← Back to Leaderboard

The AI CUDA Engineer 👷

100_ConvTranspose3d_Clamp_Min_Divide • balanced_workload_distribution_base

Level 2 • Task 100
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    stride: int,
    padding: int,
    min_value: float,
    divisor: float,
    conv_transpose: torch.Tensor,
    conv_transpose_bias: torch.Tensor,
) -> torch.Tensor:
    """
    Applies a transposed 3D convolution, clamps output to min value, and divides by constant.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_channels, depth, height, width)
        stride (int): Stride of the transposed convolution
        padding (int): Padding of the transposed convolution
        min_value (float): Minimum value for clamping
        divisor (float): Value to divide output by
        conv_transpose (torch.Tensor): Transposed convolution weight tensor
        conv_transpose_bias (torch.Tensor): Bias tensor for transposed convolution

    Returns:
        torch.Tensor: Output tensor after applying transposed convolution, clamping and division
    """
    x = F.conv_transpose3d(
        x, conv_transpose, bias=conv_transpose_bias, stride=stride, padding=padding
    )
    x = torch.clamp(x, min=min_value)
    x = x / divisor
    return x


class Model(nn.Module):
    """
    A model that performs a transposed 3D convolution, clamps the output to a minimum value,
    and then divides the result by a constant.
    """

    def __init__(
        self,
        in_channels,
        out_channels,
        kernel_size,
        stride,
        padding,
        min_value,
        divisor,
    ):
        super(Model, self).__init__()
        conv_transpose = nn.ConvTranspose3d(
            in_channels, out_channels, kernel_size, stride, padding
        )
        self.conv_transpose_parameter = conv_transpose.weight
        self.conv_transpose_bias = conv_transpose.bias

    def forward(self, x, stride, padding, min_value, divisor, fn=module_fn):
        return fn(
            x,
            stride,
            padding,
            min_value,
            divisor,
            self.conv_transpose_parameter,
            self.conv_transpose_bias,
        )


batch_size = 16
in_channels = 32
out_channels = 16
depth, height, width = 16, 32, 32
kernel_size = 3
stride = 2
padding = 1
min_value = -1.0
divisor = 2.0


def get_inputs():
    return [
        torch.randn(batch_size, in_channels, depth, height, width),
        stride,
        padding,
        min_value,
        divisor,
    ]


def get_init_inputs():
    return [in_channels, out_channels, kernel_size, stride, padding, min_value, divisor]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that performs a transposed 3D convolution, clamps the output to a minimum value, 
    and then divides the result by a constant.
    """
    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, min_value, divisor):
        super(Model, self).__init__()
        self.conv_transpose = nn.ConvTranspose3d(in_channels, out_channels, kernel_size, stride=stride, padding=padding)
        self.min_value = min_value
        self.divisor = divisor

    def forward(self, x):
        x = self.conv_transpose(x)
        x = torch.clamp(x, min=self.min_value)
        x = x / self.divisor
        return x

batch_size = 16
in_channels = 32
out_channels = 16
depth, height, width = 16, 32, 32
kernel_size = 3
stride = 2
padding = 1
min_value = -1.0
divisor = 2.0

def get_inputs():
    return [torch.randn(batch_size, in_channels, depth, height, width)]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size, stride, padding, min_value, divisor]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 100_ConvTranspose3d_Clamp_Min_Divide
Level ID 2
Task ID 100
Kernel Name balanced_workload_distribution_base
CUDA Speedup (Native) 1.131x
CUDA Speedup (Compile) 0.839x
CUDA Runtime 0.572 ms
PyTorch Runtime (Native) 0.647 ms
PyTorch Runtime (Compile) 0.480 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.00
View Experiment Progress Details

