Merge pull request #50 from ROCm/hipify_runtime

Hipify runtime API

cupy/_core/_accelerator.pyx

@@ -1,6 +1,9 @@
 import os
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 
 cdef list _elementwise_accelerators = []

cupy/_core/_cub_reduction.pyx

@@ -10,7 +10,10 @@ from cupy._core cimport internal
 from cupy.cuda cimport cub
 from cupy.cuda cimport function
 from cupy.cuda cimport memory
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 import math
 import string

cupy/_core/_dtype.pyx

@@ -2,7 +2,10 @@ cimport cython  # NOQA
 import numpy
 import warnings
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 
 all_type_chars = '?bhilqBHILQefdFD'

cupy/_core/_fusion_kernel.pyx

@@ -10,7 +10,10 @@ from cupy._core.core cimport _ndarray_base
 from cupy._core cimport internal
 from cupy._core cimport _routines_manipulation as _manipulation
 from cupy_backends.cuda.api cimport driver
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 import cupy as _cupy
 from cupy._core import _dtype

cupy/_core/_kernel.pyx

@@ -27,7 +27,10 @@ from cupy._core.core cimport _ndarray_init
 from cupy._core.core cimport compile_with_cache
 from cupy._core.core cimport _ndarray_base
 from cupy._core cimport internal
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 try:
     import cupy_backends.cuda.libs.cutensor as cuda_cutensor

cupy/_core/_reduction.pyx

@@ -25,7 +25,10 @@ from cupy._core.core cimport _ndarray_base
 from cupy._core cimport internal
 from cupy.cuda cimport device
 from cupy.cuda cimport function
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 import math
 import string

cupy/_core/_routines_linalg.pyx

@@ -26,7 +26,10 @@ from cupy._core cimport _memory_range
 from cupy._core cimport _routines_manipulation as _manipulation
 from cupy._core cimport _routines_math as _math
 from cupy.cuda cimport device
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 
 cdef extern from '../../cupy_backends/cupy_complex.h':

cupy/_core/_routines_math.pyx

@@ -11,7 +11,10 @@ import cupy._core.core as core
 from cupy._core cimport internal
 from cupy import _util
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy._core cimport _accelerator
 from cupy._core._dtype cimport get_dtype
 from cupy._core.core cimport _ndarray_init

cupy/_core/core.pyx

@@ -47,7 +47,10 @@ from cupy.cuda cimport pinned_memory
 from cupy.cuda cimport memory
 from cupy.cuda cimport stream as stream_module
 from cupy_backends.cuda cimport stream as _stream_module
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 
 # If rop of cupy.ndarray is called, cupy's op is the last chance.

cupy/_core/dlpack.pyx

@@ -9,7 +9,10 @@ from libc.stdint cimport uint64_t
 from libc.stdint cimport intptr_t
 from libcpp.vector cimport vector
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda cimport stream as stream_module
 from cupy._core.core cimport _ndarray_base
 from cupy.cuda cimport memory

cupy/_core/raw.pyx

@@ -3,7 +3,10 @@ import pickle
 import cupy
 
 from cupy_backends.cuda.api cimport driver
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy.cuda.function cimport Function, Module
 
 

cupy/cuda/common.pyx

@@ -1,4 +1,7 @@
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy.cuda cimport device
 
 import numpy

cupy/cuda/cub.pyx

@@ -5,7 +5,10 @@
 from cpython cimport sequence
 from libc.stdint cimport intptr_t
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy._core.core cimport _internal_ascontiguousarray
 from cupy._core.internal cimport _contig_axes, is_in
 from cupy.cuda cimport common

cupy/cuda/device.pyx

@@ -3,7 +3,10 @@
 import threading
 
 from cupy._core import syncdetect
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda.api import runtime as runtime_module
 from cupy import _util
 

cupy/cuda/function.pyx

@@ -14,7 +14,10 @@ from libcpp cimport vector
 from cupy._core cimport _carray
 from cupy._core.core cimport _ndarray_base
 from cupy_backends.cuda.api cimport driver
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy.cuda cimport stream as stream_module
 from cupy.cuda.memory cimport MemoryPointer
 from cupy.cuda.texture cimport TextureObject, SurfaceObject

cupy/cuda/graph.pyx

@@ -1,4 +1,7 @@
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda cimport stream as stream_module
 
