← Back to Leaderboard

The AI CUDA Engineer 👷

53_Min_reduction_over_a_dimension • min_reduction_warp_shared_hybrid_base

Level 1 • Task 53
import torch
import torch.nn as nn
import torch.functional as F


def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
    """
    Applies min reduction over the specified dimension to the input tensor.

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

    Returns:
        torch.Tensor: Output tensor after min reduction over the specified dimension
    """
    return torch.min(x, dim)[0]


class Model(nn.Module):
    """
    Simple model that performs min 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:
        """
        Applies min reduction over the specified dimension to the input tensor.

        Args:
            x (torch.Tensor): Input tensor
            fn: Function to apply (defaults to module_fn)

        Returns:
            torch.Tensor: Output tensor after min reduction over the specified dimension
        """
        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]  # Example, change to desired dimension

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs min 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:
        """
        Applies min reduction over the specified dimension to the input tensor.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Output tensor after min reduction over the specified dimension.
        """
        return torch.min(x, dim=self.dim)[0]

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] # Example, change to desired dimension
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 53_Min_reduction_over_a_dimension
Level ID 1
Task ID 53
Kernel Name min_reduction_warp_shared_hybrid_base
CUDA Speedup (Native) 1.718x
CUDA Speedup (Compile) 2.429x
CUDA Runtime 0.014 ms
PyTorch Runtime (Native) 0.024 ms
PyTorch Runtime (Compile) 0.034 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
Temperature 0.50
View Experiment Progress Details

Related Kernels (Level 1, Task 53 • 53_Min_reduction_over_a_dimension)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 min_reduction_warp_base_base 0.01 2.19 3.09
🥇 efficient_min_reduce_kernel_base 0.01 2.19 3.09
🥇 min_reduce_block_size_tuning_base 0.01 2.19 3.09
🥇 min_reduce_optimized_base 0.01 2.19 3.09
🥇 min_reduction_optimized_memory_base 0.01 2.19 3.09
🥇 min_reduce_tunable_blocksize_base 0.01 2.19 3.09
🥇 min_reduce_dynamic_block_base_base 0.01 2.19 3.09
🥇 modular_min_reduce_kernel_base_base 0.01 2.19 3.09
🥇 min_reduce_adaptive_blocks_base_base 0.01 2.19 3.09
🥇 min_reduce_fused_warp_base 0.01 2.19 3.09
🥇 min_reduce_combined_base 0.01 2.19 3.09
🥇 min_reduce_warp_unroll_base 0.01 2.19 3.09
🥇 min_reduce_combined_kernel_base 0.01 2.19 3.09
14 balanced_min_reduction_base 0.01 1.85 2.62
15 min_reduction_warp_shared_hybrid_base 0.01 1.72 2.43
15 optimized_block_size_experiment_base_base 0.01 1.72 2.43
15 min_reduction_shared_base 0.01 1.72 2.43
15 modular_min_reduction_base 0.01 1.72 2.43
15 fast_min_reduction_edit_1 0.01 1.72 2.43
15 vector_load_min_reduction_edit_1 0.01 1.72 2.43 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <vector>
#include <limits>
#include <c10/cuda/CUDAStream.h>

template <typename scalar_t>
__global__ void min_reduce_warp_shared_kernel(
    const scalar_t* __restrict__ input,
    scalar_t* __restrict__ output,
    const int outer,
    const int r,
    const int inner) {
    
    extern __shared__ char shared_mem[];
    scalar_t* shared_data = reinterpret_cast<scalar_t*>(shared_mem);
    
    const int warp_size = 32;
    const int warps_per_block = blockDim.x / warp_size;
    const int warp_id = threadIdx.x / warp_size;
    const int lane_id = threadIdx.x % warp_size;
    const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
    
    if (global_warp_id >= outer * inner) return;
    
    const int outer_idx = global_warp_id / inner;
    const int inner_idx = global_warp_id % inner;
    const int base = outer_idx * (r * inner) + inner_idx;
    
    const int tile_size = 128;
    scalar_t warp_min = std::numeric_limits<scalar_t>::max();
    
    #pragma unroll 1
    for (int tile_start = 0; tile_start < r; tile_start += tile_size) {
        const int tile_end = min(tile_start + tile_size, r);
        
        for (int j = tile_start + lane_id; j < tile_end; j += warp_size) {
            const int shared_idx = (j - tile_start) + warp_id * tile_size;
            const int global_idx = base + j * inner;
            shared_data[shared_idx] = input[global_idx];
        }
        __syncthreads();
        
        for (int j = 0; j < (tile_end - tile_start); j++) {
            const int shared_idx = j + warp_id * tile_size;
            scalar_t val = shared_data[shared_idx];
            warp_min = min(warp_min, val);
        }
        __syncthreads();
    }
    
