Update/dpctl memcpy async (#529)

* Initial update to accommodate async dpctl memcpy
* Update dpnp implementations
* Fix dpctl version
* Use dpctl 0.10*

Co-authored-by: Sergey Pokhodenko <sergey.pokhodenko@intel.com>

.github/workflows/conda-package.yml

@@ -86,7 +86,7 @@ jobs:
             integration_channels: -c dppy/label/dev
             artifact_name: -c dppy_label_dev
             experimental: false  # current stable
-            dependencies: dpctl=0.9.0=*_21 dpnp=0.7.1=*_41
+            dependencies: dpnp=0.7.1=*_41
     continue-on-error: ${{ matrix.experimental }}
     env:
       # conda-forge: llvm-spirv 11 not on intel channel yet

conda-recipe/meta.yaml

@@ -19,13 +19,13 @@ requirements:
         - setuptools
         - cython
         - numba 0.54*
-        - dpctl >=0.9*
+        - dpctl >=0.10*
         - dpnp >=0.7*  # [linux]
         - wheel
     run:
         - python
         - numba 0.54*
-        - dpctl >=0.9*
+        - dpctl >=0.10*
         - spirv-tools
         - llvm-spirv 11.*
         - dpnp >=0.7* # [linux]

numba_dppy/dpctl_functions.py

@@ -29,18 +29,39 @@ def dpctl_malloc_shared():
 
 
 def dpctl_queue_memcpy():
-    ret_type = types.void
+    ret_type = types.voidptr
     sig = signature(ret_type, types.voidptr, types.voidptr, types.voidptr, types.int64)
     return types.ExternalFunction("DPCTLQueue_Memcpy", sig)
 
 
+def dpctl_event_wait():
+    ret_type = types.voidptr
+    sig = signature(ret_type, types.voidptr)
+    return types.ExternalFunction("DPCTLEvent_Wait", sig)
+
+
+def dpctl_event_delete():
+    ret_type = types.void
+    sig = signature(ret_type, types.voidptr)
+    return types.ExternalFunction("DPCTLEvent_Delete", sig)
+
+
 def dpctl_free_with_queue():
     ret_type = types.void
     sig = signature(ret_type, types.voidptr, types.voidptr)
     return types.ExternalFunction("DPCTLfree_with_queue", sig)
 
 
+def dpctl_queue_wait():
+    ret_type = types.void
+    sig = signature(ret_type, types.voidptr)
+    return types.ExternalFunction("DPCTLQueue_Wait", sig)
+
+
 get_current_queue = dpctl_get_current_queue()
 malloc_shared = dpctl_malloc_shared()
 queue_memcpy = dpctl_queue_memcpy()
 free_with_queue = dpctl_free_with_queue()
+event_wait = dpctl_event_wait()
+event_delete = dpctl_event_delete()
+queue_wait = dpctl_queue_wait()

numba_dppy/dpnp_glue/dpnp_array_creations_impl.py

@@ -31,15 +31,21 @@ def common_impl(a, b, out, dpnp_func, PRINT_DEBUG):
     sycl_queue = dpctl_functions.get_current_queue()
 
     b_usm = dpctl_functions.malloc_shared(b.size * b.itemsize, sycl_queue)
-    dpctl_functions.queue_memcpy(sycl_queue, b_usm, b.ctypes, b.size * b.itemsize)
+    event = dpctl_functions.queue_memcpy(
+        sycl_queue, b_usm, b.ctypes, b.size * b.itemsize
+    )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
     dpnp_func(out_usm, b_usm, a.size)
 
-    dpctl_functions.queue_memcpy(
+    event = dpctl_functions.queue_memcpy(
         sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
     )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     dpctl_functions.free_with_queue(b_usm, sycl_queue)
     dpctl_functions.free_with_queue(out_usm, sycl_queue)
@@ -58,15 +64,21 @@ def common_shape_impl(a, out, dpnp_func, PRINT_DEBUG):
     sycl_queue = dpctl_functions.get_current_queue()
 
     a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-    dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+    event = dpctl_functions.queue_memcpy(
+        sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+    )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
     dpnp_func(a_usm, out_usm, a.shapeptr, a.ndim)
 
