
CUDA Speedup (Compile) 2.841x
CUDA Runtime 0.115 ms
PyTorch Runtime (Native) 0.209 ms
PyTorch Runtime (Compile) 0.327 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.50
View Experiment Progress Details

Related Kernels (Level 1, Task 45 • 45_Average_Pooling_2D)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 unrolled_avg_pool2d_base 0.11 1.94 3.03
🥇 modular_avg_pool2d_base_base 0.11 1.94 3.03
🥇 grid_stride_manual_unroll_base 0.11 1.94 3.03
🥇 optimized_avg_pool2d_base 0.11 1.94 3.03
🥇 manual_unroll_avg_pool2d_base 0.11 1.94 3.03
🥇 efficient_avg_pool_base 0.11 1.94 3.03
🥇 constant_memory_avg_pool2d_base 0.11 1.94 3.03
🥇 even_workload_avg_pool2d_base 0.11 1.94 3.03
🥇 unrolled_optimized_avg_pool2d_base 0.11 1.94 3.03
10 warp_uniform_flow_base_base 0.11 1.83 2.87
10 optimized_avg_pool2d_base 0.11 1.83 2.87
10 grid_stride_avg_pool2d_base_base 0.11 1.83 2.87
10 warp_uniform_flow_base_edit_1 0.11 1.83 2.87
14 stride_optimized_avg_pool2d_base 0.12 1.82 2.84
14 warp_divergence_avg_pool2d_base 0.12 1.82 2.84
14 stride_loop_avg_pool2d_base 0.12 1.82 2.84
14 grid_unrolled_avg_pool2d_base 0.12 1.82 2.84
18 combined_avg_pool_base 0.12 1.80 2.82
18 spatial_block_optimized_base_base 0.12 1.80 2.82
18 spatial_block_optimized_base_edit_1 0.12 1.80 2.82 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Kernel that minimizes warp divergence by restructuring conditional logic
// for uniform control flow within warps.

template <typename scalar_t>
__global__ void warp_divergence_avg_pool2d_kernel(
    const scalar_t* __restrict__ input,
    scalar_t* __restrict__ output,
    const int N,
    const int C,
    const int H,
    const int W,
    const int outH,
    const int outW,
    const int kernel_size,
    const int stride,
    const int padding
) {
    int out_x = blockIdx.x * blockDim.x + threadIdx.x;
    int out_y = blockIdx.y * blockDim.y + threadIdx.y;
    int nc = blockIdx.z;
    int n = nc / C;
    int c = nc % C;

    if (out_x >= outW || out_y >= outH)
        return;

    int in_x_start = out_x * stride - padding;
    int in_y_start = out_y * stride - padding;
    scalar_t sum = scalar_t(0);

    // Calculate boundaries and clamp to input size
    int h_start = max(0, in_y_start);
    int w_start = max(0, in_x_start);
    int h_end = min(H, in_y_start + kernel_size);
    int w_end = min(W, in_x_start + kernel_size);

    // Use a single path with pre-calculated bounds to avoid warp divergence
    for (int h = h_start; h < h_end; ++h) {
        int row_offset = ((n * C + c) * H + h) * W;
        for (int w = w_start; w < w_end; ++w) {
            sum += input[row_offset + w];
        }
    }

    int out_index = ((n * C + c) * outH + out_y) * outW + out_x;
    output[out_index] = sum / static_cast<scalar_t>(kernel_size * kernel_size);
}

// Forward function exposed to PyTorch

torch::Tensor warp_divergence_avg_pool2d_forward(
    torch::Tensor x,
    int kernel_size,
    int stride,
    int padding
) {
    TORCH_CHECK(x.dim() == 4, "Input must be a 4D tensor.");
    
    const int N = x.size(0);
    const int C = x.size(1);
    const int H = x.size(2);
    const int W = x.size(3);
    const int outH = (H + 2 * padding - kernel_size) / stride + 1;
    const int outW = (W + 2 * padding - kernel_size) / stride + 1;
    
    auto x_cont = x.contiguous();
    auto options = x.options();
    auto output = torch::empty({N, C, outH, outW}, options);
    
    dim3 threads(32, 8);
    dim3 blocks(
        (outW + threads.x - 1) / threads.x,
        (outH + threads.y - 1) / threads.y,
        N * C
    );
    
    AT_DISPATCH_FLOATING_TYPES(x.scalar_type(), "warp_divergence_avg_pool2d_kernel", ([&] {
        warp_divergence_avg_pool2d_kernel<scalar_t><<<blocks, threads>>>(
            x_cont.data_ptr<scalar_t>(),
            output.data_ptr<scalar_t>(),
            N, C, H, W,
            outH, outW,
            kernel_size, stride, padding
        );
    }));
    
    cudaError_t err = cudaGetLastError();
    TORCH_CHECK(err == cudaSuccess, "CUDA Error: ", cudaGetErrorString(err));
    
