← Back to Leaderboard

The AI CUDA Engineer 👷

98_KLDivLoss • kldiv_optimized_stride_base_base_base

Level 1 • Task 98
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
    """
    Computes the Kullback-Leibler Divergence for comparing two distributions.

    Args:
        predictions (torch.Tensor): Predicted values.
        targets (torch.Tensor): Target values.

    Returns:
        torch.Tensor: Kullback-Leibler Divergence.
    """
    return F.kl_div(torch.log(predictions), targets, reduction="batchmean")


class Model(nn.Module):
    """
    A model that computes Kullback-Leibler Divergence for comparing two distributions.

    Parameters:
        None
    """

    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets, fn=module_fn):
        return fn(predictions, targets)


batch_size = 128
input_shape = (4096,)
dim = 1


def get_inputs():
    return [
        torch.randn(batch_size, *input_shape).softmax(dim=-1),
        torch.randn(batch_size, *input_shape).softmax(dim=-1),
    ]


def get_init_inputs():
    return []

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that computes Kullback-Leibler Divergence for comparing two distributions.

    Parameters:
        None
    """
    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets):
        return torch.nn.functional.kl_div(torch.log(predictions), targets, reduction='batchmean')

batch_size = 128
input_shape = (4096, )
dim = 1

def get_inputs():
    return [torch.randn(batch_size, *input_shape).softmax(dim=-1), torch.randn(batch_size, *input_shape).softmax(dim=-1)]

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

Kernel Information

Operation Name 98_KLDivLoss
Level ID 1
Task ID 98
Kernel Name kldiv_optimized_stride_base_base_base
CUDA Speedup (Native) 2.595x
CUDA Speedup (Compile) 2.931x
CUDA Runtime 0.012 ms
PyTorch Runtime (Native) 0.031 ms
PyTorch Runtime (Compile) 0.035 ms
Correct True
Max Diff (vs. Reference) 0.003000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
Temperature 1.00
View Experiment Progress Details

Related Kernels (Level 1, Task 98 • 98_KLDivLoss)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 optimized_kl_div_cuda_base 0.01 2.83 3.20
🥈 kl_div_sync_optimized_base 0.01 2.59 2.93
🥈 optimized_kl_div_kernel_base 0.01 2.59 2.93
🥈 kl_div_balanced_workload_base 0.01 2.59 2.93
🥈 kl_div_warp_reduce_base_base 0.01 2.59 2.93
🥈 optimized_kl_div_base 0.01 2.59 2.93
🥈 kl_div_modular_reduce_base_base 0.01 2.59 2.93
🥈 kldiv_optimized_stride_base_base_base 0.01 2.59 2.93
🥈 vectorized_aligned_kl_base 0.01 2.59 2.93
🥈 98_KLDivLoss_optimal_reduce_edit_1 0.01 2.59 2.93
🥈 strided_warp_kl_base_base 0.01 2.59 2.93
🥈 fast_strided_kl_base 0.01 2.59 2.93
🥈 coalesced_chunked_kl_base 0.01 2.59 2.93
🥈 kldiv_modular_per_thread_base_base 0.01 2.59 2.93
🥈 kldiv_unrolled_reduction_base_base 0.01 2.59 2.93
🥈 kl_div_unrolled_reduce_base_base 0.01 2.59 2.93
🥈 warp_block_vec4_opt_base 0.01 2.59 2.93
🥈 vectorized_kldiv_base_base 0.01 2.59 2.93
🥈 kl_div_even_workload_distribution_base 0.01 2.59 2.93
🥈 adaptive_kl_div_cuda_base 0.01 2.59 2.93 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

__device__ __forceinline__ float warp_reduce(float val) {
    #pragma unroll
    for (int offset = 16; offset > 0; offset /= 2) {
        val += __shfl_down_sync(0xffffffff, val, offset);
    }
    return val;
}

__global__ void kl_div_kernel(
    const float* __restrict__ log_predictions,
    const float* __restrict__ targets,
    float* __restrict__ output,
    const int n) {
    
    extern __shared__ float smem[];
    
    const int tid = threadIdx.x;
    const int wid = tid >> 5;
    const int lane = tid & 31;
    
    // Calculate initial index and strides
    int idx = blockIdx.x * blockDim.x + tid;
    const int grid_stride = gridDim.x * blockDim.x;
    
    // Each thread maintains its own sum
    float thread_sum = 0.0f;
    
    // Fast path for aligned access (processes 8 elements per iteration)
    const int aligned_n = n & ~7;
    
    #pragma unroll
    for (; idx < aligned_n; idx += grid_stride * 8) {
        #pragma unroll
        for (int i = 0; i < 8; i++) {
            const int curr_idx = idx + i * grid_stride;
            if (curr_idx < n) {
                const float log_pred = log_predictions[curr_idx];
                const float target = targets[curr_idx];
                thread_sum += expf(log_pred) - target * log_pred;
            }
        }
    }
    
    // Handle remaining elements
    for (; idx < n; idx += grid_stride) {
        const float log_pred = log_predictions[idx];
        const float target = targets[idx];
        thread_sum += expf(log_pred) - target * log_pred;
    }
    
