Actual source code: cupmblasinterface.hpp

  1: #pragma once

  3: #include <petsc/private/cupminterface.hpp>

  5: #include <limits> // std::numeric_limits

  7: namespace Petsc
  8: {

 10: namespace device
 11: {

 13: namespace cupm
 14: {

 16: namespace impl
 17: {

 19: #define PetscCallCUPMBLAS_(__abort_fn__, __comm__, ...) \
 20:   do { \
 21:     PetscStackUpdateLine; \
 22:     const cupmBlasError_t cberr_p_ = __VA_ARGS__; \
 23:     if (PetscUnlikely(cberr_p_ != CUPMBLAS_STATUS_SUCCESS)) { \
 24:       if ((cberr_p_ == CUPMBLAS_STATUS_NOT_INITIALIZED || cberr_p_ == CUPMBLAS_STATUS_ALLOC_FAILED) && PetscDeviceInitialized(PETSC_DEVICE_CUPM())) { \
 25:         __abort_fn__(__comm__, PETSC_ERR_GPU_RESOURCE, \
 26:                      "%s error %d (%s). Reports not initialized or alloc failed; " \
 27:                      "this indicates the GPU may have run out resources", \
 28:                      cupmBlasName(), static_cast<PetscErrorCode>(cberr_p_), cupmBlasGetErrorName(cberr_p_)); \
 29:       } \
 30:       __abort_fn__(__comm__, PETSC_ERR_GPU, "%s error %d (%s)", cupmBlasName(), static_cast<PetscErrorCode>(cberr_p_), cupmBlasGetErrorName(cberr_p_)); \
 31:     } \
 32:   } while (0)

 34: #define PetscCallCUPMBLAS(...)             PetscCallCUPMBLAS_(SETERRQ, PETSC_COMM_SELF, __VA_ARGS__)
 35: #define PetscCallCUPMBLASAbort(comm_, ...) PetscCallCUPMBLAS_(SETERRABORT, comm_, __VA_ARGS__)

 37: // given cupmBlasaxpy() then
 38: // T = PETSC_CUPBLAS_FP_TYPE
 39: // given cupmBlasnrm2() then
 40: // T = PETSC_CUPMBLAS_FP_INPUT_TYPE
 41: // u = PETSC_CUPMBLAS_FP_RETURN_TYPE
 42: #if PetscDefined(USE_COMPLEX)
 43:   #if PetscDefined(USE_REAL_SINGLE)
 44:     #define PETSC_CUPMBLAS_FP_TYPE_U       C
 45:     #define PETSC_CUPMBLAS_FP_TYPE_L       c
 46:     #define PETSC_CUPMBLAS_FP_INPUT_TYPE_U S
 47:     #define PETSC_CUPMBLAS_FP_INPUT_TYPE_L s
 48:   #elif PetscDefined(USE_REAL_DOUBLE)
 49:     #define PETSC_CUPMBLAS_FP_TYPE_U       Z
 50:     #define PETSC_CUPMBLAS_FP_TYPE_L       z
 51:     #define PETSC_CUPMBLAS_FP_INPUT_TYPE_U D
 52:     #define PETSC_CUPMBLAS_FP_INPUT_TYPE_L d
 53:   #endif
 54:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE_U PETSC_CUPMBLAS_FP_TYPE_U
 55:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE_L PETSC_CUPMBLAS_FP_TYPE_L
 56: #else
 57:   #if PetscDefined(USE_REAL_SINGLE)
 58:     #define PETSC_CUPMBLAS_FP_TYPE_U S
 59:     #define PETSC_CUPMBLAS_FP_TYPE_L s
 60:   #elif PetscDefined(USE_REAL_DOUBLE)
 61:     #define PETSC_CUPMBLAS_FP_TYPE_U D
 62:     #define PETSC_CUPMBLAS_FP_TYPE_L d
 63:   #endif
 64:   #define PETSC_CUPMBLAS_FP_INPUT_TYPE_U PETSC_CUPMBLAS_FP_TYPE_U
 65:   #define PETSC_CUPMBLAS_FP_INPUT_TYPE_L PETSC_CUPMBLAS_FP_TYPE_L
 66:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE_U
 67:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE_L
 68: #endif // USE_COMPLEX

