← Back to Leaderboard

The AI CUDA Engineer 👷

68_Matmul_Min_Subtract • efficient_thread_block_mapping_base

Level 2 • Task 68
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    linear_weight: torch.Tensor,
    linear_bias: torch.Tensor,
    constant: torch.Tensor,
) -> torch.Tensor:
    """
    Performs matrix multiplication, applies minimum with constant, and subtracts constant.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_features)
        linear_weight (torch.Tensor): Weight matrix of shape (out_features, in_features)
        linear_bias (torch.Tensor): Bias vector of shape (out_features)
        constant (torch.Tensor): Scalar constant tensor

    Returns:
        torch.Tensor: Output tensor of shape (batch_size, out_features)
    """
    x = F.linear(x, linear_weight, linear_bias)
    x = torch.min(x, constant)
    x = x - constant
    return x


class Model(nn.Module):
    """
    Simple model that performs a matrix multiplication, applies minimum, and subtracts a constant.
    """

    def __init__(self, in_features, out_features, constant):
        super(Model, self).__init__()
        gemm = nn.Linear(in_features, out_features)
        self.linear_weight = nn.Parameter(gemm.weight)
        self.linear_bias = nn.Parameter(gemm.bias)
        self.constant = nn.Parameter(torch.tensor(constant))

    def forward(self, x, fn=module_fn):
        return fn(x, self.linear_weight, self.linear_bias, self.constant)


batch_size = 128
in_features = 10
out_features = 5
constant = 2.0


def get_inputs():
    return [torch.randn(batch_size, in_features)]


def get_init_inputs():
    return [in_features, out_features, constant]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a matrix multiplication, applies minimum, and subtracts a constant.
    """
    def __init__(self, in_features, out_features, constant):
        super(Model, self).__init__()
        self.linear = nn.Linear(in_features, out_features)
        self.constant = nn.Parameter(torch.tensor(constant))

    def forward(self, x):
        x = self.linear(x)
        x = torch.min(x, self.constant)
        x = x - self.constant
        return x

batch_size = 128
in_features = 10
out_features = 5
constant = 2.0

def get_inputs():
    return [torch.randn(batch_size, in_features)]

def get_init_inputs():
    return [in_features, out_features, constant]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 68_Matmul_Min_Subtract
Level ID 2
Task ID 68
Kernel Name efficient_thread_block_mapping_base
CUDA Speedup (Native) 2.947x
CUDA Speedup (Compile) 1.920x
CUDA Runtime 0.008 ms
PyTorch Runtime (Native) 0.024 ms
PyTorch Runtime (Compile) 0.015 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 68 • 68_Matmul_Min_Subtract)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 optimized_shared_memory_sync_base 0.01 2.95 1.92
🥇 optimized_thread_block_indexing_base_base 0.01 2.95 1.92
🥇 modularized_device_functions_base_base 0.01 2.95 1.92
🥇 aligned_memory_access_ldg_base_base 0.01 2.95 1.92
🥇 stride_loop_optimization_thread_base_base 0.01 2.95 1.92
🥇 tiled_shared_memory_matmul_base_base 0.01 2.95 1.92
🥇 unrolled_loop_optimization_base 0.01 2.95 1.92
🥇 optimized_warp_coalesced_base 0.01 2.95 1.92
🥇 grid_2d_mapping_base 0.01 2.95 1.92
🥇 aligned_memory_access_base_edit_1 0.01 2.95 1.92
🥇 modular_device_functions_edit_1 0.01 2.95 1.92
🥇 aligned_memory_access_base_base 0.01 2.95 1.92
🥇 grid_2d_mapping_edit_1 0.01 2.95 1.92
🥇 efficient_thread_block_mapping_base 0.01 2.95 1.92
🥇 stride_loop_optimization_base_base 0.01 2.95 1.92
🥇 tiled_gemm_thread_mapping_base 0.01 2.95 1.92
17 modular_device_functions_base 0.01 2.62 1.71
17 aligned_ldg_vectorized_edit_1 0.01 2.62 1.71
17 aligned_ldg_vectorized_base 0.01 2.62 1.71
17 branchless_min_dot_edit_1 0.01 2.62 1.71 
#include <torch/extension.h>