Related Kernels (Level 2, Task 100 • 100_ConvTranspose3d_Clamp_Min_Divide)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 optimized_clamp_divide_base 0.57 1.13 0.84
🥇 stride_loop_implementation_edit_1 0.57 1.13 0.84
🥇 modular_device_funcs3_base 0.57 1.13 0.84
🥇 balanced_workload_distribution_base 0.57 1.13 0.84
🥇 memory_coalescing_optimization_base 0.57 1.13 0.84
6 unrolled_convtranspose3d_edit_1 0.57 1.13 0.84
6 minimize_warp_divergence_edit_base 0.57 1.13 0.84
6 branchless_clamp_divide_opt_edit_1 0.57 1.13 0.84
6 modular_device_funcs2_edit_1 0.57 1.13 0.84
10 modular_device_funcs2_base 0.57 1.13 0.84
10 optimized_thread_block_indexing_base 0.57 1.13 0.84
10 optimized_thread_block_indexing_edit_1 0.57 1.13 0.84
13 vectorized_ldg_kernel_128_base 0.58 1.12 0.83
13 vectorized_ldg_kernel_128_edit_1 0.58 1.12 0.83
13 vectorized_modular_kernel_edit_1 0.58 1.12 0.83
13 vectorized_stride_loop_edit_1 0.58 1.12 0.83
13 vectorized_stride_loop_base 0.58 1.12 0.83
13 vectorized_modular_kernel_base 0.58 1.12 0.83
19 atomic_free_stride_loop_edit_1 0.58 1.12 0.83
19 modular_conv3d_clamp_divide_base 0.58 1.12 0.83 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

template <typename scalar_t>
__global__ void clamp_and_divide_kernel(
    scalar_t* __restrict__ output,
    const int64_t numel,
    const float min_value,
    const float divisor) {
    
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    const int total_threads = gridDim.x * blockDim.x;
    
    // Distribute workload evenly across threads
    for (int idx = tid; idx < numel; idx += total_threads) {
        output[idx] = max(output[idx], static_cast<scalar_t>(min_value)) / static_cast<scalar_t>(divisor);
    }
}

torch::Tensor forward(
    torch::Tensor input,
    int stride,
    int padding,
    float min_value,
    float divisor,
    torch::Tensor weight,
    torch::Tensor bias) {
    
    auto output = torch::conv_transpose3d(
        input,
        weight,
        bias,
        stride,
        padding
    );

    const int threads = 256;
    const int max_blocks = 65535;
    const int blocks = min(max_blocks, static_cast<int>((output.numel() + threads - 1) / threads));

    AT_DISPATCH_FLOATING_TYPES(output.scalar_type(), "clamp_and_divide", ([&] {
        clamp_and_divide_kernel<scalar_t><<<blocks, threads>>>(
            output.data_ptr<scalar_t>(),
            output.numel(),
            min_value,
            divisor
        );
    }));

