← Back to Leaderboard

The AI CUDA Engineer 👷

35_LTSM • 35_lstm_warp_aligned_base

Level 3 • Task 35
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import _VF


def module_fn(
    x: torch.Tensor,
    lstm_weights_ih: torch.Tensor,
    lstm_weights_hh: torch.Tensor,
    lstm_biases_ih: torch.Tensor,
    lstm_biases_hh: torch.Tensor,
    fc_weight: torch.Tensor,
    fc_bias: torch.Tensor,
    h0: torch.Tensor,
    c0: torch.Tensor,
    is_training: bool,
) -> torch.Tensor:
    """
    LSTM forward pass

    Args:
        x: Input tensor of shape (batch_size, sequence_length, input_size)
        lstm_weights_ih: List of input-hidden weight tensors for each LSTM layer
        lstm_weights_hh: List of hidden-hidden weight tensors for each LSTM layer
        lstm_biases_ih: List of input-hidden bias tensors for each LSTM layer
        lstm_biases_hh: List of hidden-hidden bias tensors for each LSTM layer
        fc_weight: Weight tensor for final linear layer
        fc_bias: Bias tensor for final linear layer
        h0: Initial hidden state
        c0: Initial cell state
        is_training: Whether in training mode

    Returns:
        Output tensor of shape (batch_size, output_size)
    """
    h0 = h0.to(x.device)
    c0 = c0.to(x.device)

    # Run LSTM layers
    out = x

    for i in range(len(lstm_weights_ih)):
        params = (
            lstm_weights_ih[i],
            lstm_weights_hh[i],
            lstm_biases_ih[i],
            lstm_biases_hh[i],
        )
        out = _VF.lstm(
            out,
            (h0[i : i + 1], c0[i : i + 1]),
            params,
            True,  # has_biases
            1,  # num_layers
            0.0 if not is_training else dropout,  # dropout
            is_training,  # training
            False,  # bidirectional
            True,
        )[
            0
        ]  # batch_first, only keep output

    # Get last timestep and apply final linear layer
    out = F.linear(out[:, -1, :], fc_weight, fc_bias)

    return out


class Model(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, output_size, dropout=0.0):
        """
        Initialize the LSTM model.

        :param input_size: The number of expected features in the input `x`
        :param hidden_size: The number of features in the hidden state `h`
        :param num_layers: Number of recurrent layers
        :param output_size: The number of output features
        :param dropout: If non-zero, introduces a Dropout layer on the outputs of each LSTM layer except the last layer
        """
        super(Model, self).__init__()

        # Initialize hidden states
        self.h0 = torch.randn((num_layers, batch_size, hidden_size))
        self.c0 = torch.randn((num_layers, batch_size, hidden_size))

        # Extract LSTM parameters
        lstm = nn.LSTM(
            input_size,
            hidden_size,
            num_layers,
            batch_first=True,
            dropout=dropout,
            bidirectional=False,
        )

        # Get weights and biases for each layer
        self.lstm_weights_ih = nn.ParameterList()
        self.lstm_weights_hh = nn.ParameterList()
        self.lstm_biases_ih = nn.ParameterList()
        self.lstm_biases_hh = nn.ParameterList()

        for i in range(num_layers):
            self.lstm_weights_ih.append(
                nn.Parameter(getattr(lstm, f"weight_ih_l{i}").data.clone())
            )
            self.lstm_weights_hh.append(
                nn.Parameter(getattr(lstm, f"weight_hh_l{i}").data.clone())
            )
            self.lstm_biases_ih.append(
                nn.Parameter(getattr(lstm, f"bias_ih_l{i}").data.clone())
            )
            self.lstm_biases_hh.append(
                nn.Parameter(getattr(lstm, f"bias_hh_l{i}").data.clone())
            )

        # Extract linear layer parameters
        fc = nn.Linear(hidden_size, output_size)
        self.fc_weight = nn.Parameter(fc.weight.data.clone())
        self.fc_bias = nn.Parameter(fc.bias.data.clone())

    def forward(self, x, fn=module_fn):
        return fn(
            x,
            self.lstm_weights_ih,
            self.lstm_weights_hh,
            self.lstm_biases_ih,
            self.lstm_biases_hh,
            self.fc_weight,
            self.fc_bias,
            self.h0,
            self.c0,
            self.training,
        )


# Test code
batch_size = 10
sequence_length = 512
input_size = 128
hidden_size = 256
num_layers = 6
output_size = 10
dropout = 0.0


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


def get_init_inputs():
    return [input_size, hidden_size, num_layers, output_size, dropout]

import torch
import torch.nn as nn

class Model(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, output_size, dropout=0.0):
        """
        Initialize the LSTM model.

