| 1 |
! Minimal check whether HIP works - by checking whether the API routines |
| 2 |
! seem to work. This includes a fallback if hipfort is not available |
| 3 |
|
| 4 |
#ifndef HAVE_HIPFORT |
| 5 |
#ifndef USE_HIP_FALLBACK_MODULE |
| 6 |
#if USE_CUDA_NAMES |
| 7 |
#warning "Using fallback implementation for module hipfort as HAVE_HIPFORT is undefined (for NVIDA/CUDA)" |
| 8 |
#else |
| 9 |
#warning "Using fallback implementation for module hipfort as HAVE_HIPFORT is undefined - assume AMD as USE_CUDA_NAMES is unset" |
| 10 |
#endif |
| 11 |
#endif |
| 12 |
module hipfort ! Minimal implementation for the testsuite |
| 13 |
implicit none |
| 14 |
|
| 15 |
enum, bind(c) |
| 16 |
enumerator :: hipSuccess = 0 |
| 17 |
enumerator :: hipErrorNotSupported = 801 |
| 18 |
end enum |
| 19 |
|
| 20 |
enum, bind(c) |
| 21 |
enumerator :: hipDeviceAttributeClockRate = 5 |
| 22 |
enumerator :: hipDeviceAttributeMaxGridDimX = 29 |
| 23 |
end enum |
| 24 |
|
| 25 |
interface |
| 26 |
integer(kind(hipSuccess)) function hipDeviceGetAttribute (ip, attr, dev) & |
| 27 |
#if USE_CUDA_NAMES |
| 28 |
bind(c, name="cudaDeviceGetAttribute") |
| 29 |
#else |
| 30 |
bind(c, name="hipDeviceGetAttribute") |
| 31 |
#endif |
| 32 |
use iso_c_binding, only: c_ptr, c_int |
| 33 |
import |
| 34 |
implicit none |
| 35 |
type(c_ptr), value :: ip |
| 36 |
integer(kind(hipDeviceAttributeClockRate)), value :: attr |
| 37 |
integer(c_int), value :: dev |
| 38 |
end |
| 39 |
|
| 40 |
integer(kind(hipSuccess)) function hipCtxGetApiVersion (ctx, ip) & |
| 41 |
#if USE_CUDA_NAMES |
| 42 |
bind(c, name="cudaCtxGetApiVersion") |
| 43 |
#else |
| 44 |
bind(c, name="hipCtxGetApiVersion") |
| 45 |
#endif |
| 46 |
use iso_c_binding, only: c_ptr |
| 47 |
import |
| 48 |
implicit none |
| 49 |
type(c_ptr), value :: ctx, ip |
| 50 |
end |
| 51 |
|
| 52 |
integer(kind(hipSuccess)) function hipStreamQuery (stream) & |
| 53 |
#if USE_CUDA_NAMES |
| 54 |
bind(c, name="cudaStreamQuery") |
| 55 |
#else |
| 56 |
bind(c, name="hipStreamQuery") |
| 57 |
#endif |
| 58 |
use iso_c_binding, only: c_ptr |
| 59 |
import |
| 60 |
implicit none |
| 61 |
type(c_ptr), value :: stream |
| 62 |
end |
| 63 |
|
| 64 |
integer(kind(hipSuccess)) function hipStreamGetFlags (stream, flags) & |
| 65 |
#if USE_CUDA_NAMES |
| 66 |
bind(c, name="cudaStreamGetFlags") |
| 67 |
#else |
| 68 |
bind(c, name="hipStreamGetFlags") |
| 69 |
#endif |
| 70 |
use iso_c_binding, only: c_ptr |
| 71 |
import |
| 72 |
implicit none |
| 73 |
type(c_ptr), value :: stream |
| 74 |
type(c_ptr), value :: flags |
| 75 |
end |
| 76 |
end interface |
| 77 |
end module |
| 78 |
#endif |
| 79 |
|
| 80 |
program main |
| 81 |