    // Warp-level reduction
    thread_sum = warp_reduce(thread_sum);
    
    // Store warp results to shared memory
    if (lane == 0) {
        smem[wid] = thread_sum;
    }
    __syncthreads();
    
    // Final reduction using first warp
    if (wid == 0) {
        float warp_sum = 0.0f;
        if (lane < (blockDim.x >> 5)) { // number of warps
            warp_sum = smem[lane];
        }
        
        warp_sum = warp_reduce(warp_sum);
        
        if (lane == 0) {
            atomicAdd(output, warp_sum);
        }
    }
}

torch::Tensor kl_div_cuda_forward(
    torch::Tensor log_predictions,
    torch::Tensor targets) {
    
    const int n = log_predictions.numel();
    auto output = torch::zeros({1}, log_predictions.options());
    
    // Optimize thread and block count based on input size
    const int threads = 256;
    const int min_elements_per_thread = 8;
    const int blocks = min((n + threads * min_elements_per_thread - 1) / 
                          (threads * min_elements_per_thread), 1024);
    
    // Shared memory for warp results
    const int shared_mem = (threads / 32) * sizeof(float);
    
    kl_div_kernel<<<blocks, threads, shared_mem>>>(
        log_predictions.data_ptr<float>(),
        targets.data_ptr<float>(),
        output.data_ptr<float>(),
        n
    );
    
    return output / static_cast<float>(n);
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &kl_div_cuda_forward, "KL divergence forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.498 inst/cycle 0.001 5
Executed Ipc Elapsed 0.274 inst/cycle 0.000 5
Issue Slots Busy 13.266 % 0.553 5
Issued Ipc Active 0.532 inst/cycle 0.001 5
SM Busy 13.266 % 0.553 5
Memory Throughput 824675029475.532 byte/second 100796186396015230976.000 5
Mem Busy 14.320 % 0.038 5
Max Bandwidth 24.800 % 0.132 5
L1/TEX Hit Rate 0.000 % 0.000 5
L2 Hit Rate 18.618 % 0.001 5
Mem Pipes Busy 9.326 % 0.016 5
Warp Cycles Per Issued Instruction 26.804 cycle 0.515 5
Warp Cycles Per Executed Instruction 28.498 cycle 0.583 5
Avg. Active Threads Per Warp 31.890 0.000 5
Avg. Not Predicated Off Threads Per Warp 30.210 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 8.000 block 0.000 5
Block Limit Shared Mem 28.000 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 22.576 % 0.014 5
Achieved Active Warps Per SM 14.452 warp 0.005 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 (22.6%) 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::zeros
CPU Time 5247528.06 μs
Device Time 206960.61 μs
Self CPU Time 127603.93 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::zero_
CPU Time 5544516.89 μs
Device Time 7347891.01 μs
Self CPU Time 321668.37 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 5222849.69 μs
Device Time 7347891.01 μs
Self CPU Time 368389.47 μs
Self Device Time 7347891.01 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 5496076.63 μs
Device Time 2056.21 μs
Self CPU Time 5496076.63 μs
Self Device Time 2056.21 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
kl_div_kernel(float const*, float const*, float*, int)
CPU Time 0.00 μs
Device Time 531209.49 μs
Self CPU Time 0.00 μs
Self Device Time 531209.49 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::div
CPU Time 826859.25 μs
Device Time 263296.13 μs
Self CPU Time 496205.13 μs
Self Device Time 263296.13 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::BUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, at::detail::Array<char*, 2> >(int, at::native::BUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, at::detail::Array<char*, 2>)
CPU Time 0.00 μs
Device Time 263319.04 μs
Self CPU Time 0.00 μs
Self Device Time 263319.04 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 7140930.39 μs
Self CPU Time 0.00 μs
Self Device Time 7140930.39 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 
45247 warnings and 1 error generated when compiling for host.
Error while processing /home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu.
Suppressed 45287 warnings (45240 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/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:14:5 bugprone-easily-swappable-parameters
14 | const float* __restrict__ log_predictions,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
15 | const float* __restrict__ targets,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:14:31: note: the first parameter in the range is 'log_predictions'
14 | const float* __restrict__ log_predictions,
| ^~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:15:31: note: the last parameter in the range is 'targets'
15 | const float* __restrict__ targets,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:21:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
21 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:26:15: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
26 | int idx = blockIdx.x * blockDim.x + tid;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:27:29: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
27 | const int grid_stride = gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:80:19: warning: the parameter 'log_predictions' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
80 | torch::Tensor log_predictions,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:81:19: warning: the parameter 'targets' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
81 | torch::Tensor targets) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:83:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
83 | const int n = log_predictions.numel();
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_models_1/level_1/task_98/b7_s3_kldiv_optimized_stride_base_base/base/base.cu:89:24: error: no matching function for call to 'min' [clang-diagnostic-error]
89 | const int blocks = min((n + threads * min_elements_per_thread - 1) /
| ^~~
/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)
| ^