 70: #if !defined(PETSC_CUPMBLAS_FP_TYPE_U) && !PetscDefined(USE_REAL___FLOAT128)
 71:   #error "Unsupported floating-point type for CUDA/HIP BLAS"
 72: #endif

 74: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_MODIFIED() - Helper macro to build a "modified"
 75: // blas function whose return type does not match the input type
 76: //
 77: // input param:
 78: // func - base suffix of the blas function, e.g. nrm2
 79: //
 80: // notes:
 81: // requires PETSC_CUPMBLAS_FP_INPUT_TYPE to be defined as the blas floating point input type
 82: // letter ("S" for real/complex single, "D" for real/complex double).
 83: //
 84: // requires PETSC_CUPMBLAS_FP_RETURN_TYPE to be defined as the blas floating point output type
 85: // letter ("c" for complex single, "z" for complex double and  for real
 86: // single/double).
 87: //
 88: // In their infinite wisdom nvidia/amd have made the upper-case vs lower-case scheme
 89: // infuriatingly inconsistent...
 90: //
 91: // example usage:
 92: // #define PETSC_CUPMBLAS_FP_INPUT_TYPE  S
 93: // #define PETSC_CUPMBLAS_FP_RETURN_TYPE
 94: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_MODIFIED(nrm2) -> Snrm2
 95: //
 96: // #define PETSC_CUPMBLAS_FP_INPUT_TYPE  D
 97: // #define PETSC_CUPMBLAS_FP_RETURN_TYPE z
 98: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_MODIFIED(nrm2) -> Dznrm2
 99: #define PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_MODIFIED(func) PetscConcat(PetscConcat(PETSC_CUPMBLAS_FP_INPUT_TYPE, PETSC_CUPMBLAS_FP_RETURN_TYPE), func)

101: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_IFPTYPE() - Helper macro to build Iamax and Iamin
102: // because they are both extra special
103: //
104: // input param:
105: // func - base suffix of the blas function, either amax or amin
106: //
107: // notes:
108: // The macro name literally stands for "I" ## "floating point type" because shockingly enough,
109: // that's what it does.
110: //
111: // requires PETSC_CUPMBLAS_FP_TYPE_L to be defined as the lower-case blas floating point input type
112: // letter ("s" for complex single, "z" for complex double, "s" for real single, and "d" for
113: // real double).
114: //
115: // example usage:
116: // #define PETSC_CUPMBLAS_FP_TYPE_L s
117: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_IFPTYPE(amax) -> Isamax
118: //
119: // #define PETSC_CUPMBLAS_FP_TYPE_L z
120: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_IFPTYPE(amin) -> Izamin
121: #define PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_IFPTYPE(func) PetscConcat(I, PetscConcat(PETSC_CUPMBLAS_FP_TYPE_L, func))

123: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_STANDARD() - Helper macro to build a "standard"
124: // blas function name
125: //
126: // input param:
127: // func - base suffix of the blas function, e.g. axpy, scal
128: //
129: // notes:
130: // requires PETSC_CUPMBLAS_FP_TYPE to be defined as the blas floating-point letter ("C" for
131: // complex single, "Z" for complex double, "S" for real single, "D" for real double).
132: //
133: // example usage:
134: // #define PETSC_CUPMBLAS_FP_TYPE S
135: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_STANDARD(axpy) -> Saxpy
136: //
137: // #define PETSC_CUPMBLAS_FP_TYPE Z
138: // PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_STANDARD(axpy) -> Zaxpy
139: #define PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_STANDARD(func) PetscConcat(PETSC_CUPMBLAS_FP_TYPE, func)

