Initial FPGA Support

.github/workflows/cpu-ci.yml

@@ -54,7 +54,7 @@ jobs:
     - name: Test with pytest
       env:
         ORT_RELEASE: ${{ github.workspace }}/onnxruntime-daceml-patched
-        PYTEST_ARGS: --cov=daceml --cov-report=term --cov-report xml --cov-config=.coveragerc -m "not slow" -m "not gpu"
+        PYTEST_ARGS: --cov=daceml --cov-report=term --cov-report xml --cov-config=.coveragerc -m "not slow and not fpga and not gpu"
       run: make test
 
     - name: Test with doctest

.github/workflows/docs-no-trigger.yml

@@ -0,0 +1,41 @@
+name: Docs
+
+on:
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  build-doc:
+    runs-on: [self-hosted, linux, gpu]
+    env:
+      ORT_ROOT: '/opt/onnxruntime'
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+          submodules: 'recursive'
+
+      - name: Install dependencies
+        env:
+          UPDATE_PIP: 'true'
+        run: |
+          rm -rf .dacecache tests/.dacecache
+          . /opt/setupenv
+          make clean install
+
+      - name: Build docs
+        run: make doc
+        env:
+          DACEML_DOC_BUILD_FPGA: 'True'
+          DACE_compiler_fpga_vendor: intel_fpga
+          DACE_compiler_use_cache: 0
+          DACE_compiler_default_data_types: C
+          DACE_compiler_intel_fpga_mode: emulator
+          DACE_optimizer_transform_on_call: 0
+          DACE_optimizer_autooptimize: 0
+
+      - uses: actions/upload-artifact@v2
+        with:
+          name: auto_examples_${{ github.sha }}
+          path: doc/auto_examples/

.github/workflows/docs.yml

@@ -27,6 +27,7 @@ jobs:
       - name: Build docs
         run: make doc
         env:
+          DACEML_DOC_BUILD_FPGA: 'True'
           DACEML_DOC_BUILD_CUDA: 'True'
 
       - uses: actions/upload-artifact@v2

.github/workflows/fpga-ci.yml

@@ -0,0 +1,41 @@
+name: FPGA CI
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  test-fpga:
+    runs-on: [self-hosted, linux, intel-fpga]
+    env:
+      ORT_ROOT: '/opt/onnxruntime'
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+          submodules: 'recursive'
+
+      - name: Install dependencies
+        env:
+          UPDATE_PIP: 'true'
+        run: |
+          rm -rf .dacecache tests/.dacecache
+          . /opt/setupenv
+          make clean install
+
+      - name: Run Intel FPGA tests
+        env:
+          PYTEST_ARGS: --cov=daceml --cov-report=term --cov-report xml --cov-config=.coveragerc -s -m "not slow and fpga"
+          DACE_compiler_fpga_vendor: intel_fpga
+          DACE_compiler_use_cache: 0
+          DACE_compiler_default_data_types: C
+          DACE_compiler_intel_fpga_mode: emulator
+          DACE_optimizer_transform_on_call: 0
+          DACE_optimizer_autooptimize: 0
+        run: make test-intel-fpga
+
+      - name: Upload coverage
+        run: make codecov

.github/workflows/gpu-ci.yml

@@ -28,7 +28,7 @@ jobs:
 
       - name: Test with pytest
         env:
-          PYTEST_ARGS: --cov=daceml --cov-report=term --cov-report xml --cov-config=.coveragerc --gpu-only -m "not slow"
+          PYTEST_ARGS: --cov=daceml --cov-report=term --cov-report xml --cov-config=.coveragerc --gpu-only -m "not slow and not fpga"
         run: make test
 
       - name: Upload coverage

Makefile

@@ -1,5 +1,6 @@
 VENV_PATH ?= venv
 PYTHON ?= python
+PYTHON_BINARY ?= python
 PYTEST ?= pytest
 PIP ?= pip
 YAPF ?= yapf
@@ -53,6 +54,9 @@ test-parallel:
 test-gpu: 
 	$(ACTIVATE) $(PYTEST) $(PYTEST_ARGS) tests --gpu
 
+test-intel-fpga:
+	$(ACTIVATE) $(PYTEST) $(PYTEST_ARGS) tests/pytorch/fpga/
+
 codecov:
 	curl -s https://codecov.io/bash | bash
 

daceml/autodiff/backward_pass_generator.py

@@ -382,8 +382,13 @@ def _expand_nodes(self, subgraph: dstate.StateSubgraphView) -> bool:
 
