#include <hip/hip_runtime.h>#include <hipcub/hipcub.hpp>#include <hipcub/device/device_radix_sort.hpp>#include <omp.h>
Go to the source code of this file.
Classes | |
| class | arch::buf< T > |
Namespaces | |
| namespace | arch |
Macros | |
| #define | gpuGetLastError hipGetLastError |
| #define | gpuGetErrorString hipGetErrorString |
| #define | gpuPeekAtLastError hipPeekAtLastError |
| #define | gpuSetDevice hipSetDevice |
| #define | gpuGetDevice hipGetDevice |
| #define | gpuGetDeviceCount hipGetDeviceCount |
| #define | gpuGetDeviceProperties hipGetDeviceProperties |
| #define | gpuDeviceGetAttribute hipDeviceGetAttribute |
| #define | gpuDeviceSynchronize hipDeviceSynchronize |
| #define | gpuDeviceReset hipDeviceReset |
| #define | gpuCpuDeviceId hipCpuDeviceId |
| #define | gpuMemGetInfo hipMemGetInfo |
| #define | gpuDevAttrMaxBlocksPerMultiprocessor hipDeviceAttributeMaxBlocksPerMultiProcessor |
| #define | gpuFree hipFree |
| #define | gpuFreeHost hipHostFree |
| #define | gpuFreeAsync(ptr, stream) |
| #define | gpuMalloc hipMalloc |
| #define | gpuMallocHost hipHostMalloc |
| #define | gpuMallocAsync hipMallocAsync |
| #define | gpuMallocManaged hipMallocManaged |
| #define | gpuHostAlloc hipHostMalloc |
| #define | gpuHostAllocPortable hipHostAllocPortable |
| #define | gpuMemcpy hipMemcpy |
| #define | gpuMemcpyAsync hipMemcpyAsync |
| #define | gpuMemset hipMemset |
| #define | gpuMemsetAsync hipMemsetAsync |
| #define | gpuMemAdviseSetAccessedBy hipMemAdviseSetAccessedBy |
| #define | gpuMemAdviseSetPreferredLocation hipMemAdviseSetPreferredLocation |
| #define | gpuMemAttachSingle hipMemAttachSingle |
| #define | gpuMemAttachGlobal hipMemAttachGlobal |
| #define | gpuMemPrefetchAsync hipMemPrefetchAsync |
| #define | gpuStreamCreate hipStreamCreate |
| #define | gpuHostRegister hipHostRegister |
| #define | gpuHostRegisterPortable hipHostRegisterPortable |
| #define | gpuStreamDestroy hipStreamDestroy |
| #define | gpuStreamWaitEvent hipStreamWaitEvent |
| #define | gpuStreamSynchronize hipStreamSynchronize |
| #define | gpuStreamAttachMemAsync hipStreamAttachMemAsync |
| #define | gpuDeviceGetStreamPriorityRange hipDeviceGetStreamPriorityRange |
| #define | gpuStreamCreateWithPriority hipStreamCreateWithPriority |
| #define | gpuStreamDefault hipStreamDefault |
| #define | gpuEventCreate hipEventCreate |
| #define | gpuEventCreateWithFlags hipEventCreateWithFlags |
| #define | gpuEventDestroy hipEventDestroy |
| #define | gpuEventQuery hipEventQuery |
| #define | gpuEventRecord hipEventRecord |
| #define | gpuEventSynchronize hipEventSynchronize |
| #define | gpuEventElapsedTime hipEventElapsedTime |
| #define | gpuError_t hipError_t |
| #define | gpuSuccess hipSuccess |
| #define | gpuStream_t hipStream_t |
| #define | gpuDeviceProp hipDeviceProp_t |
| #define | gpuEvent_t hipEvent_t |
| #define | gpuEventDefault hipEventDefault |
| #define | gpuEventBlockingSync hipEventBlockingSync |
| #define | gpuEventDisableTiming hipEventDisableTiming |
| #define | gpuMemcpyKind hipMemcpyKind |
| #define | gpuMemcpyDeviceToHost hipMemcpyDeviceToHost |
| #define | gpuMemcpyHostToDevice hipMemcpyHostToDevice |
| #define | gpuMemcpyDeviceToDevice hipMemcpyDeviceToDevice |
| #define | gpuMemcpyHostToHost hipMemcpyHostToHost |
| #define | gpuMemcpyToSymbol hipMemcpyToSymbol |
| #define | gpuKernelBallot(mask, input) |
| #define | gpuKernelAny(mask, input) |
| #define | gpuKernelShfl(input, source, mask) |
| #define | gpuKernelShflDown(val, offset) |
| #define | gpuWarpSync() |
| #define | ARCH_LOOP_LAMBDA [=] __host__ __device__ |