        :param input_size: The number of expected features in the input `x`
        :param hidden_size: The number of features in the hidden state `h`
        :param num_layers: Number of recurrent layers
        :param output_size: The number of output features
        :param dropout: If non-zero, introduces a Dropout layer on the outputs of each LSTM layer except the last layer, with dropout probability equal to `dropout`
        """
        super(Model, self).__init__()
        # Initialize hidden state with random values
        self.h0 = torch.randn((num_layers, batch_size, hidden_size))
        self.c0 = torch.randn((num_layers, batch_size, hidden_size))
        self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, dropout=dropout, bidirectional=False)
        self.fc = nn.Linear(hidden_size, output_size)
    
    def forward(self, x):
        """
        Forward pass through the LSTM model.

        :param x: The input tensor, shape (batch_size, sequence_length, input_size)
        :return: The output tensor, shape (batch_size, sequence_length, output_size)
        """
        self.h0 = self.h0.to(x.device)
        self.c0 = self.h0.to(x.device)
        
        # Forward propagate LSTM
        out, hn = self.lstm(x, (self.h0, self.c0))  # out: tensor of shape (batch_size, seq_length, hidden_size)
        
        # Decode the hidden state of the last time step
        out = self.fc(out[:, -1, :])  # out: tensor of shape (batch_size, output_size)
        
        return out

# Test code
batch_size = 10
sequence_length = 512
input_size = 128
hidden_size = 256
num_layers = 6
output_size = 10
dropout = 0.0

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

def get_init_inputs():
    return [input_size, hidden_size, num_layers, output_size, dropout]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 35_LTSM
Level ID 3
Task ID 35
Kernel Name 35_lstm_warp_aligned_base
CUDA Speedup (Native) 0.356x
CUDA Speedup (Compile) 0.673x
CUDA Runtime 89.808 ms
PyTorch Runtime (Native) 31.945 ms
PyTorch Runtime (Compile) 60.432 ms
Correct True
Max Diff (vs. Reference) 0.001000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
Temperature 0.00
View Experiment Progress Details

Related Kernels (Level 3, Task 35 • 35_LTSM)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 35_lstm_grid_stride_base_base 72.97 0.44 0.83
🥈 35_lstm_modular_device_edit_1 75.07 0.43 0.81
🥉 35_lstm_shared_memory_base 86.54 0.37 0.70
4 35_lstm_atomic_reduction_base_base 86.99 0.37 0.69
5 35_lstm_workload_balanced_base 87.77 0.36 0.69
6 35_lstm_aligned_base 88.03 0.36 0.69
7 35_lstm_tiled_unroll_edit_1 88.19 0.36 0.69
8 35_lstm_load_balancing_base 88.28 0.36 0.68
9 fused_tiled_base 88.40 0.36 0.68
10 35_lstm_ldg_aligned_v2_base 88.50 0.36 0.68
11 35_lstm_load_balancing_edit_1 88.68 0.36 0.68
12 35_LTSM 88.90 0.36 0.68
13 35_lstm_memory_coalescing_edit_1 89.05 0.36 0.68
14 modular_35_ltsm_base 89.17 0.36 0.68
15 35_lstm_shared_memory_edit_1 89.34 0.36 0.68
16 fused_tiled_edit_1 89.35 0.36 0.68
17 35_lstm_unrolled_base 89.58 0.36 0.67
18 35_lstm_memory_coalescing_base 89.77 0.36 0.67
19 35_lstm_warp_reduce_base 89.78 0.36 0.67
20 35_lstm_warp_aligned_base 89.81 0.36 0.67 
#include <torch/extension.h>
#include <vector>

