Update runtime for ROCm6.0 (SWDEV-434155)

.github/workflows/pretest-rocm-test.sh

@@ -7,7 +7,7 @@ DEBIAN_FRONTEND=noninteractive apt-get -y install python3.9-dev python3-pip
 
 hipconfig
 
-python3.9 -m pip install -U pip wheel
+pip3 install -U pip wheel
 
 export ROCM_HOME="/opt/rocm"
 export HCC_AMDGPU_TARGET="gfx900"

cupy_backends/cuda/api/_runtime_typedef.pxi

@@ -131,13 +131,13 @@ cdef extern from *:
         ctypedef struct _MemPoolProps 'cudaMemPoolProps':
             pass  # for HIP & RTD
 
-    IF CUPY_CUDA_VERSION > 0:
+    IF CUPY_CUDA_VERSION > 0 or CUPY_HIP_VERSION > 60000000:
         ctypedef struct _PointerAttributes 'cudaPointerAttributes':
             int type
             int device
             void* devicePointer
             void* hostPointer
-    ELIF CUPY_HIP_VERSION > 0:
+    ELIF 0 < CUPY_HIP_VERSION < 60000000:
         ctypedef struct _PointerAttributes 'cudaPointerAttributes':
             int memoryType
             int device

cupy_backends/cuda/api/runtime.pyx

@@ -295,7 +295,6 @@ cpdef getDeviceProperties(int device):
         properties['clockInstructionRate'] = props.clockInstructionRate
         properties['maxSharedMemoryPerMultiProcessor'] = (
             props.maxSharedMemoryPerMultiProcessor)
-        properties['gcnArch'] = props.gcnArch
         properties['hdpMemFlushCntl'] = <intptr_t>(props.hdpMemFlushCntl)
         properties['hdpRegFlushCntl'] = <intptr_t>(props.hdpRegFlushCntl)
         properties['memPitch'] = props.memPitch
@@ -328,6 +327,8 @@ cpdef getDeviceProperties(int device):
         arch['has3dGrid'] = props.arch.has3dGrid
         arch['hasDynamicParallelism'] = props.arch.hasDynamicParallelism
         properties['arch'] = arch
+    IF 0 < CUPY_HIP_VERSION < 310:  # gcnArchName used after ROCm 3.1+
+        properties['gcnArch'] = props.gcnArch
     IF CUPY_HIP_VERSION >= 310:
         properties['gcnArchName'] = props.gcnArchName
         properties['asicRevision'] = props.asicRevision
@@ -697,13 +698,13 @@ cpdef PointerAttributes pointerGetAttributes(intptr_t ptr):
     cdef _PointerAttributes attrs
     status = cudaPointerGetAttributes(&attrs, <void*>ptr)
     check_status(status)
-    IF CUPY_CUDA_VERSION > 0:
+    IF CUPY_CUDA_VERSION > 0 or CUPY_HIP_VERSION > 60000000:
         return PointerAttributes(
             attrs.device,
             <intptr_t>attrs.devicePointer,
             <intptr_t>attrs.hostPointer,
             attrs.type)
-    ELIF CUPY_HIP_VERSION > 0:
+    ELIF 0 < CUPY_HIP_VERSION < 60000000:
         return PointerAttributes(
             attrs.device,
             <intptr_t>attrs.devicePointer,

cupy_backends/hip/cupy_hip_runtime.h

@@ -269,6 +269,21 @@ cudaError_t cudaMemPrefetchAsync(const void *devPtr, size_t count,
 cudaError_t cudaPointerGetAttributes(cudaPointerAttributes *attributes,
                                      const void* ptr) {
     cudaError_t status = hipPointerGetAttributes(attributes, ptr);
+#if HIP_VERSION >= 60000000
+    if (status == cudaSuccess) {
+        switch (attributes->type) {
+            case 0 /* hipMemoryTypeHost */:
+                attributes->type = (hipMemoryType)1; /* cudaMemoryTypeHost */
+                return status;
+            case 1 /* hipMemoryTypeDevice */:
+                attributes->type = (hipMemoryType)2; /* cudaMemoryTypeDevice */
+                return status;
+            default:
+                /* we don't care the rest of possibilities */
+                return status;
+        }
+    }
+#else
     if (status == cudaSuccess) {
         switch (attributes->memoryType) {
             case 0 /* hipMemoryTypeHost */:
@@ -281,7 +296,9 @@ cudaError_t cudaPointerGetAttributes(cudaPointerAttributes *attributes,
                 /* we don't care the rest of possibilities */
                 return status;
         }
-    } else {
+    }
+#endif
+    else {
         return status;
     }
 }