| #define | ARCH_INNER_BODY2(i, j, aggregate) |
| #define | ARCH_INNER_BODY3(i, j, k, aggregate) |
| #define | ARCH_INNER_BODY4(i, j, k, l, aggregate) |
| #define | HIPCUB_STDERR |
| #define | ARCH_BLOCKSIZE_R 512 |
| #define | ARCH_BLOCKSIZE_R_SMALL 64 |
| #define | GPUTHREADS (64) |
| #define | WARPSPERBLOCK (16) |
| #define | FULL_MASK 0xffffffffffffffff |
| #define | CHK_ERR(err) |
Functions | |
| static void | hip_error (hipError_t err, const char *file, int line) |
| __device__ static __forceinline__ void | arch::atomicMax (double *address, double val2) |
| __device__ static __forceinline__ void | arch::atomicMin (double *address, double val2) |
| __device__ static __forceinline__ void | arch::atomicMax (float *address, float val2) |
| __device__ static __forceinline__ void | arch::atomicMin (float *address, float val2) |
| __host__ static __forceinline__ void | arch::device_mempool_check (uint64_t threshold_new) |
| __host__ static __forceinline__ void * | arch::allocate (size_t bytes) |
| __host__ static __forceinline__ void * | arch::allocate (size_t bytes, hipStream_t stream) |
| template<typename T> | |
| __host__ static __forceinline__ void | arch::free (T *ptr) |
| template<typename T> | |
| __host__ static __forceinline__ void | arch::free (T *ptr, hipStream_t stream) |
| template<typename T> | |
| static __forceinline__ void | arch::memcpy_h2d (T *dst, T *src, size_t bytes) |
| template<typename T> | |
| static __forceinline__ void | arch::memcpy_h2d (T *dst, T *src, size_t bytes, hipStream_t stream) |
| template<typename T> | |
| static __forceinline__ void | arch::memcpy_d2h (T *dst, T *src, size_t bytes) |
| template<typename T> | |
| static __forceinline__ void | arch::memcpy_d2h (T *dst, T *src, size_t bytes, hipStream_t stream) |
| template<typename T> | |
| static __forceinline__ void | arch::host_register (T *ptr, size_t bytes) |
| template<typename T> | |
| static __forceinline__ void | arch::host_unregister (T *ptr) |
| template<typename Lambda, typename T> | |
| __device__ static __forceinline__ void | arch::lambda_eval (const uint(&idx)[1], T *__restrict__ thread_data, Lambda loop_body) |
| template<typename Lambda, typename T> | |
| __device__ static __forceinline__ void | arch::lambda_eval (const uint(&idx)[2], T *__restrict__ thread_data, Lambda loop_body) |
| template<typename Lambda, typename T> | |
| __device__ static __forceinline__ void | arch::lambda_eval (const uint(&idx)[3], T *__restrict__ thread_data, Lambda loop_body) |
| template<typename Lambda, typename T> | |
| __device__ static __forceinline__ void | arch::lambda_eval (const uint(&idx)[4], T *__restrict__ thread_data, Lambda loop_body) |
| template<uint NDim, typename Lambda, typename T> | |
| __device__ static __forceinline__ void | arch::loop_eval (const uint idx_glob, const uint *__restrict__ lims, T *__restrict__ thread_data, Lambda loop_body) |
| template<uint Blocksize, reduce_op Op, uint NDim, uint NReduStatic, typename Lambda, typename T> | |
| static __global__ void | arch::__launch_bounds__ (ARCH_BLOCKSIZE_R) reduction_kernel(Lambda loop_body |
| arch::if (Op==reduce_op::null) | |
| arch::for (uint i=0;i< n_reductions;i++) | |
| arch::if (idx_glob< n_total) | |
| template<reduce_op Op, uint NReduStatic, uint NDim, typename Lambda, typename T> | |
| static __forceinline__ void | arch::parallel_reduce_driver (const uint(&limits)[NDim], Lambda loop_body, T *sum, const uint n_redu_dynamic) |
Variables | |
| hipStream_t | gpuStreamList [] |
Macro Definition Documentation
◆ ARCH_BLOCKSIZE_R
| #define ARCH_BLOCKSIZE_R 512 |
Definition at line 104 of file arch_device_hip.h.