    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "3D Transposed convolution with clamp and divide (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.128 inst/cycle 0.000 5
Executed Ipc Elapsed 1.092 inst/cycle 0.000 5
Issue Slots Busy 28.210 % 0.003 5
Issued Ipc Active 1.128 inst/cycle 0.000 5
SM Busy 28.210 % 0.003 5
Memory Throughput 2355059259746.092 byte/second 55386006150357114880.000 5
Mem Busy 39.346 % 0.011 5
Max Bandwidth 70.284 % 0.044 5
L1/TEX Hit Rate 49.664 % 0.000 5
L2 Hit Rate 50.312 % 0.001 5
Mem Pipes Busy 12.222 % 0.001 5
Warp Cycles Per Issued Instruction 44.780 cycle 0.014 5
Warp Cycles Per Executed Instruction 44.840 cycle 0.014 5
Avg. Active Threads Per Warp 32.000 0.000 5
Avg. Not Predicated Off Threads Per Warp 30.500 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 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 79.776 % 0.000 5
Achieved Active Warps Per SM 51.056 warp 0.000 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (21.4%) 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 (79.8%) 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::conv_transpose3d
CPU Time 2452260.00 μs
Device Time 3471933.26 μs
Self CPU Time 15460.35 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::convolution
CPU Time 2436799.65 μs
Device Time 3471933.26 μs
Self CPU Time 22074.80 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::_convolution
CPU Time 2414724.85 μs
Device Time 3471933.26 μs
Self CPU Time 44886.58 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::cudnn_convolution_transpose
CPU Time 850774.43 μs
Device Time 2665070.96 μs
Self CPU Time 221558.83 μs
Self Device Time 2665070.96 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 3374165.41 μs
Device Time 60888.74 μs
Self CPU Time 3374165.41 μs
Self Device Time 60888.74 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
sm90_xmma_dgrad_implicit_gemm_indexed_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x64x32_warpgroupsize1x1x1_g1_strided_execute_kernel__5x_cudnn
CPU Time 0.00 μs
Device Time 1595083.48 μs
Self CPU Time 0.00 μs
Self Device Time 1595083.48 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::add_
CPU Time 1505002.79 μs
Device Time 806862.30 μs
Self CPU Time 37979.23 μs
Self Device Time 806862.30 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Failed 
45254 warnings and 1 error generated when compiling for host.
Error while processing /home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu.
Suppressed 45292 warnings (45245 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.
Found compiler error(s).
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:8:5 bugprone-easily-swappable-parameters
8 | const int64_t numel,
| ^~~~~~~~~~~~~~~~~~~~
9 | const float min_value,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:8:19: note: the first parameter in the range is 'numel'
8 | const int64_t numel,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:9:17: note: the last parameter in the range is 'min_value'
9 | const float min_value,
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:8:5: note:
8 | const int64_t numel,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:9:5: note: 'const int64_t' and 'const float' may be implicitly converted: 'const int64_t' (as 'long') -> 'const float' (as 'float'), 'const float' (as 'float') -> 'const int64_t' (as 'long')
9 | const float min_value,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:12:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
12 | const int tid = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:13:31: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
13 | const int total_threads = gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:22:19: warning: the parameter 'input' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
22 | torch::Tensor input,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:24:5: warning: 2 adjacent parameters of 'forward' of convertible types are easily swapped by mistake [bugprone-easily-swappable-parameters]
24 | int padding,
| ^~~~~~~~~~~~
25 | float min_value,
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:24:9: note: the first parameter in the range is 'padding'
24 | int padding,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:25:11: note: the last parameter in the range is 'min_value'
25 | float min_value,
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:25:5: note: 'int' and 'float' may be implicitly converted
25 | float min_value,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:27:19: warning: the parameter 'weight' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
27 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:40:24: error: no matching function for call to 'min' [clang-diagnostic-error]
40 | const int blocks = min(max_blocks, static_cast<int>((output.numel() + threads - 1) / threads));
| ^~~
/home/common_modules/clang-tidy/20.0.0git/lib/clang/20/include/__clang_cuda_math.h:201:16: note: candidate function not viable: call to __device__ function from __host__ function
201 | __DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
| ^
/usr/local/cuda/include/crt/math_functions.hpp:868:38: note: candidate function not viable: call to __device__ function from __host__ function
868 | __MATH_FUNCTIONS_DECL__ unsigned int min(const unsigned int a, const unsigned int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:873:38: note: candidate function not viable: call to __device__ function from __host__ function
873 | __MATH_FUNCTIONS_DECL__ unsigned int min(const int a, const unsigned int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:878:38: note: candidate function not viable: call to __device__ function from __host__ function
878 | __MATH_FUNCTIONS_DECL__ unsigned int min(const unsigned int a, const int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:883:34: note: candidate function not viable: call to __device__ function from __host__ function
883 | __MATH_FUNCTIONS_DECL__ long int min(const long int a, const long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:902:43: note: candidate function not viable: call to __device__ function from __host__ function
902 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const unsigned long int a, const unsigned long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:919:43: note: candidate function not viable: call to __device__ function from __host__ function
919 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const long int a, const unsigned long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:936:43: note: candidate function not viable: call to __device__ function from __host__ function
936 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const unsigned long int a, const long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:953:39: note: candidate function not viable: call to __device__ function from __host__ function
953 | __MATH_FUNCTIONS_DECL__ long long int min(const long long int a, const long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:958:48: note: candidate function not viable: call to __device__ function from __host__ function
958 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const unsigned long long int a, const unsigned long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:963:48: note: candidate function not viable: call to __device__ function from __host__ function
963 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const long long int a, const unsigned long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:968:48: note: candidate function not viable: call to __device__ function from __host__ function
968 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const unsigned long long int a, const long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:973:31: note: candidate function not viable: call to __device__ function from __host__ function
973 | __MATH_FUNCTIONS_DECL__ float min(const float a, const float b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:978:32: note: candidate function not viable: call to __device__ function from __host__ function
978 | __MATH_FUNCTIONS_DECL__ double min(const double a, const double b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:983:32: note: candidate function not viable: call to __device__ function from __host__ function
983 | __MATH_FUNCTIONS_DECL__ double min(const float a, const double b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:988:32: note: candidate function not viable: call to __device__ function from __host__ function
988 | __MATH_FUNCTIONS_DECL__ double min(const double a, const float b)
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b5_s1_balanced_workload_distribution/base/base.cu:42:5: 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]
42 | AT_DISPATCH_FLOATING_TYPES(output.scalar_type(), "clamp_and_divide", ([&] {
| ^
/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__, \
| ^