// Optimized CUDA kernel with reduced warp divergence
__global__ void lstm_elementwise_forward(
    const float* __restrict__ gates,
    const float* __restrict__ prev_c,
    float* __restrict__ h,
    float* __restrict__ c,
    int batch_size,
    int hidden_size
) {
    // Align to warp size (32 threads)
    const int warp_size = 32;
    const int lane_id = threadIdx.x % warp_size;
    const int warp_id = threadIdx.x / warp_size;
    const int warps_per_block = blockDim.x / warp_size;
    const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
    
    // Process elements in warp-aligned chunks
    for (int base_idx = global_warp_id * warp_size; 
         base_idx < batch_size * hidden_size; 
         base_idx += gridDim.x * warps_per_block * warp_size) {
        
        int idx = base_idx + lane_id;
        
        // Pre-compute bounds check for entire warp
        bool valid = idx < batch_size * hidden_size;
        
        if (valid) {
            int b = idx / hidden_size;
            int n = idx % hidden_size;
            int gate_idx = b * hidden_size * 4 + n;

            float i_gate = gates[gate_idx];
            float f_gate = gates[gate_idx + hidden_size];
            float g_gate = gates[gate_idx + 2 * hidden_size];
            float o_gate = gates[gate_idx + 3 * hidden_size];

            // Vectorized activation computations
            i_gate = 1.0f / (1.0f + expf(-i_gate));
            f_gate = 1.0f / (1.0f + expf(-f_gate));
            g_gate = tanhf(g_gate);
            o_gate = 1.0f / (1.0f + expf(-o_gate));

            float c_new = f_gate * prev_c[idx] + i_gate * g_gate;
            float h_new = o_gate * tanhf(c_new);

            c[idx] = c_new;
            h[idx] = h_new;
        }
    }
}

// Optimized linear kernel with reduced warp divergence
__global__ void linear_forward_kernel(
    const float* __restrict__ input,
    const float* __restrict__ weight,
    const float* __restrict__ bias,
    float* __restrict__ output,
    int batch_size,
    int input_dim,
    int output_dim
) {
    const int warp_size = 32;
    const int lane_id = threadIdx.x % warp_size;
    const int warp_id = threadIdx.x / warp_size;
    const int warps_per_block = blockDim.x / warp_size;
    const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
    
    for (int base_idx = global_warp_id * warp_size;
         base_idx < batch_size * output_dim;
         base_idx += gridDim.x * warps_per_block * warp_size) {
        
        int idx = base_idx + lane_id;
        bool valid = idx < batch_size * output_dim;
        
        if (valid) {
            int batch_idx = idx / output_dim;
            int out_idx = idx % output_dim;
            float sum = 0.0f;
            
            // Coalesced memory access pattern
            #pragma unroll 4
            for (int k = 0; k < input_dim; k++) {
                sum += input[batch_idx * input_dim + k] * 
                       weight[out_idx * input_dim + k];
            }
            
            if (bias != nullptr) {
                sum += bias[out_idx];
            }
            output[idx] = sum;
        }
    }
}

torch::Tensor lstm_forward_cuda(
    torch::Tensor input,
    torch::Tensor w_ih,
    torch::Tensor w_hh,
    torch::Tensor b_ih,
    torch::Tensor b_hh,
    torch::Tensor h0,
    torch::Tensor c0
) {
    auto options = torch::TensorOptions().dtype(input.dtype()).device(input.device());
    int batch_size = input.size(0);
    int seq_len = input.size(1);
    int hidden_size = h0.size(1);

    torch::Tensor h = h0.clone();
    torch::Tensor c = c0.clone();
    std::vector<torch::Tensor> outputs;

    const int threads_per_block = 256;
    const int num_warps = threads_per_block / 32;
    const int total_elements = batch_size * hidden_size;
    const int blocks = (total_elements + threads_per_block - 1) / threads_per_block;

    for (int t = 0; t < seq_len; ++t) {
        torch::Tensor xt = input.select(1, t);
        torch::Tensor gates = torch::addmm(b_ih, xt, w_ih.t());
        gates = torch::addmm(gates, h, w_hh.t());
        gates += b_hh;

