acquire lock before any changes can be made

guard algo benchmark results

mark mutex as mutable

re-add /bin/sh (?)

use 3.10

use 6.1.2

machine-learning/0001-fix-avoid-race-condition-for-rocm-conv-algo-caching.patch

@@ -1,25 +0,0 @@
-From e267bc9bab8b3873dba57323ddcd9a9d09a1211e Mon Sep 17 00:00:00 2001
-From: mertalev <101130780+mertalev@users.noreply.github.com>
-Date: Fri, 20 Dec 2024 00:59:21 -0500
-Subject: [PATCH] fix: avoid race condition for rocm conv algo caching
-
----
- onnxruntime/core/providers/rocm/nn/conv.cc | 2 ++
- 1 file changed, 2 insertions(+)
-
-diff --git a/onnxruntime/core/providers/rocm/nn/conv.cc b/onnxruntime/core/providers/rocm/nn/conv.cc
-index d7f47d07a8..ec438287ac 100644
---- a/onnxruntime/core/providers/rocm/nn/conv.cc
-+++ b/onnxruntime/core/providers/rocm/nn/conv.cc
-@@ -278,6 +278,8 @@ Status Conv<T, NHWC>::UpdateState(OpKernelContext* context, bool bias_expected)
-       HIP_CALL_THROW(hipMemsetAsync(s_.b_zero, 0, malloc_size, Stream(context)));
-     }
- 
-+    // lock is needed to avoid race condition during algo search
-+    std::lock_guard<OrtMutex> lock(s_.mutex);
-     if (!s_.cached_benchmark_fwd_results.contains(x_dims_miopen)) {
-       miopenConvAlgoPerf_t perf;
-       int algo_count = 1;
--- 
-2.43.0
-

machine-learning/0001-guard-algo-benchmark-results.patch

@@ -0,0 +1,58 @@
+From 1f5d6323fa69ee16feab25f8e1398c1aed03ee08 Mon Sep 17 00:00:00 2001
+From: mertalev <101130780+mertalev@users.noreply.github.com>
+Date: Sun, 29 Dec 2024 14:07:51 -0500
+Subject: [PATCH] guard algo benchmark results
+
+---
+ onnxruntime/core/providers/rocm/nn/conv.h | 6 ++++++
+ 1 file changed, 6 insertions(+)
+
+diff --git a/onnxruntime/core/providers/rocm/nn/conv.h b/onnxruntime/core/providers/rocm/nn/conv.h
+index bc9846203e..0086e064f1 100644
+--- a/onnxruntime/core/providers/rocm/nn/conv.h
++++ b/onnxruntime/core/providers/rocm/nn/conv.h
+@@ -52,6 +52,7 @@ class lru_unordered_map {
+   lru_unordered_map(size_t max_size) : max_size_(max_size) {}
+ 
+   void insert(const Key& key, const T& value) {
++    std::lock_guard<std::mutex> guard(mutex_);
+     auto it = items_.find(key);
+     if (it != items_.end()) {
+       it->second.value = value;
+@@ -69,6 +70,7 @@ class lru_unordered_map {
+   }
+ 
+   T& at(const Key& key) {
++    std::lock_guard<std::mutex> guard(mutex_);
+     auto it = items_.find(key);
+     if (it == items_.end()) {
+       throw std::out_of_range("There is no such key in cache");
+@@ -78,14 +80,17 @@ class lru_unordered_map {
+   }
+ 
+   bool contains(const Key& key) const {
++    std::lock_guard<std::mutex> guard(mutex_);
+     return items_.find(key) != items_.end();
+   }
+ 
+   size_t size() const {
++    std::lock_guard<std::mutex> guard(mutex_);
+     return items_.size();
+   }
+ 
+   void clear() {
++    std::lock_guard<std::mutex> guard(mutex_);
+     items_.clear();
+     lru_list_.clear();
+   }
+@@ -106,6 +111,7 @@ class lru_unordered_map {
+   size_t max_size_;
+   std::unordered_map<Key, value_type, Hash, KeyEqual, MapAllocator> items_;
+   list_type lru_list_;
++  mutable std::mutex mutex_;
+ };
+ 
+ // cached miopen descriptors
+-- 
+2.43.0
+

machine-learning/Dockerfile

@@ -17,11 +17,11 @@ RUN mkdir /opt/armnn && \
 
 # Warning: 26.3Gb of disk space required to pull this image
 # https://github.com/microsoft/onnxruntime/blob/main/dockerfiles/Dockerfile.rocm
-FROM rocm/dev-ubuntu-24.04:6.2.4-complete AS builder-rocm
+FROM rocm/dev-ubuntu-22.04:6.1.2-complete AS builder-rocm
 
 WORKDIR /code
 
-RUN apt-get update && apt-get install -y --no-install-recommends wget git python3.12-venv
+RUN apt-get update && apt-get install -y --no-install-recommends wget git python3.10-venv
 # Install same version as the Dockerfile provided by onnxruntime
 RUN wget -nv https://github.com/Kitware/CMake/releases/download/v3.27.3/cmake-3.27.3-linux-x86_64.sh && \
     chmod +x cmake-3.27.3-linux-x86_64.sh && \
@@ -32,12 +32,11 @@ RUN wget -nv https://github.com/Kitware/CMake/releases/download/v3.27.3/cmake-3.
 ENV PATH /code/cmake-3.27.3-linux-x86_64/bin:${PATH}
 
 # Prepare onnxruntime repository & build onnxruntime
-# Note: cannot upgrade from 1.19.2 as of writing until upstream updates the ROCm CI
 RUN git clone --single-branch --branch v1.19.2 --recursive "https://github.com/Microsoft/onnxruntime" onnxruntime
 WORKDIR /code/onnxruntime
 # Fix for multi-threading based on comments in https://github.com/microsoft/onnxruntime/pull/19567
-COPY ./0001-fix-avoid-race-condition-for-rocm-conv-algo-caching.patch /tmp/
+COPY ./0001-guard-algo-benchmark-results.patch /tmp/
-RUN git apply /tmp/0001-fix-avoid-race-condition-for-rocm-conv-algo-caching.patch
+RUN git apply /tmp/0001-guard-algo-benchmark-results.patch
 
 RUN /bin/sh ./dockerfiles/scripts/install_common_deps.sh
 # Note: the `parallel` setting uses a substantial amount of RAM