← Back to Leaderboard

The AI CUDA Engineer 👷

48_Mean_reduction_over_a_dimension • modularized_mean_reduce_base

Level 1 • Task 48
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
    """
    Reduces the input tensor along the specified dimension by taking the mean.

    Args:
        x (torch.Tensor): Input tensor of arbitrary shape.
        dim (int): The dimension to reduce over.

    Returns:
        torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
    """
    return torch.mean(x, dim=dim)


class Model(nn.Module):
    """
    Simple model that performs mean reduction over a specific dimension.
    """

    def __init__(self, dim: int):
        """
        Initializes the model with the dimension to reduce over.

        Args:
            dim (int): The dimension to reduce over.
        """
        super(Model, self).__init__()
        self.dim = dim

    def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
        """
        Reduces the input tensor along the specified dimension by taking the mean.

        Args:
            x (torch.Tensor): Input tensor of arbitrary shape.

        Returns:
            torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
        """
        return fn(x, self.dim)


batch_size = 16
dim1 = 256
dim2 = 256


def get_inputs():
    x = torch.randn(batch_size, dim1, dim2)
    return [x]


def get_init_inputs():
    return [1]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs mean reduction over a specific dimension.
    """
    def __init__(self, dim: int):
        """
        Initializes the model with the dimension to reduce over.

        Args:
            dim (int): The dimension to reduce over.
        """
        super(Model, self).__init__()
        self.dim = dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Reduces the input tensor along the specified dimension by taking the mean.

        Args:
            x (torch.Tensor): Input tensor of arbitrary shape.

        Returns:
            torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
        """
        return torch.mean(x, dim=self.dim)

batch_size = 16
dim1 = 256
dim2 = 256

def get_inputs():
    x = torch.randn(batch_size, dim1, dim2)
    return [x]

def get_init_inputs():
    return [1]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 48_Mean_reduction_over_a_dimension
Level ID 1
Task ID 48
Kernel Name modularized_mean_reduce_base
CUDA Speedup (Native) 1.132x
CUDA Speedup (Compile) 2.329x
CUDA Runtime 0.014 ms
PyTorch Runtime (Native) 0.016 ms
PyTorch Runtime (Compile) 0.033 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model o3-mini-2025-01-31
Temperature 1.00
View Experiment Progress Details

Related Kernels (Level 1, Task 48 • 48_Mean_reduction_over_a_dimension)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 evenly_distributed_mean_base 0.01 1.76 3.62
🥈 modular_warp_reduce_base 0.01 1.32 2.72
🥈 thread_block_optimization_base 0.01 1.32 2.72
🥈 warp_reduce_shfl_base_edit_1 0.01 1.32 2.72
🥈 warp_reduce_shfl_base_base 0.01 1.32 2.72
6 shared_mean_reduction_edit_1 0.01 1.22 2.51
6 shared_mean_reduction_base 0.01 1.22 2.51
8 mean_reduce_unroll_base 0.01 1.13 2.33
8 hybrid_reduce_warp_shared_edit_1 0.01 1.13 2.33
8 mean_reduce_balanced_threads_edit_1 0.01 1.13 2.33
8 hybrid_mean_reduce_base 0.01 1.13 2.33
8 hybrid_reduce_warp_shared_base 0.01 1.13 2.33
8 mean_reduce_warp_base 0.01 1.13 2.33
8 mean_reduce_unroll_optimized_edit_1 0.01 1.13 2.33
8 modularized_mean_reduce_base 0.01 1.13 2.33
8 mean_reduce_opt_base 0.01 1.13 2.33
8 mean_reduce_warp_edit_1 0.01 1.13 2.33
8 mean_reduce_unroll_optimized_base 0.01 1.13 2.33
8 mean_reduce_memory_optimized_base 0.01 1.13 2.33
8 hybrid_warp_block_mean_reduce_base 0.01 1.13 2.33 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Device function to perform block-level reduction using shared memory
template <typename scalar_t>
__device__ __forceinline__ scalar_t block_reduce_sum(scalar_t* sdata, int tid, int blockDim) {
    // Reduce within the block using loop unrolling
    for (int stride = blockDim / 2; stride > 32; stride >>= 1) {
        if (tid < stride) {
            sdata[tid] += sdata[tid + stride];
        }
        __syncthreads();
    }
    if (tid < 32) {
        volatile scalar_t* vsmem = sdata;
        if (blockDim >= 64) vsmem[tid] += vsmem[tid + 32];
        if (blockDim >= 32) vsmem[tid] += vsmem[tid + 16];
        if (blockDim >= 16) vsmem[tid] += vsmem[tid + 8];
        if (blockDim >= 8)  vsmem[tid] += vsmem[tid + 4];
        if (blockDim >= 4)  vsmem[tid] += vsmem[tid + 2];
        if (blockDim >= 2)  vsmem[tid] += vsmem[tid + 1];
    }
    return sdata[0];
}

// CUDA kernel for mean reduction over a specified dimension
template <typename scalar_t>
__global__ void mean_reduce_kernel(
    const scalar_t* __restrict__ input,
    scalar_t* __restrict__ output,
    int64_t outer_size,
    int64_t dim_size,
    int64_t inner_size) {

    extern __shared__ char shared_mem[];
    scalar_t* sdata = reinterpret_cast<scalar_t*>(shared_mem);

    const int tid = threadIdx.x;
    const int out_idx = blockIdx.x;

    // Check if the current block corresponds to a valid output element
    if (out_idx >= outer_size * inner_size)
        return;

    int outer_idx = out_idx / inner_size;
    int inner_idx = out_idx % inner_size;
    int base_idx = outer_idx * dim_size * inner_size + inner_idx;

