← Back to Leaderboard

The AI CUDA Engineer 👷

45_Average_Pooling_2D • spatial_block_optimized_base_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 spatial_block_optimized_base_base
CUDA Speedup (Native) 1.802x
CUDA Speedup (Compile) 2.817x
CUDA Runtime 0.116 ms
PyTorch Runtime (Native) 0.209 ms
PyTorch Runtime (Compile) 0.327 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model openrouter-deepseek-reasoner
Temperature 1.00
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>

template <typename scalar_t>
__global__ void avg_pool2d_forward_kernel(
    const scalar_t* __restrict__ input,
    scalar_t* __restrict__ output,
    int N, int C, int H, int W,
    int outH, int outW,
    int kernel_size, int stride, int padding
) {
    int nc = blockIdx.z;
    int n = nc / C;
    int c = nc % C;
    
    int h_out = blockIdx.x * blockDim.y + threadIdx.y;
    int w_out = blockIdx.y * blockDim.x + threadIdx.x;
    
    if (n >= N || c >= C || h_out >= outH || w_out >= outW) return;

    int h_start = h_out * stride - padding;
    int w_start = w_out * stride - padding;

    scalar_t sum_val = scalar_t(0);
    for (int i = 0; i < kernel_size; ++i) {
        for (int j = 0; j < kernel_size; ++j) {
            int h_in = h_start + i;
            int w_in = w_start + j;
            if (h_in >= 0 && h_in < H && w_in >= 0 && w_in < W) {
                sum_val += input[((n * C + c) * H + h_in) * W + w_in];
            }
        }
    }
    output[((n * C + c) * outH + h_out) * outW + w_out] = 
        sum_val / static_cast<scalar_t>(kernel_size * kernel_size);
}

torch::Tensor avg_pool2d_forward(
    torch::Tensor x,
    int kernel_size,
    int stride,
    int padding
) {
    TORCH_CHECK(x.dim() == 4, "Input must be a 4D tensor.");
    auto N = x.size(0);
    auto C = x.size(1);
    auto H = x.size(2);
    auto W = x.size(3);

    int outH = (H + 2 * padding - kernel_size)/stride + 1;
    int outW = (W + 2 * padding - kernel_size)/stride + 1;

    auto x_cont = x.contiguous();
    auto out = torch::empty({N, C, outH, outW}, x.options());

    dim3 block(32, 4); // Optimized for spatial access
    dim3 grid(
        (outH + block.y - 1) / block.y,
        (outW + block.x - 1) / block.x,
        N * C
    );

    AT_DISPATCH_FLOATING_TYPES(x.scalar_type(), "avg_pool_forward", ([&] {
        avg_pool2d_forward_kernel<scalar_t><<<grid, block>>>(
            x_cont.data_ptr<scalar_t>(),
            out.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 out;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &avg_pool2d_forward, "2D Average Pooling forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 2.522 inst/cycle 0.000 5
Executed Ipc Elapsed 2.432 inst/cycle 0.000 5
Issue Slots Busy 63.098 % 0.036 5
Issued Ipc Active 2.526 inst/cycle 0.000 5
SM Busy 63.098 % 0.036 5
Memory Throughput 2653791368028.072 byte/second 51862839540401020928.000 5
Mem Busy 45.562 % 0.013 5
Max Bandwidth 79.196 % 0.047 5
L1/TEX Hit Rate 63.404 % 0.000 5
L2 Hit Rate 13.246 % 0.001 5
Mem Pipes Busy 29.904 % 0.040 5
Warp Cycles Per Issued Instruction 21.868 cycle 0.007 5
Warp Cycles Per Executed Instruction 21.878 cycle 0.007 5
Avg. Active Threads Per Warp 29.060 0.000 5
Avg. Not Predicated Off Threads Per Warp 25.380 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 32.000 block 0.000 5
Block Limit Warps 16.000 block 0.000 5
Theoretical Active Warps per SM 64.000 warp 0.000 5
Theoretical Occupancy 100.000 % 0.000 5
Achieved Occupancy 86.766 % 0.001 5
Achieved Active Warps Per SM 55.530 warp 0.000 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (42.3%) 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 (86.8%) 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 314290.93 μs
Device Time 29831.34 μs
Self CPU Time 38.77 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 314252.16 μs
Device Time 29831.34 μs
Self CPU Time 99.34 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 1246392.97 μs
Device Time 57894.44 μs
Self CPU Time 1246392.97 μs
Self Device Time 57894.44 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void avg_pool2d_forward_kernel<float>(float const*, float*, int, int, int, int, int, int, int, int, int)
CPU Time 0.00 μs
Device Time 866275.40 μs
Self CPU Time 0.00 μs
Self Device Time 866275.40 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 791424.11 μs
Device Time 592524.78 μs
Self CPU Time 13398.43 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 778029.29 μs
Device Time 592524.78 μs
Self CPU Time 17020.72 μs
Self Device Time 592524.78 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 592524.78 μs
Self CPU Time 0.00 μs
Self Device Time 592524.78 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 
45289 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/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:9:26 bugprone-easily-swappable-parameters
9 | int N, int C, int H, int W,
| ^~~~~~
10 | int outH, int outW,
| ~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:9:30: note: the first parameter in the range is 'W'
9 | int N, int C, int H, int W,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:10:9: note: the last parameter in the range is 'outH'
10 | int outH, int outW,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:11:5: warning: 2 adjacent parameters of 'avg_pool2d_forward_kernel' of similar type ('int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
11 | int kernel_size, int stride, int padding
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:11:9: note: the first parameter in the range is 'kernel_size'
11 | int kernel_size, int stride, int padding
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:11:26: note: the last parameter in the range is 'stride'
11 | int kernel_size, int stride, int padding
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:13:14: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
13 | int nc = blockIdx.z;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:17:17: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
17 | int h_out = blockIdx.x * blockDim.y + threadIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:18:17: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
18 | int w_out = blockIdx.y * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:51:16: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
51 | int outH = (H + 2 * padding - kernel_size)/stride + 1;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:51:21: warning: performing an implicit widening conversion to type 'int64_t' (aka 'long') of a multiplication performed in type 'int' [bugprone-implicit-widening-of-multiplication-result]
51 | int outH = (H + 2 * padding - kernel_size)/stride + 1;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:51:21: note: make conversion explicit to silence this warning
4 | int outH = (H + 2 * padding - kernel_size)/stride + 1;
| ^~~~~~~~~~~
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:51:21: note: perform multiplication in a wider type
51 | int outH = (H + 2 * padding - kernel_size)/stride + 1;
| ^
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:52:16: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
52 | int outW = (W + 2 * padding - kernel_size)/stride + 1;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:52:21: warning: performing an implicit widening conversion to type 'int64_t' (aka 'long') of a multiplication performed in type 'int' [bugprone-implicit-widening-of-multiplication-result]
52 | int outW = (W + 2 * padding - kernel_size)/stride + 1;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:52:21: note: make conversion explicit to silence this warning
52 | int outW = (W + 2 * padding - kernel_size)/stride + 1;
| ^~~~~~~~~~~
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:52:21: note: perform multiplication in a wider type
52 | int outW = (W + 2 * padding - kernel_size)/stride + 1;
| ^
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_1/task_45/b4_s3_spatial_block_optimized_base/base/base.cu:64:5: 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]
64 | AT_DISPATCH_FLOATING_TYPES(x.scalar_type(), "avg_pool_forward", ([&] {
| ^
/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__, \
| ^