-    dpctl_functions.queue_memcpy(
+    event = dpctl_functions.queue_memcpy(
         sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
     )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     dpctl_functions.free_with_queue(a_usm, sycl_queue)
     dpctl_functions.free_with_queue(out_usm, sycl_queue)
@@ -203,16 +215,22 @@ def dpnp_impl(a, b):
         sycl_queue = dpctl_functions.get_current_queue()
 
         b_usm = dpctl_functions.malloc_shared(b.size * b.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(sycl_queue, b_usm, b.ctypes, b.size * b.itemsize)
+        event = dpctl_functions.queue_memcpy(
+            sycl_queue, b_usm, b.ctypes, b.size * b.itemsize
+        )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         out = np.arange(0, a.size, 1, res_dtype)
         out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
         dpnp_func(b_usm, out_usm, a.size)
 
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         dpctl_functions.free_with_queue(b_usm, sycl_queue)
         dpctl_functions.free_with_queue(out_usm, sycl_queue)

numba_dppy/dpnp_glue/dpnp_array_ops_impl.py

@@ -30,15 +30,21 @@ def common_impl(a, out, dpnp_func, print_debug):
 
     sycl_queue = dpctl_functions.get_current_queue()
     a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-    dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+    event = dpctl_functions.queue_memcpy(
+        sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+    )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     out_usm = dpctl_functions.malloc_shared(a.itemsize, sycl_queue)
 
     dpnp_func(a_usm, out_usm, a.size)
 
-    dpctl_functions.queue_memcpy(
+    event = dpctl_functions.queue_memcpy(
         sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
     )
+    dpctl_functions.event_wait(event)
+    dpctl_functions.event_delete(event)
 
     dpctl_functions.free_with_queue(a_usm, sycl_queue)
     dpctl_functions.free_with_queue(out_usm, sycl_queue)
@@ -130,16 +136,22 @@ def dpnp_impl(a):
         sycl_queue = dpctl_functions.get_current_queue()
 
         a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+        event = dpctl_functions.queue_memcpy(
+            sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+        )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         out = np.arange(0, a.size, 1, res_dtype)
         out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
         dpnp_func(a_usm, out_usm, a.size)
 
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         dpctl_functions.free_with_queue(a_usm, sycl_queue)
         dpctl_functions.free_with_queue(out_usm, sycl_queue)
@@ -180,16 +192,22 @@ def dpnp_impl(a):
         sycl_queue = dpctl_functions.get_current_queue()
 
         a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+        event = dpctl_functions.queue_memcpy(
+            sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+        )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         out = np.arange(0, a.size, 1, res_dtype)
         out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
         dpnp_func(a_usm, out_usm, a.size)
 
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         dpctl_functions.free_with_queue(a_usm, sycl_queue)
         dpctl_functions.free_with_queue(out_usm, sycl_queue)
@@ -228,21 +246,29 @@ def dpnp_impl(a, ind):
         sycl_queue = dpctl_functions.get_current_queue()
 
         a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+        event = dpctl_functions.queue_memcpy(
+            sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+        )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         ind_usm = dpctl_functions.malloc_shared(ind.size * ind.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, ind_usm, ind.ctypes, ind.size * ind.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         out = np.arange(0, ind.size, 1, res_dtype).reshape(ind.shape)
         out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
         dpnp_func(a_usm, ind_usm, out_usm, ind.size)
 
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         dpctl_functions.free_with_queue(a_usm, sycl_queue)
         dpctl_functions.free_with_queue(ind_usm, sycl_queue)

numba_dppy/dpnp_glue/dpnp_indexing.py

@@ -83,15 +83,21 @@ def dpnp_impl(a, offset=0):
         sycl_queue = dpctl_functions.get_current_queue()
 
         a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
-        dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes, a.size * a.itemsize)
+        event = dpctl_functions.queue_memcpy(
+            sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
+        )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize, sycl_queue)
 
         dpnp_func(a_usm, out_usm, offset, a.shapeptr, out.shapeptr, out.ndim)
 
-        dpctl_functions.queue_memcpy(
+        event = dpctl_functions.queue_memcpy(
             sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
         )
+        dpctl_functions.event_wait(event)
+        dpctl_functions.event_delete(event)
 
         dpctl_functions.free_with_queue(a_usm, sycl_queue)
         dpctl_functions.free_with_queue(out_usm, sycl_queue)