tests: switch sensitive amd gpu tests to safe math

.github/workflows/docker-bases.yml

@@ -222,6 +222,10 @@ jobs:
           file: './docker/Dockerfile.amd'
           push: true
           target: 'amdclang'
+          build-args: |
+            ROCM_VERSION=5.5.1
+            UCX_BRANCH=v1.13.1
+            OMPI_BRANCH=v4.1.4
           tags: devitocodes/bases:amd
 
       - name: AMD HIP image
@@ -231,6 +235,4 @@ jobs:
           file: './docker/Dockerfile.amd'
           push: true
           target: 'hip'
-          build-args: |
-            arch=hip
           tags: devitocodes/bases:amd-hip

devito/arch/compiler.py

@@ -563,7 +563,7 @@ def __init_finalize__(self, **kwargs):
             if language in ['C', 'openmp']:
                 self.ldflags += ['-target', 'x86_64-pc-linux-gnu']
                 self.ldflags += ['-fopenmp']
-                self.ldflags += ['--offload-arch=native']
+                self.ldflags += ['--offload-arch=%s' % platform.march]
         elif platform in [POWER8, POWER9]:
             # It doesn't make much sense to use AOMP on Power, but it should work
             self.cflags.append('-mcpu=native')

docker/Dockerfile.amd

@@ -45,6 +45,10 @@ ENV ROCM_HOME=/opt/rocm \
     UCX_HOME=/opt/ucx \
     OMPI_HOME=/opt/ompi
 
+# Adding ROCM
+ENV PATH=$ROCM_HOME/bin:$PATH \
+    LD_LIBRARY_PATH=$ROCM_HOME/lib:$ROCM_HOME/lib/llvm/lib:$LD_LIBRARY_PATH
+
 # Until rocm base has it fixed
 RUN ln -s /opt/rocm/llvm/bin/offload-arch /opt/rocm/bin/offload-arch | echo "offload-arch already exis"
 

docker/Dockerfile.devito

@@ -12,42 +12,46 @@ ARG USER_ID=1000
 ARG GROUP_ID=1000
 
 ##################  Install devito ############################################
-# Copy Devito
-ADD . /app/devito
 
-# Update if outdated
-RUN apt-get update
+# Update if outdated and install extras
+RUN apt-get update && \
+    apt-get install -y git cmake libncurses5-dev libncursesw5-dev libdrm-dev libsystemd-dev
 
-# Remove git files
-RUN rm -rf /app/devito/.git
+# Usefull utilities
+# Nvtop
+RUN git clone https://github.com/Syllo/nvtop.git /app/nvtop && \
+    mkdir -p /app/nvtop/build && cd /app/nvtop/build && \
+    cmake .. -DNVIDIA_SUPPORT=ON -DAMDGPU_SUPPORT=ON -DINTEL_SUPPORT=ON && \
+    make && make install
 
-# Install pip dependencies and devito as a pip package
+# Install pip dependencies
 RUN python3 -m venv /venv && \
     /venv/bin/pip install --no-cache-dir --upgrade pip && \
     /venv/bin/pip install --no-cache-dir jupyter && \
     /venv/bin/pip install --no-cache-dir wheel && \
-    eval "$MPI4PY_FLAGS /venv/bin/pip install --no-cache-dir -r /app/devito/requirements-mpi.txt" && \
-    /venv/bin/pip install --no-cache-dir -e /app/devito[extras,tests] && \
-    rm -rf ~/.cache/pip
-
-# Usefull utilities
-# Nvtop
-RUN apt-get install -y git cmake libncurses5-dev libncursesw5-dev libdrm-dev libsystemd-dev cmake && \
-    git clone https://github.com/Syllo/nvtop.git /app/nvtop && \
-    mkdir -p /app/nvtop/build && cd /app/nvtop/build && \
-    cmake .. -DNVIDIA_SUPPORT=ON -DAMDGPU_SUPPORT=ON -DINTEL_SUPPORT=ON && \
-    make && \
-    make install && \
     ln -fs /app/nvtop/build/src/nvtop /venv/bin/nvtop
 
-# Safety cleanup
-RUN apt-get clean && apt-get autoclean && apt-get autoremove -y && \
-    rm -rf /var/lib/apt/lists/*
+# Copy Devito
+ADD . /app/devito
+
+# Remove git files
+RUN rm -rf /app/devito/.git
+
+# Mpi4py
+RUN eval "$MPI4PY_FLAGS /venv/bin/pip install --no-cache-dir --verbose -r /app/devito/requirements-mpi.txt"
+
+# Devito
+RUN /venv/bin/pip install --no-cache-dir -e /app/devito[extras,tests] && rm -rf ~/.cache/pip
 
 FROM $base as user
 # COPY is much faster than RUN chown by order of magnitude so we have a final step that
 # just copies the built image into the user.
 
+# Last installs (such as gdb needed in user mode) and cleanup
+RUN apt-get update && apt install gdb -y && \
+    apt-get clean && apt-get autoclean && apt-get autoremove -y && \
+    rm -rf /var/lib/apt/lists/*
+
 # User/Group Ids
 ARG USER_ID=1000
 ARG GROUP_ID=1000

tests/test_gpu_common.py

@@ -643,7 +643,6 @@ def test_streaming_multi_input(self, opt, ntmps):
 
         assert np.all(grad.data == grad1.data)
 
-    @switchconfig(safe_math=True)
     def test_streaming_multi_input_conddim_foward(self):
         nt = 10
         grid = Grid(shape=(4, 4))
@@ -672,7 +671,6 @@ def test_streaming_multi_input_conddim_foward(self):
 
         assert np.all(v.data == v1.data)
 
-    @switchconfig(safe_math=True)
     def test_streaming_multi_input_conddim_backward(self):
         nt = 10
         grid = Grid(shape=(4, 4))