use iso_c_binding, only: c_ptr, c_int, c_loc |
| 82 |
use omp_lib |
| 83 |
use hipfort |
| 84 |
implicit none (type, external) |
| 85 |
|
| 86 |
! Only supported since CUDA 12.8 - skip for better compatibility |
| 87 |
! ! Manally implement hipStreamGetDevice as hipfort misses it |
| 88 |
! ! -> https://github.com/ROCm/hipfort/issues/238 |
| 89 |
! interface |
| 90 |
! integer(kind(hipSuccess)) function my_hipStreamGetDevice(stream, dev) & |
| 91 |
!#if USE_CUDA_NAMES |
| 92 |
! bind(c, name="cudaStreamGetDevice") |
| 93 |
!#else |
| 94 |
! bind(c, name="hipStreamGetDevice") |
| 95 |
!#endif |
| 96 |
! use iso_c_binding, only: c_ptr, c_int |
| 97 |
! import |
| 98 |
! implicit none |
| 99 |
! type(c_ptr), value :: stream |
| 100 |
! integer(c_int) :: dev |
| 101 |
! end |
| 102 |
! end interface |
| 103 |
|
| 104 |
integer(c_int), target :: ivar |
| 105 |
integer(omp_interop_rc_kind) :: res |
| 106 |
integer(omp_interop_kind) :: obj |
| 107 |
integer(omp_interop_fr_kind) :: fr |
| 108 |
integer(kind(hipSuccess)) :: hip_err |
| 109 |
integer(c_int) :: hip_dev, dev_stream |
| 110 |
type(c_ptr) :: hip_ctx, hip_sm |
| 111 |
|
| 112 |
logical :: vendor_is_amd |
| 113 |
|
| 114 |
obj = omp_interop_none |
| 115 |
|
| 116 |
!$omp interop init(target, targetsync, prefer_type("hip") : obj) |
| 117 |
|
| 118 |
fr = omp_get_interop_int (obj, omp_ipr_fr_id, res) |
| 119 |
if (res /= omp_irc_success) error stop 1 |
| 120 |
if (fr /= omp_ifr_hip) error stop 1 |
| 121 |
|
| 122 |
ivar = omp_get_interop_int (obj, omp_ipr_vendor, res) |
| 123 |
if (ivar == 1) then ! AMD |
| 124 |
vendor_is_amd = .true. |
| 125 |
else if (ivar == 11) then ! Nvidia |
| 126 |
vendor_is_amd = .false. |
| 127 |
else |
| 128 |
error stop 1 ! Unknown |
| 129 |
endif |
| 130 |
#if USE_CUDA_NAMES |
| 131 |
if (vendor_is_amd) error stop 1 |
| 132 |
#else |
| 133 |
if (.not. vendor_is_amd) error stop 1 |
| 134 |
#endif |
| 135 |
|
| 136 |
! Check whether the omp_ipr_device -> hipDevice_t yields a valid device. |
| 137 |
|
| 138 |
hip_dev = omp_get_interop_int (obj, omp_ipr_device, res) |
| 139 |
if (res /= omp_irc_success) error stop 1 |
| 140 |
|
| 141 |
! AMD messed up in Fortran with the attribute handling, missing the |
| 142 |
! translation table it has for C. |
| 143 |
block |
| 144 |
enum, bind(c) |
| 145 |
enumerator :: cudaDevAttrClockRate = 13 |
| 146 |
enumerator :: cudaDevAttrMaxGridDimX = 5 |
| 147 |
end enum |
| 148 |
|
| 149 |
! Assume a clock size is available and > 1 GHz; value is in kHz. |
| 150 |
! c_loc is completely bogus, but as AMD messed up the interface ... |
| 151 |
! Cf. https://github.com/ROCm/hipfort/issues/239 |
| 152 |
if (vendor_is_amd) then |