        lstm_elementwise_forward<<<blocks, threads_per_block>>>(
            gates.data_ptr<float>(),
            c.data_ptr<float>(),
            h.data_ptr<float>(),
            c.data_ptr<float>(),
            batch_size,
            hidden_size
        );

        outputs.push_back(h.unsqueeze(1));
    }

    return torch::cat(outputs, 1);
}

torch::Tensor linear_forward_cuda(
    torch::Tensor input,
    torch::Tensor weight,
    torch::Tensor bias
) {
    const auto batch_size = input.size(0);
    const auto input_dim = input.size(1);
    const auto output_dim = weight.size(0);

    auto options = torch::TensorOptions().dtype(input.dtype()).device(input.device());
    auto output = torch::empty({batch_size, output_dim}, options);

    const int threads_per_block = 256;
    const int blocks = (batch_size * output_dim + threads_per_block - 1) / threads_per_block;

    linear_forward_kernel<<<blocks, threads_per_block>>>(
        input.data_ptr<float>(),
        weight.data_ptr<float>(),
        bias.defined() ? bias.data_ptr<float>() : nullptr,
        output.data_ptr<float>(),
        batch_size,
        input_dim,
        output_dim
    );

    return output;
}

torch::Tensor forward(
    torch::Tensor x,
    std::vector<torch::Tensor> lstm_weights_ih,
    std::vector<torch::Tensor> lstm_weights_hh,
    std::vector<torch::Tensor> lstm_biases_ih,
    std::vector<torch::Tensor> lstm_biases_hh,
    torch::Tensor fc_weight,
    torch::Tensor fc_bias,
    torch::Tensor h0,
    torch::Tensor c0,
    bool is_training
) {
    h0 = h0.to(x.device());
    c0 = c0.to(x.device());
    torch::Tensor out = x;
    
    for (int i = 0; i < lstm_weights_ih.size(); ++i) {
        auto w_ih = lstm_weights_ih[i].to(x.device());
        auto w_hh = lstm_weights_hh[i].to(x.device());
        auto b_ih = lstm_biases_ih[i].to(x.device());
        auto b_hh = lstm_biases_hh[i].to(x.device());
        
        torch::Tensor h_i = h0.narrow(0, i, 1).squeeze(0);
        torch::Tensor c_i = c0.narrow(0, i, 1).squeeze(0);
        