    #pragma unroll
    for (int offset = 16; offset > 0; offset /= 2) {
        scalar_t other = __shfl_down_sync(0xffffffff, warp_min, offset);
        warp_min = min(warp_min, other);
    }
    
    if (lane_id == 0) {
        output[global_warp_id] = warp_min;
    }
}

torch::Tensor forward(torch::Tensor input, int64_t dim) {
    TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor");
    if (!input.is_contiguous()) {
        input = input.contiguous();
    }

    int ndim = input.dim();
    TORCH_CHECK(dim >= 0 && dim < ndim, "dim out of range");

    int outer = 1;
    for (int i = 0; i < dim; i++) {
        outer *= input.size(i);
    }
    int r = input.size(dim);
    int inner = 1;
    for (int i = dim + 1; i < ndim; i++) {
        inner *= input.size(i);
    }

    std::vector<int64_t> output_shape;
    for (int i = 0; i < ndim; i++) {
        if (i != dim) {
            output_shape.push_back(input.size(i));
        }
    }
    
    auto output = torch::empty(output_shape, input.options());

    const int threads_per_block = 128;
    const int warps_per_block = threads_per_block / 32;
    const int num_warps = outer * inner;
    const int num_blocks = (num_warps + warps_per_block - 1) / warps_per_block;
    
    const int shared_mem_size = 128 * warps_per_block * sizeof(float);

    AT_DISPATCH_ALL_TYPES(input.scalar_type(), "min_reduce_warp_shared_cuda", ([&] {
        min_reduce_warp_shared_kernel<scalar_t><<<num_blocks, threads_per_block, shared_mem_size,
            c10::cuda::getCurrentCUDAStream().stream()>>>(
            input.data_ptr<scalar_t>(),
            output.data_ptr<scalar_t>(),
            outer,
            r,
            inner);
    }));

    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Min reduction over a specified dimension using warp-level primitives with shared memory (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.270 inst/cycle 0.000 5
Executed Ipc Elapsed 1.008 inst/cycle 0.000 5
Issue Slots Busy 32.186 % 0.014 5
Issued Ipc Active 1.286 inst/cycle 0.000 5
SM Busy 39.446 % 0.021 5
Memory Throughput 333111371020.562 byte/second 5838846536938113024.000 5
Mem Busy 62.464 % 0.163 5
Max Bandwidth 30.940 % 0.039 5
L1/TEX Hit Rate 74.940 % 0.000 5
L2 Hit Rate 63.590 % 0.189 5
Mem Pipes Busy 13.546 % 0.007 5
Warp Cycles Per Issued Instruction 22.406 cycle 0.020 5
Warp Cycles Per Executed Instruction 22.696 cycle 0.020 5
Avg. Active Threads Per Warp 31.760 0.000 5
Avg. Not Predicated Off Threads Per Warp 30.570 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 33.000 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 45.164 % 0.001 5
Achieved Active Warps Per SM 28.902 warp 0.000 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (39.5%) 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 (45.1%) 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 562175.24 μs
Device Time 373.12 μs
Self CPU Time 36.64 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 562138.60 μs
Device Time 373.12 μs
Self CPU Time 100.28 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 561405.78 μs
Device Time 0.00 μs
Self CPU Time 80.29 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 546401.03 μs
Device Time 0.00 μs
Self CPU Time 546401.03 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 550234.81 μs
Device Time 705.27 μs
Self CPU Time 550234.81 μs
Self Device Time 705.27 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void min_reduce_warp_shared_kernel<float>(float const*, float*, int, int, int)
CPU Time 0.00 μs
Device Time 81564.02 μs
Self CPU Time 0.00 μs
Self Device Time 81564.02 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 69204.45 μs
Device Time 639469.84 μs
Self CPU Time 14642.95 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 54566.03 μs
Device Time 639469.84 μs
Self CPU Time 15600.34 μs
Self Device Time 639469.84 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 639469.84 μs
Self CPU Time 0.00 μs
Self Device Time 639469.84 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 
45299 warnings generated when compiling for host.
Suppressed 45327 warnings (45280 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/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:12:5 bugprone-easily-swappable-parameters
12 | const int outer,
| ^~~~~~~~~~~~~~~~
13 | const int r,
| ~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:12:15: note: the first parameter in the range is 'outer'
12 | const int outer,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:13:15: note: the last parameter in the range is 'r'
13 | const int r,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:20:33: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
20 | const int warps_per_block = blockDim.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:21:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
21 | const int warp_id = threadIdx.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:22:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
22 | const int lane_id = threadIdx.x % warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:23:32: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
23 | const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:70:16: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
70 | int ndim = input.dim();
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:75:18: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
75 | outer *= input.size(i);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:77:13: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
77 | int r = input.size(dim);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:79:18: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
79 | for (int i = dim + 1; i < ndim; i++) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:80:18: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
80 | inner *= input.size(i);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:97:33: warning: performing an implicit widening conversion to type 'unsigned long' of a multiplication performed in type 'int' [bugprone-implicit-widening-of-multiplication-result]
97 | const int shared_mem_size = 128 * warps_per_block * sizeof(float);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:97:33: note: make conversion explicit to silence this warning
4 | const int shared_mem_size = 128 * warps_per_block * sizeof(float);
| ^~~~~~~~~~~~~~~~~~~~~
| static_cast<unsigned long>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:97:33: note: perform multiplication in a wider type
97 | const int shared_mem_size = 128 * warps_per_block * sizeof(float);
| ^~~
| static_cast<long>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_53/b3_s2_min_reduction_warp_shared_hybrid/base/base.cu:99:5: 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]
99 | AT_DISPATCH_ALL_TYPES(input.scalar_type(), "min_reduce_warp_shared_cuda", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:482:34: note: expanded from macro 'AT_DISPATCH_ALL_TYPES'
482 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:478:3: note: expanded from macro 'AT_DISPATCH_CASE_ALL_TYPES'
478 | AT_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__) \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:458:3: note: expanded from macro 'AT_DISPATCH_CASE_INTEGRAL_TYPES'
458 | AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__) \
| ^
note: (skipping 2 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__, \
| ^