← Back to Leaderboard

The AI CUDA Engineer 👷

100_ConvTranspose3d_Clamp_Min_Divide • modular_conv3d_clamp_divide_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 modular_conv3d_clamp_divide_base
CUDA Speedup (Native) 1.119x
CUDA Speedup (Compile) 0.830x
CUDA Runtime 0.578 ms
PyTorch Runtime (Native) 0.647 ms
PyTorch Runtime (Compile) 0.480 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
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>
__device__ __forceinline__ scalar_t clamp_and_divide_op(
    scalar_t val,
    const float min_value,
    const float divisor) {
    return max(val, static_cast<scalar_t>(min_value)) / static_cast<scalar_t>(divisor);
}

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 idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < numel) {
        output[idx] = clamp_and_divide_op(
            output[idx],
            min_value,
            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 blocks = (output.numel() + threads - 1) / threads;

    AT_DISPATCH_FLOATING_TYPES(output.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 0.960 inst/cycle 0.000 5
Executed Ipc Elapsed 0.930 inst/cycle 0.000 5
Issue Slots Busy 24.024 % 0.016 5
Issued Ipc Active 0.962 inst/cycle 0.000 5
SM Busy 24.024 % 0.016 5
Memory Throughput 2162217218081.316 byte/second 1551726112271306240.000 5
Mem Busy 36.118 % 0.001 5
Max Bandwidth 64.524 % 0.002 5
L1/TEX Hit Rate 49.748 % 0.000 5
L2 Hit Rate 50.424 % 0.008 5
Mem Pipes Busy 26.570 % 0.004 5
Warp Cycles Per Issued Instruction 47.376 cycle 0.016 5
Warp Cycles Per Executed Instruction 47.414 cycle 0.016 5
Avg. Active Threads Per Warp 32.000 0.000 5
Avg. Not Predicated Off Threads Per Warp 30.480 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 72.176 % 0.003 5
Achieved Active Warps Per SM 46.192 warp 0.001 5
Analysis Rules
Rule Description
WRN HighPipeUtilization All compute pipelines are under-utilized. Either this kernel is very small or it doesn't issue enough warps per scheduler. Check the Launch Statistics and Scheduler Statistics sections for further details.
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 (72.2%) 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 2393218.74 μs
Device Time 3362246.13 μs
Self CPU Time 12591.65 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 2380627.09 μs
Device Time 3362246.13 μs
Self CPU Time 17039.23 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 2363587.86 μs
Device Time 3362246.13 μs
Self CPU Time 36448.38 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 844662.45 μs
Device Time 2579691.37 μs
Self CPU Time 196277.01 μs
Self Device Time 2579691.37 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 3410600.32 μs
Device Time 55638.26 μs
Self CPU Time 3410600.32 μs
Self Device Time 55638.26 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 1545587.85 μs
Self CPU Time 0.00 μs
Self Device Time 1545587.85 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 1471668.14 μs
Device Time 782554.76 μs
Self CPU Time 28820.94 μs
Self Device Time 782554.76 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 
45291 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/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:16:5 bugprone-easily-swappable-parameters
16 | const int64_t numel,
| ^~~~~~~~~~~~~~~~~~~~
17 | const float min_value,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:16:19: note: the first parameter in the range is 'numel'
16 | const int64_t numel,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:17:17: note: the last parameter in the range is 'min_value'
17 | const float min_value,
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:16:5: note:
16 | const int64_t numel,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:17:5: 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')
17 | const float min_value,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:20:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
20 | const int idx = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:31:19: 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]
31 | torch::Tensor input,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:33:5: warning: 2 adjacent parameters of 'forward' of convertible types are easily swapped by mistake [bugprone-easily-swappable-parameters]
33 | int padding,
| ^~~~~~~~~~~~
34 | float min_value,
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:33:9: note: the first parameter in the range is 'padding'
33 | int padding,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:34:11: note: the last parameter in the range is 'min_value'
34 | float min_value,
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:34:5: note: 'int' and 'float' may be implicitly converted
34 | float min_value,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:36:19: 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]
36 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:48:24: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
48 | const int blocks = (output.numel() + threads - 1) / threads;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:50:5: 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]
50 | AT_DISPATCH_FLOATING_TYPES(output.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__, \
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:50:5: warning: 'scalar_type' is deprecated: passing at::DeprecatedTypeProperties to an AT_DISPATCH macro is deprecated, pass an at::ScalarType instead [clang-diagnostic-deprecated-declarations]
50 | AT_DISPATCH_FLOATING_TYPES(output.type(), "clamp_and_divide", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:237:3: 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:218:36: note: expanded from macro 'AT_DISPATCH_SWITCH'
218 | at::ScalarType _st = ::detail::scalar_type(the_type); \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:106:1: note: 'scalar_type' has been explicitly marked deprecated here
106 | C10_DEPRECATED_MESSAGE(
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Deprecated.h:24:43: note: expanded from macro 'C10_DEPRECATED_MESSAGE'
24 | #define C10_DEPRECATED_MESSAGE(message) [[deprecated(message)]]
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250211_optimize_b5_s4_e1_v2/level_2/task_100/b1_s3_modular_conv3d_clamp_divide/base/base.cu:50:39: warning: 'type' is deprecated: Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device(). [clang-diagnostic-deprecated-declarations]
50 | AT_DISPATCH_FLOATING_TYPES(output.type(), "clamp_and_divide", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/core/TensorBody.h:224:3: note: 'type' has been explicitly marked deprecated here
224 | C10_DEPRECATED_MESSAGE("Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device().")
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Deprecated.h:24:43: note: expanded from macro 'C10_DEPRECATED_MESSAGE'
24 | #define C10_DEPRECATED_MESSAGE(message) [[deprecated(message)]]
| ^