        out = lstm_forward_cuda(out, w_ih, w_hh, b_ih, b_hh, h_i, c_i);
    }
    
    out = out.select(1, -1);
    out = linear_forward_cuda(out, fc_weight, fc_bias);
    return out;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Module forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.240 inst/cycle 0.000 5
Executed Ipc Elapsed 0.000 inst/cycle 0.000 5
Issue Slots Busy 6.094 % 0.000 5
Issued Ipc Active 0.240 inst/cycle 0.000 5
SM Busy 6.094 % 0.000 5
Memory Throughput 1412490029.490 byte/second 232047730527368.938 5
Mem Busy 1.764 % 0.001 5
Max Bandwidth 0.928 % 0.000 5
L1/TEX Hit Rate 94.570 % 0.000 5
L2 Hit Rate 100.444 % 0.640 5
Mem Pipes Busy 0.050 % 0.000 5
Warp Cycles Per Issued Instruction 16.810 cycle 0.077 5
Warp Cycles Per Executed Instruction 16.972 cycle 0.079 5
Avg. Active Threads Per Warp 25.230 0.000 5
Avg. Not Predicated Off Threads Per Warp 25.050 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 6.424 % 0.004 5
Achieved Active Warps Per SM 4.114 warp 0.002 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 (6.4%) 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
cudaMemcpyAsync
CPU Time 806734.16 μs
Device Time 0.00 μs
Self CPU Time 806734.16 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::t
CPU Time 894788.04 μs
Device Time 0.00 μs
Self CPU Time 395296.83 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::addmm
CPU Time 6334635.71 μs
Device Time 2567736.72 μs
Self CPU Time 3953197.96 μs
Self Device Time 2567736.72 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 2612152.68 μs
Device Time 88596.82 μs
Self CPU Time 2612152.68 μs
Self Device Time 88596.82 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void gemmSN_TN_kernel<float, 128, 16, 2, 4, 10, 11, false, cublasGemvTensorStridedBatched<float const>, cublasGemvTensorStridedBatched<float const>, cublasGemvTensorStridedBatched<float> >(cublasGemmSmallNParams<cublasGemvTensorStridedBatched<float const>, cublasGemvTensorStridedBatched<float const>, cublasGemvTensorStridedBatched<float>, float>)
CPU Time 0.00 μs
Device Time 1859455.20 μs
Self CPU Time 0.00 μs
Self Device Time 1859455.20 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 1240565.28 μs
Device Time 666128.65 μs
Self CPU Time 548980.53 μs
Self Device Time 666128.65 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::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})
CPU Time 0.00 μs
Device Time 666128.65 μs
Self CPU Time 0.00 μs
Self Device Time 666128.65 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
lstm_elementwise_forward(float const*, float const*, float*, float*, int, int)
CPU Time 0.00 μs
Device Time 576201.86 μs
Self CPU Time 0.00 μs
Self Device Time 576201.86 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 
45322 warnings generated when compiling for host.
Suppressed 45330 warnings (45283 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_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:6:5 bugprone-easily-swappable-parameters
6 | const float* __restrict__ gates,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
7 | const float* __restrict__ prev_c,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:6:31: note: the first parameter in the range is 'gates'
6 | const float* __restrict__ gates,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:7:31: note: the last parameter in the range is 'prev_c'
7 | const float* __restrict__ prev_c,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:8:5: warning: 2 adjacent parameters of 'lstm_elementwise_forward' of similar type ('float *__restrict') are easily swapped by mistake [bugprone-easily-swappable-parameters]
8 | float* __restrict__ h,
| ^~~~~~~~~~~~~~~~~~~~~~
9 | float* __restrict__ c,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:8:25: note: the first parameter in the range is 'h'
8 | float* __restrict__ h,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:9:25: note: the last parameter in the range is 'c'
9 | float* __restrict__ c,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:15:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
15 | const int lane_id = threadIdx.x % warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:16:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
16 | const int warp_id = threadIdx.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:17:33: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
17 | const int warps_per_block = blockDim.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:18:32: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
18 | const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:23:22: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
23 | base_idx += gridDim.x * warps_per_block * warp_size) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:58:5: warning: 2 adjacent parameters of 'linear_forward_kernel' of similar type ('const float *__restrict') are easily swapped by mistake [bugprone-easily-swappable-parameters]
58 | const float* __restrict__ weight,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
59 | const float* __restrict__ bias,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:58:31: note: the first parameter in the range is 'weight'
58 | const float* __restrict__ weight,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:59:31: note: the last parameter in the range is 'bias'
59 | const float* __restrict__ bias,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:61:5: warning: 2 adjacent parameters of 'linear_forward_kernel' of similar type ('int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
61 | int batch_size,
| ^~~~~~~~~~~~~~~
62 | int input_dim,
| ~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:61:9: note: the first parameter in the range is 'batch_size'
61 | int batch_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:62:9: note: the last parameter in the range is 'input_dim'
62 | int input_dim,
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:66:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
66 | const int lane_id = threadIdx.x % warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:67:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
67 | const int warp_id = threadIdx.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:68:33: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
68 | const int warps_per_block = blockDim.x / warp_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:69:32: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
69 | const int global_warp_id = blockIdx.x * warps_per_block + warp_id;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:73:22: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