 

cupy/cuda/memory.pyx

@@ -20,7 +20,10 @@ from cupy.cuda cimport device
 from cupy.cuda cimport memory_hook
 from cupy.cuda cimport stream as stream_module
 from cupy_backends.cuda.api cimport driver
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 from cupy_backends.cuda.api.runtime import CUDARuntimeError
 from cupy import _util

cupy/cuda/pinned_memory.pyx

@@ -8,7 +8,10 @@ from fastrlock cimport rlock
 from cupy_backends.cuda.api import runtime
 
 from cupy._core cimport internal
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy import _util
 
 

cupy/cuda/stream.pyx

@@ -1,6 +1,9 @@
 import threading
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda cimport stream as backends_stream
 from cupy.cuda cimport graph
 

cupy/cuda/texture.pyx

@@ -5,7 +5,10 @@ import numpy
 
 from cupy._core.core cimport _ndarray_base
 from cupy._core.core cimport _internal_ascontiguousarray
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda.api.runtime cimport Array, \
     ChannelFormatDesc, ChannelFormatKind, \
     Memcpy3DParms, MemoryKind, PitchedPtr, ResourceDesc, ResourceType, \

cupy/fft/_cache.pyx

@@ -4,7 +4,10 @@ import gc
 import warnings
 import weakref
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 
 import threading
 

cupy/fft/_callback.pyx

@@ -1,6 +1,9 @@
 from libc.stdint cimport intptr_t
 
-from cupy_backends.cuda.api cimport runtime
+IF CUPY_HIP_VERSION > 0:
+    from cupy_backends.cuda.api cimport runtime_hip as runtime
+ELSE:
+    from cupy_backends.cuda.api cimport runtime
 from cupy._core.core cimport _ndarray_base
 from cupy.cuda.device cimport get_compute_capability
 

cupy_backends/cuda/api/_runtime_enum.pxd

@@ -33,9 +33,6 @@ cpdef enum:
     cudaLimitStackSize = 0x00
     cudaLimitPrintfFifoSize = 0x01
     cudaLimitMallocHeapSize = 0x02
-    cudaLimitDevRuntimeSyncDepth = 0x03
-    cudaLimitDevRuntimePendingLaunchCount = 0x04
-    cudaLimitMaxL2FetchGranularity = 0x05
 
     # cudaChannelFormatKind
     cudaChannelFormatKindSigned = 0
@@ -94,7 +91,6 @@ cpdef enum:
     # cudaMemLocationType
     cudaMemLocationTypeDevice = 1
 
-
 # This was a legacy mistake: the prefix "cuda" should have been removed
 # so that we can directly assign their C counterparts here. Now because
 # of backward compatibility and no flexible Cython macro (IF/ELSE), we
@@ -228,6 +224,10 @@ IF CUPY_HIP_VERSION > 0:
 ELSE:
     # For CUDA/RTD
     cpdef enum:
+        cudaLimitDevRuntimeSyncDepth = 0x03
+        cudaLimitDevRuntimePendingLaunchCount = 0x04
+        cudaLimitMaxL2FetchGranularity = 0x05
+
         cudaDevAttrMaxThreadsPerBlock = 1
         cudaDevAttrMaxBlockDimX
         cudaDevAttrMaxBlockDimY
@@ -343,3 +343,8 @@ ELSE:
         cudaDevAttrGPUDirectRDMAFlushWritesOptions
         cudaDevAttrGPUDirectRDMAWritesOrdering
         cudaDevAttrMemoryPoolSupportedHandleTypes
+
+IF CUPY_HIP_VERSION > 0:
+    cpdef enum:
+        hipIpcMemLazyEnablePeerAccess = 1
+        hipMemAttachGlobal = 1

cupy_backends/cuda/api/_runtime_typedef.pxi

@@ -116,6 +116,9 @@ cdef extern from *:
     ctypedef int MemAllocationType 'cudaMemAllocationType'
     ctypedef int MemAllocationHandleType 'cudaMemAllocationHandleType'
     ctypedef int MemLocationType 'cudaMemLocationType'
+    IF CUPY_HIP_VERSION > 0:
+        ctypedef int MemoryType 'hipMemoryType'
+
     IF CUPY_CUDA_VERSION > 0:
         # This is for the annoying nested struct, which is not
         # perfectly supprted in Cython