#include <cuda.h>
#include <cuda_runtime.h>

#define TILE_SIZE 16

__device__ float compute_dot_product(const float* x_row, const float* weight_row, int in_features) {
    float sum = 0.0f;
    for (int j = 0; j < in_features; ++j) {
        sum += x_row[j] * weight_row[j];
    }
    return sum;
}

__device__ float apply_min_subtract(float computed, float bias, float constant) {
    float result = computed + bias;  // Add bias
    if (result > constant) {  // Min with constant
        result = constant;
    }
    result -= constant;  // Subtract constant
    return result;
}

__global__ void my_kernel(
    const float* x,
    const float* linear_weight,
    const float* linear_bias,
    const float* constant,
    float* y,
    int batch_size,
    int in_features,
    int out_features) {
    int batch_idx = blockIdx.x * TILE_SIZE + threadIdx.x;
    int out_idx = blockIdx.y * TILE_SIZE + threadIdx.y;

    if (batch_idx < batch_size && out_idx < out_features) {
        // Pointers to rows
        const float* x_row = x + batch_idx * in_features;
        const float* weight_row = linear_weight + out_idx * in_features;
        float bias = linear_bias[out_idx];
        float cst = *constant;

        // Compute dot product using device function
        float result = compute_dot_product(x_row, weight_row, in_features);

        // Apply min and subtract using another device function
        y[batch_idx * out_features + out_idx] = apply_min_subtract(result, bias, cst);
    }
}

torch::Tensor forward(
    torch::Tensor x,
    torch::Tensor linear_weight,
    torch::Tensor linear_bias,
    torch::Tensor constant) {
    TORCH_CHECK(x.is_cuda(), "x must be a CUDA tensor");
    TORCH_CHECK(linear_weight.is_cuda(), "linear_weight must be a CUDA tensor");
    TORCH_CHECK(linear_bias.is_cuda(), "linear_bias must be a CUDA tensor");
    TORCH_CHECK(constant.is_cuda(), "constant must be a CUDA tensor");

    int batch_size = x.size(0);
    int in_features = x.size(1);
    int out_features = linear_weight.size(0);

    auto y = torch::zeros({batch_size, out_features}, x.options());

    const float* x_ptr = x.data_ptr<float>();
    const float* weight_ptr = linear_weight.data_ptr<float>();
    const float* bias_ptr = linear_bias.data_ptr<float>();
    const float* constant_ptr = constant.data_ptr<float>();
    float* y_ptr = y.data_ptr<float>();

    dim3 threads(TILE_SIZE, TILE_SIZE);
    dim3 blocks((batch_size + TILE_SIZE - 1) / TILE_SIZE, (out_features + TILE_SIZE - 1) / TILE_SIZE);

    my_kernel<<<blocks, threads>>>(
        x_ptr,
        weight_ptr,
        bias_ptr,
        constant_ptr,
        y_ptr,
        batch_size,
        in_features,
        out_features);