    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &warp_divergence_avg_pool2d_forward, "Warp Divergence Minimized 2D Average Pooling forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.830 inst/cycle 0.000 5
Executed Ipc Elapsed 1.766 inst/cycle 0.000 5
Issue Slots Busy 45.816 % 0.019 5
Issued Ipc Active 1.832 inst/cycle 0.000 5
SM Busy 45.816 % 0.019 5
Memory Throughput 2714944488251.728 byte/second 47355282895260942336.000 5
Mem Busy 46.562 % 0.013 5
Max Bandwidth 81.016 % 0.042 5
L1/TEX Hit Rate 63.578 % 0.000 5
L2 Hit Rate 12.890 % 0.001 5
Mem Pipes Busy 32.754 % 0.004 5
Warp Cycles Per Issued Instruction 28.454 cycle 0.022 5
Warp Cycles Per Executed Instruction 28.482 cycle 0.023 5
Avg. Active Threads Per Warp 28.660 0.000 5
Avg. Not Predicated Off Threads Per Warp 26.430 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 8.000 block 0.000 5
Block Limit Shared Mem 32.000 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 81.772 % 0.001 5
Achieved Active Warps Per SM 52.332 warp 0.000 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (28.3%) 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.
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 (81.8%) 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 360463.92 μs
Device Time 28452.32 μs
Self CPU Time 48.42 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::_to_copy
CPU Time 360415.50 μs
Device Time 28452.32 μs
Self CPU Time 113.99 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
cudaLaunchKernel
CPU Time 1113016.87 μs
Device Time 51185.71 μs
Self CPU Time 1113016.87 μs
Self Device Time 51185.71 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void warp_divergence_avg_pool2d_kernel<float>(float const*, float*, int, int, int, int, int, int, int, int, int)
CPU Time 0.00 μs
Device Time 763579.79 μs
Self CPU Time 0.00 μs
Self Device Time 763579.79 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 701160.87 μs
Device Time 525621.73 μs
Self CPU Time 12580.84 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 688581.84 μs
Device Time 525621.73 μs
Self CPU Time 16416.16 μs
Self Device Time 525621.73 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 525621.73 μs
Self CPU Time 0.00 μs
Self Device Time 525621.73 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Completed 
45290 warnings generated when compiling for host.
Suppressed 45324 warnings (45277 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.
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:12:5 bugprone-easily-swappable-parameters
12 | const int N,
| ^~~~~~~~~~~~
13 | const int C,
| ~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:12:15: note: the first parameter in the range is 'N'
12 | const int N,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:13:15: note: the last parameter in the range is 'C'
13 | const int C,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:15:5: warning: 2 adjacent parameters of 'warp_divergence_avg_pool2d_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
15 | const int W,
| ^~~~~~~~~~~~
16 | const int outH,
| ~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:15:15: note: the first parameter in the range is 'W'
15 | const int W,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:16:15: note: the last parameter in the range is 'outH'
16 | const int outH,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:17:5: warning: 3 adjacent parameters of 'warp_divergence_avg_pool2d_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
17 | const int outW,
| ^~~~~~~~~~~~~~~
18 | const int kernel_size,
| ~~~~~~~~~~~~~~~~~~~~~~
19 | const int stride,
| ~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:17:15: note: the first parameter in the range is 'outW'
17 | const int outW,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:19:15: note: the last parameter in the range is 'stride'
19 | const int stride,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:22:17: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
22 | int out_x = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:23:17: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
23 | int out_y = blockIdx.y * blockDim.y + threadIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:24:14: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
24 | int nc = blockIdx.z;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:63:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
63 | const int N = x.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:64:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
64 | const int C = x.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:65:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
65 | const int H = x.size(2);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:66:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
66 | const int W = x.size(3);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250201_optimize_b10_s4_e0_sweep/level_1/task_45/b7_s3_warp_divergence_avg_pool2d/base/base.cu:81:5: warning: inside a lambda, '__func__' expands to the name of the function call operator; consider capturing the name of the enclosing function explicitly [bugprone-lambda-function-name]
81 | AT_DISPATCH_FLOATING_TYPES(x.scalar_type(), "warp_divergence_avg_pool2d_kernel", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:237:34: note: expanded from macro 'AT_DISPATCH_FLOATING_TYPES'
237 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:233:3: note: expanded from macro 'AT_DISPATCH_CASE_FLOATING_TYPES'
233 | AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:74:3: note: expanded from macro 'AT_DISPATCH_CASE'
74 | AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, scalar_t, __VA_ARGS__)
| ^
note: (skipping 1 expansions in backtrace; use -fmacro-backtrace-limit=0 to see all)
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:58:7: note: expanded from macro 'AT_PRIVATE_CHECK_SELECTIVE_BUILD'
58 | AT_ERROR( \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:711:32: note: expanded from macro 'AT_ERROR'
711 | C10_EXPAND_MSVC_WORKAROUND(TORCH_CHECK(false, ::c10::str(__VA_ARGS__))); \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:536:9: note: expanded from macro 'TORCH_CHECK'
536 | __func__, \
| ^