◆ ARCH_BLOCKSIZE_R_SMALL
| #define ARCH_BLOCKSIZE_R_SMALL 64 |
Definition at line 105 of file arch_device_hip.h.
◆ ARCH_INNER_BODY2
Definition at line 96 of file arch_device_hip.h.
◆ ARCH_INNER_BODY3
Definition at line 97 of file arch_device_hip.h.
◆ ARCH_INNER_BODY4
Definition at line 98 of file arch_device_hip.h.
◆ ARCH_LOOP_LAMBDA
| #define ARCH_LOOP_LAMBDA [=] __host__ __device__ |
Definition at line 95 of file arch_device_hip.h.
◆ CHK_ERR
| #define CHK_ERR | ( | err | ) |
Definition at line 119 of file arch_device_hip.h.
◆ FULL_MASK
| #define FULL_MASK 0xffffffffffffffff |
Definition at line 114 of file arch_device_hip.h.
◆ gpuCpuDeviceId
| #define gpuCpuDeviceId hipCpuDeviceId |
Definition at line 24 of file arch_device_hip.h.
◆ gpuDevAttrMaxBlocksPerMultiprocessor
| #define gpuDevAttrMaxBlocksPerMultiprocessor hipDeviceAttributeMaxBlocksPerMultiProcessor |
Definition at line 27 of file arch_device_hip.h.
◆ gpuDeviceGetAttribute
| #define gpuDeviceGetAttribute hipDeviceGetAttribute |
Definition at line 21 of file arch_device_hip.h.
◆ gpuDeviceGetStreamPriorityRange
| #define gpuDeviceGetStreamPriorityRange hipDeviceGetStreamPriorityRange |
Definition at line 57 of file arch_device_hip.h.
◆ gpuDeviceProp
| #define gpuDeviceProp hipDeviceProp_t |
Definition at line 74 of file arch_device_hip.h.
◆ gpuDeviceReset
| #define gpuDeviceReset hipDeviceReset |
Definition at line 23 of file arch_device_hip.h.
◆ gpuDeviceSynchronize
| #define gpuDeviceSynchronize hipDeviceSynchronize |
Definition at line 22 of file arch_device_hip.h.
◆ gpuError_t
| #define gpuError_t hipError_t |
Definition at line 70 of file arch_device_hip.h.
◆ gpuEvent_t
| #define gpuEvent_t hipEvent_t |
Definition at line 76 of file arch_device_hip.h.
◆ gpuEventBlockingSync
| #define gpuEventBlockingSync hipEventBlockingSync |
Definition at line 78 of file arch_device_hip.h.
◆ gpuEventCreate
| #define gpuEventCreate hipEventCreate |
Definition at line 61 of file arch_device_hip.h.
◆ gpuEventCreateWithFlags
| #define gpuEventCreateWithFlags hipEventCreateWithFlags |
Definition at line 62 of file arch_device_hip.h.
◆ gpuEventDefault
| #define gpuEventDefault hipEventDefault |
Definition at line 77 of file arch_device_hip.h.
◆ gpuEventDestroy
| #define gpuEventDestroy hipEventDestroy |
Definition at line 63 of file arch_device_hip.h.
◆ gpuEventDisableTiming
| #define gpuEventDisableTiming hipEventDisableTiming |
Definition at line 79 of file arch_device_hip.h.
◆ gpuEventElapsedTime
| #define gpuEventElapsedTime hipEventElapsedTime |
Definition at line 67 of file arch_device_hip.h.
