ryankert01 commented on code in PR #675:
URL: https://github.com/apache/mahout/pull/675#discussion_r2578775251
##########
qdp/qdp-kernels/src/lib.rs:
##########
@@ -47,8 +47,8 @@ unsafe extern "C" {
pub fn launch_amplitude_encode(
input_d: *const f64,
state_d: *mut c_void,
- input_len: i32,
- state_len: i32,
+ input_len: usize,
+ state_len: usize,
Review Comment:
what's behind it?
##########
qdp/qdp-core/src/gpu/pipeline.rs:
##########
@@ -0,0 +1,173 @@
+//
+// Licensed to the Apache Software Foundation (ASF) under one or more
+// contributor license agreements. See the NOTICE file distributed with
+// this work for additional information regarding copyright ownership.
+// The ASF licenses this file to You under the Apache License, Version 2.0
+// (the "License"); you may not use this file except in compliance with
+// the License. You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Async Pipeline Infrastructure
+//
+// Provides generic double-buffered execution for large data processing.
+// Separates the "streaming mechanics" from the "kernel logic".
+
+use std::sync::Arc;
+use std::ffi::c_void;
+use cudarc::driver::{CudaDevice, CudaSlice, DevicePtr, safe::CudaStream};
+use crate::error::{MahoutError, Result};
+
+/// Chunk processing callback for async pipeline
+///
+/// This closure is called for each chunk with:
+/// - `stream`: The CUDA stream to launch the kernel on
+/// - `input_ptr`: Device pointer to the chunk data (already copied)
+/// - `chunk_offset`: Global offset in the original data (in elements)
+/// - `chunk_len`: Length of this chunk (in elements)
+pub type ChunkProcessor = dyn FnMut(&CudaStream, *const f64, usize, usize) ->
Result<()>;
+
+/// Executes a task using dual-stream double-buffering pattern
+///
+/// This function handles the generic pipeline mechanics:
+/// - Dual stream creation and management
+/// - Data chunking and async H2D copy
+/// - Buffer lifetime management
+/// - Stream synchronization
+///
+/// The caller provides a `kernel_launcher` closure that handles the
+/// specific kernel launch logic for each chunk.
+///
+/// # Arguments
+/// * `device` - The CUDA device
+/// * `host_data` - Full source data to process
+/// * `kernel_launcher` - Closure that launches the specific kernel for each
chunk
+///
+/// # Example
+/// ```rust,ignore
+/// run_dual_stream_pipeline(device, host_data, |stream, input_ptr, offset,
len| {
+/// // Launch your specific kernel here
+/// launch_my_kernel(input_ptr, offset, len, stream)?;
+/// Ok(())
+/// })?;
+/// ```
+#[cfg(target_os = "linux")]
+pub fn run_dual_stream_pipeline<F>(
Review Comment:
Could you provide some context of `run_dual_stream_pipeline` ? First glance,
I though why not 10 streams.
##########
qdp/qdp-core/src/gpu/pipeline.rs:
##########
@@ -0,0 +1,173 @@
+//
+// Licensed to the Apache Software Foundation (ASF) under one or more
+// contributor license agreements. See the NOTICE file distributed with
+// this work for additional information regarding copyright ownership.
+// The ASF licenses this file to You under the Apache License, Version 2.0
+// (the "License"); you may not use this file except in compliance with
+// the License. You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Async Pipeline Infrastructure
+//
+// Provides generic double-buffered execution for large data processing.
+// Separates the "streaming mechanics" from the "kernel logic".
+
+use std::sync::Arc;
+use std::ffi::c_void;
+use cudarc::driver::{CudaDevice, CudaSlice, DevicePtr, safe::CudaStream};
+use crate::error::{MahoutError, Result};
+
+/// Chunk processing callback for async pipeline
+///
+/// This closure is called for each chunk with:
+/// - `stream`: The CUDA stream to launch the kernel on
+/// - `input_ptr`: Device pointer to the chunk data (already copied)
+/// - `chunk_offset`: Global offset in the original data (in elements)
+/// - `chunk_len`: Length of this chunk (in elements)
+pub type ChunkProcessor = dyn FnMut(&CudaStream, *const f64, usize, usize) ->
Result<()>;
+
+/// Executes a task using dual-stream double-buffering pattern
+///
+/// This function handles the generic pipeline mechanics:
+/// - Dual stream creation and management
+/// - Data chunking and async H2D copy
+/// - Buffer lifetime management
+/// - Stream synchronization
+///
+/// The caller provides a `kernel_launcher` closure that handles the
+/// specific kernel launch logic for each chunk.
+///
+/// # Arguments
+/// * `device` - The CUDA device
+/// * `host_data` - Full source data to process
+/// * `kernel_launcher` - Closure that launches the specific kernel for each
chunk
+///
+/// # Example
+/// ```rust,ignore
+/// run_dual_stream_pipeline(device, host_data, |stream, input_ptr, offset,
len| {
+/// // Launch your specific kernel here
+/// launch_my_kernel(input_ptr, offset, len, stream)?;
+/// Ok(())
+/// })?;
+/// ```
+#[cfg(target_os = "linux")]
+pub fn run_dual_stream_pipeline<F>(
+ device: &Arc<CudaDevice>,
+ host_data: &[f64],
+ mut kernel_launcher: F,
+) -> Result<()>
+where
+ F: FnMut(&CudaStream, *const f64, usize, usize) -> Result<()>,
+{
+ crate::profile_scope!("GPU::AsyncPipeline");
+
+ // 1. Create dual streams for pipeline overlap
+ let stream1 = device.fork_default_stream()
+ .map_err(|e| MahoutError::Cuda(format!("Failed to create stream 1:
{:?}", e)))?;
+ let stream2 = device.fork_default_stream()
+ .map_err(|e| MahoutError::Cuda(format!("Failed to create stream 2:
{:?}", e)))?;
+ let streams = [&stream1, &stream2];
+
+ // 2. Chunk size: 8MB per chunk (balance between overhead and overlap
opportunity)
+ // Too small = launch overhead dominates, too large = less overlap
+ const CHUNK_SIZE_ELEMENTS: usize = 8 * 1024 * 1024 /
std::mem::size_of::<f64>(); // 8MB
Review Comment:
Is that a hyper-parameter that depands on what gpu we use? We can add a todo
to it if it is. (add some if-else to tune that)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