    return y;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "CUDA forward function");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.204 inst/cycle 0.000 5
Executed Ipc Elapsed 0.006 inst/cycle 0.000 5
Issue Slots Busy 5.826 % 0.014 5
Issued Ipc Active 0.232 inst/cycle 0.000 5
SM Busy 5.826 % 0.014 5
Memory Throughput 3148706495.256 byte/second 3261915124058999.000 5
Mem Busy 8.972 % 0.011 5
Max Bandwidth 4.662 % 0.006 5
L1/TEX Hit Rate 93.110 % 0.000 5
L2 Hit Rate 102.356 % 0.063 5
Mem Pipes Busy 0.114 % 0.000 5
Warp Cycles Per Issued Instruction 17.246 cycle 0.076 5
Warp Cycles Per Executed Instruction 19.470 cycle 0.097 5
Avg. Active Threads Per Warp 27.770 0.000 5
Avg. Not Predicated Off Threads Per Warp 26.560 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 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 6.454 % 0.032 5
Achieved Active Warps Per SM 4.128 warp 0.013 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 (6.5%) 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 5804205.27 μs
Device Time 222905.28 μs
Self CPU Time 157577.46 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 6150519.18 μs
Device Time 7833177.32 μs
Self CPU Time 337016.00 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 5813505.24 μs
Device Time 7833177.32 μs
Self CPU Time 413272.10 μs
Self Device Time 7833177.32 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 5793373.06 μs
Device Time 2924.24 μs
Self CPU Time 5793373.06 μs
Self Device Time 2924.24 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
my_kernel(float const*, float const*, float const*, float const*, float*, int, int, int)
CPU Time 0.00 μs
Device Time 310934.61 μs
Self CPU Time 0.00 μs
Self Device Time 310934.61 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 268491.17 μs
Device Time 1262203.17 μs
Self CPU Time 268491.17 μs
Self Device Time 1262203.17 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 7610272.04 μs
Self CPU Time 0.00 μs
Self Device Time 7610272.04 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaEventElapsedTime
CPU Time 309826.24 μs
Device Time 0.00 μs
Self CPU Time 309826.24 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
Status: Completed 
45291 warnings generated when compiling for host.
Suppressed 45324 warnings (45277 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/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:16:53 bugprone-easily-swappable-parameters
16 | __device__ float apply_min_subtract(float computed, float bias, float constant) {
| ^~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:16:59: note: the first parameter in the range is 'bias'
16 | __device__ float apply_min_subtract(float computed, float bias, float constant) {
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:16:71: note: the last parameter in the range is 'constant'
16 | __device__ float apply_min_subtract(float computed, float bias, float constant) {
| ^~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:26:5: warning: 4 adjacent parameters of 'my_kernel' of similar type ('const float *') are easily swapped by mistake [bugprone-easily-swappable-parameters]
26 | const float* x,
| ^~~~~~~~~~~~~~~
27 | const float* linear_weight,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~
28 | const float* linear_bias,
| ~~~~~~~~~~~~~~~~~~~~~~~~~
29 | const float* constant,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:26:18: note: the first parameter in the range is 'x'
26 | const float* x,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:29:18: note: the last parameter in the range is 'constant'
29 | const float* constant,
| ^~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:31:5: warning: 2 adjacent parameters of 'my_kernel' of similar type ('int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
31 | int batch_size,
| ^~~~~~~~~~~~~~~
32 | int in_features,
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:31:9: note: the first parameter in the range is 'batch_size'
31 | int batch_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:32:9: note: the last parameter in the range is 'in_features'
32 | int in_features,
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:34:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
34 | int batch_idx = blockIdx.x * TILE_SIZE + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:35:19: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
35 | int out_idx = blockIdx.y * TILE_SIZE + threadIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:39:30: warning: result of multiplication in type 'int' is used as a pointer offset after an implicit widening conversion to type 'ptrdiff_t' [bugprone-implicit-widening-of-multiplication-result]
39 | const float* x_row = x + batch_idx * in_features;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:39:34: note: make conversion explicit to silence this warning
5 | const float* x_row = x + batch_idx * in_features;
| ^~~~~~~~~~~~~~~~~~~~~~~
| static_cast<ptrdiff_t>()
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:39:34: note: perform multiplication in a wider type
39 | const float* x_row = x + batch_idx * in_features;
| ^~~~~~~~~
| static_cast<ptrdiff_t>()
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:40:35: warning: result of multiplication in type 'int' is used as a pointer offset after an implicit widening conversion to type 'ptrdiff_t' [bugprone-implicit-widening-of-multiplication-result]
40 | const float* weight_row = linear_weight + out_idx * in_features;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:40:51: note: make conversion explicit to silence this warning
40 | const float* weight_row = linear_weight + out_idx * in_features;
| ^~~~~~~~~~~~~~~~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:40:51: note: perform multiplication in a wider type
40 | const float* weight_row = linear_weight + out_idx * in_features;
| ^~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:53:19: warning: the parameter 'x' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
53 | torch::Tensor x,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:54:19: warning: the parameter 'linear_weight' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
54 | torch::Tensor linear_weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:55:19: warning: the parameter 'linear_bias' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
55 | torch::Tensor linear_bias,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:56:19: warning: the parameter 'constant' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
56 | torch::Tensor constant) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:62:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
62 | int batch_size = x.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:63:23: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
63 | int in_features = x.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_68/b7_s3_efficient_thread_block_mapping/base/base.cu:64:24: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
64 | int out_features = linear_weight.size(0);
| ^