◆ gpuEventQuery
| #define gpuEventQuery hipEventQuery |
Definition at line 64 of file arch_device_hip.h.
◆ gpuEventRecord
| #define gpuEventRecord hipEventRecord |
Definition at line 65 of file arch_device_hip.h.
◆ gpuEventSynchronize
| #define gpuEventSynchronize hipEventSynchronize |
Definition at line 66 of file arch_device_hip.h.
◆ gpuFree
| #define gpuFree hipFree |
Definition at line 29 of file arch_device_hip.h.
◆ gpuFreeAsync
| #define gpuFreeAsync | ( | ptr, | |
| stream ) |
Definition at line 31 of file arch_device_hip.h.
◆ gpuFreeHost
| #define gpuFreeHost hipHostFree |
Definition at line 30 of file arch_device_hip.h.
◆ gpuGetDevice
| #define gpuGetDevice hipGetDevice |
Definition at line 18 of file arch_device_hip.h.
◆ gpuGetDeviceCount
| #define gpuGetDeviceCount hipGetDeviceCount |
Definition at line 19 of file arch_device_hip.h.
◆ gpuGetDeviceProperties
| #define gpuGetDeviceProperties hipGetDeviceProperties |
Definition at line 20 of file arch_device_hip.h.
◆ gpuGetErrorString
| #define gpuGetErrorString hipGetErrorString |
Definition at line 14 of file arch_device_hip.h.
◆ gpuGetLastError
| #define gpuGetLastError hipGetLastError |
Definition at line 13 of file arch_device_hip.h.
◆ gpuHostAlloc
| #define gpuHostAlloc hipHostMalloc |
Definition at line 36 of file arch_device_hip.h.
◆ gpuHostAllocPortable
| #define gpuHostAllocPortable hipHostAllocPortable |
Definition at line 37 of file arch_device_hip.h.
◆ gpuHostRegister
| #define gpuHostRegister hipHostRegister |
Definition at line 50 of file arch_device_hip.h.
◆ gpuHostRegisterPortable
| #define gpuHostRegisterPortable hipHostRegisterPortable |
Definition at line 51 of file arch_device_hip.h.
◆ gpuKernelAny
| #define gpuKernelAny | ( | mask, | |
| input ) |
Definition at line 89 of file arch_device_hip.h.
◆ gpuKernelBallot
| #define gpuKernelBallot | ( | mask, | |
| input ) |
Definition at line 88 of file arch_device_hip.h.
◆ gpuKernelShfl
| #define gpuKernelShfl | ( | input, | |
| source, | |||
| mask ) |
Definition at line 90 of file arch_device_hip.h.
◆ gpuKernelShflDown
| #define gpuKernelShflDown | ( | val, | |
| offset ) |
Definition at line 91 of file arch_device_hip.h.
◆ gpuMalloc
| #define gpuMalloc hipMalloc |
Definition at line 32 of file arch_device_hip.h.
◆ gpuMallocAsync
| #define gpuMallocAsync hipMallocAsync |
Definition at line 34 of file arch_device_hip.h.
◆ gpuMallocHost
| #define gpuMallocHost hipHostMalloc |
Definition at line 33 of file arch_device_hip.h.
◆ gpuMallocManaged
| #define gpuMallocManaged hipMallocManaged |
Definition at line 35 of file arch_device_hip.h.
◆ gpuMemAdviseSetAccessedBy
| #define gpuMemAdviseSetAccessedBy hipMemAdviseSetAccessedBy |
Definition at line 43 of file arch_device_hip.h.
◆ gpuMemAdviseSetPreferredLocation
| #define gpuMemAdviseSetPreferredLocation hipMemAdviseSetPreferredLocation |
Definition at line 44 of file arch_device_hip.h.
◆ gpuMemAttachGlobal
| #define gpuMemAttachGlobal hipMemAttachGlobal |
Definition at line 46 of file arch_device_hip.h.
◆ gpuMemAttachSingle
| #define gpuMemAttachSingle hipMemAttachSingle |
Definition at line 45 of file arch_device_hip.h.
◆ gpuMemcpy
| #define gpuMemcpy hipMemcpy |
Definition at line 38 of file arch_device_hip.h.