    // Each thread computes a partial sum over elements in the reduction dimension
    scalar_t sum = 0;
    for (int i = tid; i < dim_size; i += blockDim.x) {
        sum += input[base_idx + i * inner_size];
    }
    sdata[tid] = sum;
    __syncthreads();

    // Use the modular block reduction function to sum the partial results
    scalar_t block_sum = block_reduce_sum(sdata, tid, blockDim.x);
    if (tid == 0) {
        output[out_idx] = block_sum / static_cast<scalar_t>(dim_size);
    }
}

// Host function to prepare and launch the CUDA kernel
torch::Tensor mean_reduce_cuda(torch::Tensor input, int64_t dim) {
    if (dim < 0) dim += input.dim();

    auto sizes = input.sizes().vec();
    int64_t dim_size = sizes[dim];

    int64_t outer_size = 1;
    for (int i = 0; i < dim; i++) {
        outer_size *= sizes[i];
    }

    int64_t inner_size = 1;
    for (size_t i = dim + 1; i < sizes.size(); i++) {
        inner_size *= sizes[i];
    }

    // Remove the reduced dimension from the output shape
    sizes.erase(sizes.begin() + dim);
    auto output = torch::empty(sizes, input.options());

    const int threads = 256;
    const int blocks = outer_size * inner_size;

    AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "mean_reduce_cuda", ([&] {
        const int shared_mem_size = threads * sizeof(scalar_t);
        mean_reduce_kernel<scalar_t><<<blocks, threads, shared_mem_size>>>(
            input.data_ptr<scalar_t>(),
            output.data_ptr<scalar_t>(),
            outer_size,
            dim_size,
            inner_size
        );
    }));

    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &mean_reduce_cuda, "Mean reduction (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.756 inst/cycle 0.001 5
Executed Ipc Elapsed 1.356 inst/cycle 0.000 5
Issue Slots Busy 44.260 % 0.574 5
Issued Ipc Active 1.770 inst/cycle 0.001 5
SM Busy 44.260 % 0.574 5
Memory Throughput 318146021537.816 byte/second 7293715449278270464.000 5
Mem Busy 53.894 % 0.611 5
Max Bandwidth 31.170 % 0.778 5
L1/TEX Hit Rate 6.670 % 0.642 5
L2 Hit Rate 85.536 % 6.547 5
Mem Pipes Busy 19.722 % 0.027 5
Warp Cycles Per Issued Instruction 28.566 cycle 0.037 5
Warp Cycles Per Executed Instruction 28.796 cycle 0.037 5
Avg. Active Threads Per Warp 31.490 0.000 5
Avg. Not Predicated Off Threads Per Warp 26.880 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 16.000 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 77.840 % 0.098 5
Achieved Active Warps Per SM 49.816 warp 0.040 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (26.3%) 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 (78.0%) 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 490288.15 μs
Device Time 360.41 μs
Self CPU Time 37.51 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 490250.64 μs
Device Time 360.41 μs
Self CPU Time 83.10 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 489589.46 μs
Device Time 0.00 μs
Self CPU Time 75.59 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 487484.01 μs
Device Time 0.00 μs
Self CPU Time 487484.01 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 395979.84 μs
Device Time 20501.03 μs
Self CPU Time 395979.84 μs
Self Device Time 20501.03 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void mean_reduce_kernel<float>(float const*, float*, long, long, long)
CPU Time 0.00 μs
Device Time 83391.90 μs
Self CPU Time 0.00 μs
Self Device Time 83391.90 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 25750.89 μs
Device Time 40948.90 μs
Self CPU Time 25750.89 μs
Self Device Time 40948.90 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 95000.11 μs
Device Time 610973.21 μs
Self CPU Time 21369.63 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
aten::fill_
CPU Time 73632.23 μs
Device Time 610973.21 μs
Self CPU Time 27068.86 μs
Self Device Time 610973.21 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 610973.21 μs
Self CPU Time 0.00 μs
Self Device Time 610973.21 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 
45287 warnings generated when compiling for host.
Suppressed 45322 warnings (45275 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/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:7:71 bugprone-easily-swappable-parameters
7 | __device__ __forceinline__ scalar_t block_reduce_sum(scalar_t* sdata, int tid, int blockDim) {
| ^~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:7:75: note: the first parameter in the range is 'tid'
7 | __device__ __forceinline__ scalar_t block_reduce_sum(scalar_t* sdata, int tid, int blockDim) {
| ^~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:7:84: note: the last parameter in the range is 'blockDim'
7 | __device__ __forceinline__ scalar_t block_reduce_sum(scalar_t* sdata, int tid, int blockDim) {
| ^~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:32:5: warning: 2 adjacent parameters of 'mean_reduce_kernel' of similar type ('int64_t') are easily swapped by mistake [bugprone-easily-swappable-parameters]
32 | int64_t outer_size,
| ^~~~~~~~~~~~~~~~~~~
33 | int64_t dim_size,
| ~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:32:13: note: the first parameter in the range is 'outer_size'
32 | int64_t outer_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:33:13: note: the last parameter in the range is 'dim_size'
33 | int64_t dim_size,
| ^~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:39:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
39 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:40:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
40 | const int out_idx = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:46:21: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
46 | int outer_idx = out_idx / inner_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:47:21: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
47 | int inner_idx = out_idx % inner_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:48:20: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
48 | int base_idx = outer_idx * dim_size * inner_size + inner_idx;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:52:42: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
52 | for (int i = tid; i < dim_size; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:87:24: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
87 | const int blocks = outer_size * inner_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250207_optimize_b5_s4_e1_sweep/level_1/task_48/b2_s3_modularized_mean_reduce/base/base.cu:89:5: 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]
89 | AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "mean_reduce_cuda", ([&] {
| ^
/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__, \
| ^