73 | base_idx += gridDim.x * warps_per_block * warp_size) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:99:5: warning: 2 adjacent parameters of 'lstm_forward_cuda' of similar type ('torch::Tensor') are easily swapped by mistake [bugprone-easily-swappable-parameters]
99 | torch::Tensor input,
| ^~~~~~~~~~~~~~~~~~~~
100 | torch::Tensor w_ih,
| ~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:99:19: note: the first parameter in the range is 'input'
99 | torch::Tensor input,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:100:19: note: the last parameter in the range is 'w_ih'
100 | torch::Tensor w_ih,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:99:19: 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]
99 | torch::Tensor input,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:100:19: warning: the parameter 'w_ih' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
100 | torch::Tensor w_ih,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:101:5: warning: 4 adjacent parameters of 'lstm_forward_cuda' of similar type ('torch::Tensor') are easily swapped by mistake [bugprone-easily-swappable-parameters]
101 | torch::Tensor w_hh,
| ^~~~~~~~~~~~~~~~~~~
102 | torch::Tensor b_ih,
| ~~~~~~~~~~~~~~~~~~~
103 | torch::Tensor b_hh,
| ~~~~~~~~~~~~~~~~~~~
104 | torch::Tensor h0,
| ~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:101:19: note: the first parameter in the range is 'w_hh'
101 | torch::Tensor w_hh,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:104:19: note: the last parameter in the range is 'h0'
104 | torch::Tensor h0,
| ^~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:101:19: warning: the parameter 'w_hh' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
101 | torch::Tensor w_hh,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:102:19: warning: the parameter 'b_ih' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
102 | torch::Tensor b_ih,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:103:19: warning: the parameter 'b_hh' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
103 | torch::Tensor b_hh,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:104:19: warning: the parameter 'h0' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
104 | torch::Tensor h0,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:105:19: warning: the parameter 'c0' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
105 | torch::Tensor c0
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:107:10: warning: Value stored to 'options' during its initialization is never read [clang-analyzer-deadcode.DeadStores]
107 | auto options = torch::TensorOptions().dtype(input.dtype()).device(input.device());
| ^~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:107:10: note: Value stored to 'options' during its initialization is never read
107 | auto options = torch::TensorOptions().dtype(input.dtype()).device(input.device());
| ^~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:108:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
108 | int batch_size = input.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:109:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
109 | int seq_len = input.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:110:23: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
110 | int hidden_size = h0.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:143:19: 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]
143 | torch::Tensor input,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:144:19: 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]
144 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:145:19: warning: the parameter 'bias' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
145 | torch::Tensor bias
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:155:24: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
155 | const int blocks = (batch_size * output_dim + threads_per_block - 1) / threads_per_block;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:162:9: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
162 | batch_size,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:163:9: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
163 | input_dim,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:164:9: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
164 | output_dim
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:171:19: 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]
171 | torch::Tensor x,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:172:5: warning: 4 adjacent parameters of 'forward' of similar type ('std::vector<torch::Tensor>') are easily swapped by mistake [bugprone-easily-swappable-parameters]
172 | std::vector<torch::Tensor> lstm_weights_ih,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
173 | std::vector<torch::Tensor> lstm_weights_hh,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
174 | std::vector<torch::Tensor> lstm_biases_ih,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
175 | std::vector<torch::Tensor> lstm_biases_hh,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:172:32: note: the first parameter in the range is 'lstm_weights_ih'
172 | std::vector<torch::Tensor> lstm_weights_ih,
| ^~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:175:32: note: the last parameter in the range is 'lstm_biases_hh'
175 | std::vector<torch::Tensor> lstm_biases_hh,
| ^~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:177:5: warning: 2 adjacent parameters of 'forward' of similar type ('torch::Tensor') are easily swapped by mistake [bugprone-easily-swappable-parameters]
177 | torch::Tensor fc_bias,
| ^~~~~~~~~~~~~~~~~~~~~~
178 | torch::Tensor h0,
| ~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:177:19: note: the first parameter in the range is 'fc_bias'
177 | torch::Tensor fc_bias,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:178:19: note: the last parameter in the range is 'h0'
178 | torch::Tensor h0,
| ^~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:199:36: warning: parameter 'fc_weight' is passed by value and only copied once; consider moving it to avoid unnecessary copies [performance-unnecessary-value-param]
2 | out = linear_forward_cuda(out, fc_weight, fc_bias);
| ^
| std::move( )
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_3/task_35/b1_s2_35_lstm_warp_aligned/base/base.cu:199:47: warning: parameter 'fc_bias' is passed by value and only copied once; consider moving it to avoid unnecessary copies [performance-unnecessary-value-param]
199 | out = linear_forward_cuda(out, fc_weight, fc_bias);
| ^
| std::move( )