             # only check others if we didn't break out of the above loop
             if isinstance(node, ONNXOp):
-                for impl in ONNXForward.registered_implementations(
-                        node.schema.name):
+                impls = ONNXForward.registered_implementations(
+                    node.schema.name)
+
+                # order the implementations so that implementations containing "pure" are tried first
+                impls = [i for name, i in impls if "pure" in name
+                         ] + [i for name, i in impls if "pure" not in name]
+                for impl in impls:
                     if impl.forward_can_be_applied(node, state, self.sdfg):
                         # try to apply the expansion
                         class Expansion(xf.ExpandTransformation):
@@ -403,7 +408,7 @@ def annotates_memlets() -> bool:
                                            verify=False,
                                            _match_node=node)
                         expanded_something = True
-                        continue
+                        break
 
             # This could later on be changed to check if the expansion is differentiable and if not, move
             # on to the next expansion. For now we will just apply the first one that matches, prioritizing ones that

daceml/onnx/binary_utilities/op_checker.py

@@ -21,7 +21,7 @@ def check_op(sdfg, state, node, cuda=False) -> Tuple[List[bool], List[bool]]:
     log.debug(f"Checking node {node}")
 
     with ORTCAPIInterface() as api,\
-            KernelSession(api) as session,\
+            KernelSession(api, cuda=cuda) as session,\
             ExecutableKernelContext(api, session, node.name, node.schema.name) as context:
 
         for attribute, onnx_attribute in node.schema.attributes.items():

daceml/onnx/forward_implementation_abc.py

@@ -39,12 +39,15 @@ def forward(node: ONNXOp, state: SDFGState,
         """
         ...
 
-    @staticmethod
-    def registered_implementations(op_name: str) -> typing.List["ONNXForward"]:
+    @classmethod
+    def registered_implementations(
+            cls,
+            op_name: str) -> typing.List[typing.Tuple[str, "ONNXForward"]]:
         impls = []
-        for impl, args in ONNXForward.extensions().items():
+        for impl, args in cls.extensions().items():
             if "op" in args and args["op"] == op_name:
-                impls.append(impl)
+                impls.append((args["name"], impl))
+
         return impls
 
 

daceml/onnx/nodes/onnx_op.py

@@ -376,11 +376,13 @@ def validate(self, sdfg: SDFG, state: SDFGState):
 
             edge_data = edge.data.data
             edge_dtype = sdfg.arrays[edge_data].dtype
+            # edge_dtype can be a vector type
             if matched.param_type == ONNXParameterType.Variadic and not matched.homogeneous:
                 # non homogeneous parameters don't need to be consistent
                 pass
-            elif matched.type_str in assigned_params and assigned_params[
-                    matched.type_str] != edge_dtype:
+            elif matched.type_str in assigned_params and (
+                    assigned_params[matched.type_str] != edge_dtype and
+                    assigned_params[matched.type_str] != edge_dtype.base_type):
                 raise ValueError(
                     "Could not solve type constraints;"
                     " excepted type '{expected}' for {param_type} '{conn_name}', got type '{actual}'"
@@ -391,14 +393,14 @@ def validate(self, sdfg: SDFG, state: SDFGState):
 
             # otherwise, matched.type_str was not assigned a type yet: try to assign it
             cons = self.schema.type_constraints[matched.type_str]
-            if edge_dtype not in cons.types:
+            if edge_dtype not in cons.types and edge_dtype.base_type not in cons.types:
                 raise ValueError(
                     "Expected type in '{possible}' for {param_type} '{conn_name}', got type '{actual}'"
                     .format(possible=cons.types,
                             param_type="input" if is_input else "output",
                             conn_name=matched.name,
                             actual=edge_dtype))
-            assigned_params[matched.type_str] = edge_dtype
+            assigned_params[matched.type_str] = edge_dtype.base_type
 
         # check that we have all required attributes
         ##########################################

daceml/onnx/op_implementations/__init__.py

@@ -1,4 +1,5 @@
 from .utils import *
 from .pure_implementations import *
+from .fpga_implementations import *
 from .img_op_implementations import *
 from .cudnn_implementations import *