| 153 |
hip_err = hipDeviceGetAttribute (c_loc(ivar), hipDeviceAttributeClockRate, hip_dev) |
| 154 |
else |
| 155 |
hip_err = hipDeviceGetAttribute (c_loc(ivar), cudaDevAttrClockRate, hip_dev) |
| 156 |
endif |
| 157 |
if (hip_err /= hipSuccess) error stop 1 |
| 158 |
if (ivar <= 1000000) error stop 1 ! in kHz |
| 159 |
|
| 160 |
! Assume that the MaxGridDimX is available and > 1024 |
| 161 |
! c_loc is completely bogus, but as AMD messed up the interface ... |
| 162 |
! Cf. https://github.com/ROCm/hipfort/issues/239 |
| 163 |
if (vendor_is_amd) then |
| 164 |
hip_err = hipDeviceGetAttribute (c_loc(ivar), hipDeviceAttributeMaxGridDimX, hip_dev) |
| 165 |
else |
| 166 |
hip_err = hipDeviceGetAttribute (c_loc(ivar), cudaDevAttrMaxGridDimX, hip_dev) |
| 167 |
endif |
| 168 |
if (hip_err /= hipSuccess) error stop 1 |
| 169 |
if (ivar <= 1024) error stop 1 |
| 170 |
end block |
| 171 |
|
| 172 |
|
| 173 |
! Check whether the omp_ipr_device_context -> hipCtx_t yields a context. |
| 174 |
|
| 175 |
hip_ctx = omp_get_interop_ptr (obj, omp_ipr_device_context, res) |
| 176 |
if (res /= omp_irc_success) error stop 1 |
| 177 |
|
| 178 |
! ! Assume API Version > 0 for Nvidia, hipErrorNotSupported for AMD. */ |
| 179 |
! ivar = -99 |
| 180 |
! ! AMD deprectated hipCtxGetApiVersion (in C/C++) |
| 181 |
! hip_err = hipCtxGetApiVersion (hip_ctx, c_loc(ivar)) |
| 182 |
! |
| 183 |
! if (vendor_is_amd) then |
| 184 |
! if (hip_err /= hipErrorNotSupported .or. ivar /= -99) error stop 1 |
| 185 |
! else |
| 186 |
! if (hip_err /= hipSuccess) error stop 1 |
| 187 |
! if (ivar <= 0) error stop 1 |
| 188 |
! end if |
| 189 |
|
| 190 |
|
| 191 |
! Check whether the omp_ipr_targetsync -> hipStream_t yields a stream. |
| 192 |
|
| 193 |
hip_sm = omp_get_interop_ptr (obj, omp_ipr_targetsync, res) |
| 194 |
if (res /= omp_irc_success) error stop 1 |
| 195 |
|
| 196 |
! Skip as this is only in CUDA 12.8 |
| 197 |
! dev_stream = 99 |
| 198 |
! ! Not (yet) implemented: https://github.com/ROCm/hipfort/issues/238 |
| 199 |
! ! hip_err = hipStreamGetDevice (hip_sm, dev_stream) |
| 200 |
! hip_err = my_hipStreamGetDevice (hip_sm, dev_stream) |
| 201 |
! if (hip_err /= hipSuccess) error stop 1 |
| 202 |
! if (dev_stream /= hip_dev) error stop 1 |
| 203 |
|
| 204 |
! Get flags of the stream |
| 205 |
hip_err = hipStreamGetFlags (hip_sm, c_loc (ivar)) |
| 206 |
if (hip_err /= hipSuccess) error stop 1 |
| 207 |
! Accept any value |
| 208 |
|
| 209 |
! All jobs should have been completed (as there were none none) |
| 210 |
hip_err = hipStreamQuery (hip_sm) |
| 211 |
if (hip_err /= hipSuccess) error stop 1 |
| 212 |
|
| 213 |
!$omp interop destroy(obj) |
| 214 |
end |