◆ gpuMemcpyAsync
| #define gpuMemcpyAsync hipMemcpyAsync |
Definition at line 39 of file arch_device_hip.h.
◆ gpuMemcpyDeviceToDevice
| #define gpuMemcpyDeviceToDevice hipMemcpyDeviceToDevice |
Definition at line 84 of file arch_device_hip.h.
◆ gpuMemcpyDeviceToHost
| #define gpuMemcpyDeviceToHost hipMemcpyDeviceToHost |
Definition at line 82 of file arch_device_hip.h.
◆ gpuMemcpyHostToDevice
| #define gpuMemcpyHostToDevice hipMemcpyHostToDevice |
Definition at line 83 of file arch_device_hip.h.
◆ gpuMemcpyHostToHost
| #define gpuMemcpyHostToHost hipMemcpyHostToHost |
Definition at line 85 of file arch_device_hip.h.
◆ gpuMemcpyKind
| #define gpuMemcpyKind hipMemcpyKind |
Definition at line 81 of file arch_device_hip.h.
◆ gpuMemcpyToSymbol
| #define gpuMemcpyToSymbol hipMemcpyToSymbol |
Definition at line 86 of file arch_device_hip.h.
◆ gpuMemGetInfo
| #define gpuMemGetInfo hipMemGetInfo |
Definition at line 25 of file arch_device_hip.h.
◆ gpuMemPrefetchAsync
| #define gpuMemPrefetchAsync hipMemPrefetchAsync |
Definition at line 47 of file arch_device_hip.h.
◆ gpuMemset
| #define gpuMemset hipMemset |
Definition at line 40 of file arch_device_hip.h.
◆ gpuMemsetAsync
| #define gpuMemsetAsync hipMemsetAsync |
Definition at line 41 of file arch_device_hip.h.
◆ gpuPeekAtLastError
| #define gpuPeekAtLastError hipPeekAtLastError |
Definition at line 15 of file arch_device_hip.h.
◆ gpuSetDevice
| #define gpuSetDevice hipSetDevice |
Definition at line 17 of file arch_device_hip.h.
◆ gpuStream_t
| #define gpuStream_t hipStream_t |
Definition at line 73 of file arch_device_hip.h.
◆ gpuStreamAttachMemAsync
| #define gpuStreamAttachMemAsync hipStreamAttachMemAsync |
Definition at line 56 of file arch_device_hip.h.
◆ gpuStreamCreate
| #define gpuStreamCreate hipStreamCreate |
Definition at line 49 of file arch_device_hip.h.
◆ gpuStreamCreateWithPriority
| #define gpuStreamCreateWithPriority hipStreamCreateWithPriority |
Definition at line 58 of file arch_device_hip.h.
◆ gpuStreamDefault
| #define gpuStreamDefault hipStreamDefault |
Definition at line 59 of file arch_device_hip.h.
◆ gpuStreamDestroy
| #define gpuStreamDestroy hipStreamDestroy |
Definition at line 53 of file arch_device_hip.h.
◆ gpuStreamSynchronize
| #define gpuStreamSynchronize hipStreamSynchronize |
Definition at line 55 of file arch_device_hip.h.
◆ gpuStreamWaitEvent
| #define gpuStreamWaitEvent hipStreamWaitEvent |
Definition at line 54 of file arch_device_hip.h.
◆ gpuSuccess
| #define gpuSuccess hipSuccess |
Definition at line 71 of file arch_device_hip.h.
◆ GPUTHREADS
| #define GPUTHREADS (64) |
Definition at line 109 of file arch_device_hip.h.
◆ gpuWarpSync
| #define gpuWarpSync | ( | ) |
Definition at line 92 of file arch_device_hip.h.
◆ HIPCUB_STDERR
| #define HIPCUB_STDERR |
Definition at line 101 of file arch_device_hip.h.
◆ WARPSPERBLOCK
| #define WARPSPERBLOCK (16) |
Definition at line 112 of file arch_device_hip.h.
Function Documentation
◆ hip_error()
|
inlinestatic |
Definition at line 120 of file arch_device_hip.h.
Variable Documentation
◆ gpuStreamList
|
extern |
Definition at line 51 of file gpu_base.cpp.