141: // PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT() - In case CUDA/HIP don't agree with our suffix
142: // one can provide both here
143: //
144: // input params:
145: // MACRO_SUFFIX - suffix to one of the above blas function builder macros, e.g. STANDARD or
146: // IFPTYPE
147: // our_suffix   - the suffix of the alias function
148: // their_suffix - the suffix of the function being aliased
149: //
150: // notes:
151: // requires PETSC_CUPMBLAS_PREFIX to be defined as the specific CUDA/HIP blas function
152: // prefix. requires any other specific definitions required by the specific builder macro to
153: // also be defined. See PETSC_CUPM_ALIAS_FUNCTION_EXACT() for the exact expansion of the
154: // function alias.
155: //
156: // example usage:
157: // #define PETSC_CUPMBLAS_PREFIX  cublas
158: // #define PETSC_CUPMBLAS_FP_TYPE C
159: // PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD,dot,dotc) ->
160: // template 
161: // static constexpr auto cupmBlasXdot(T&&... args) *noexcept and returntype detection*
162: // {
163: //   return cublasCdotc(std::forward(args)...);
164: // }
165: #define PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(MACRO_SUFFIX, our_suffix, their_suffix) \
166:   PETSC_CUPM_ALIAS_FUNCTION(PetscConcat(cupmBlasX, our_suffix), PetscConcat(PETSC_CUPMBLAS_PREFIX, PetscConcat(PETSC_CUPMBLAS_BUILD_BLAS_FUNCTION_ALIAS_, MACRO_SUFFIX)(their_suffix)))

168: // PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION() - Alias a CUDA/HIP blas function
169: //
170: // input params:
171: // MACRO_SUFFIX - suffix to one of the above blas function builder macros, e.g. STANDARD or
172: // IFPTYPE
173: // suffix       - the common suffix between CUDA and HIP of the alias function
174: //
175: // notes:
176: // see PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(), this macro just calls that one with "suffix" as
177: // "our_prefix" and "their_prefix"
178: #define PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(MACRO_SUFFIX, suffix) PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(MACRO_SUFFIX, suffix, suffix)

180: // PETSC_CUPMBLAS_ALIAS_FUNCTION() - Alias a CUDA/HIP library function
181: //
182: // input params:
183: // suffix - the common suffix between CUDA and HIP of the alias function
184: //
185: // notes:
186: // requires PETSC_CUPMBLAS_PREFIX to be defined as the specific CUDA/HIP blas library
187: // prefix. see PETSC_CUPMM_ALIAS_FUNCTION_EXACT() for the precise expansion of this macro.
188: //
189: // example usage:
190: // #define PETSC_CUPMBLAS_PREFIX hipblas
191: // PETSC_CUPMBLAS_ALIAS_FUNCTION(Create) ->
192: // template 
193: // static constexpr auto cupmBlasCreate(T&&... args) *noexcept and returntype detection*
194: // {
195: //   return hipblasCreate(std::forward(args)...);
196: // }
197: #define PETSC_CUPMBLAS_ALIAS_FUNCTION(suffix) PETSC_CUPM_ALIAS_FUNCTION(PetscConcat(cupmBlas, suffix), PetscConcat(PETSC_CUPMBLAS_PREFIX, suffix))

199: template <DeviceType>
200: struct BlasInterfaceImpl;

202: // Exists because HIP (for whatever godforsaken reason) has elected to define both their
203: // hipBlasHandle_t and hipSolverHandle_t as void *. So we cannot disambiguate them for overload
204: // resolution and hence need to wrap their types int this mess.
205: template <typename T, std::size_t I>
206: class cupmBlasHandleWrapper {
207: public:
208:   constexpr cupmBlasHandleWrapper() noexcept = default;
209:   constexpr cupmBlasHandleWrapper(T h) noexcept : handle_{std::move(h)} { static_assert(std::is_standard_layout<cupmBlasHandleWrapper<T, I>>::value, ""); }

211:   cupmBlasHandleWrapper &operator=(std::nullptr_t) noexcept
212:   {
213:     handle_ = nullptr;
214:     return *this;
215:   }

217:   operator T() const { return handle_; }

219:   const T *ptr_to() const { return &handle_; }
220:   T       *ptr_to() { return &handle_; }

222: private:
223:   T handle_{};
224: };

226: #if PetscDefined(HAVE_CUDA)
227:   #define PETSC_CUPMBLAS_PREFIX         cublas
228:   #define PETSC_CUPMBLAS_PREFIX_U       CUBLAS
229:   #define PETSC_CUPMBLAS_FP_TYPE        PETSC_CUPMBLAS_FP_TYPE_U
230:   #define PETSC_CUPMBLAS_FP_INPUT_TYPE  PETSC_CUPMBLAS_FP_INPUT_TYPE_U
231:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE PETSC_CUPMBLAS_FP_RETURN_TYPE_L
232: template <>
233: struct PETSC_SINGLE_LIBRARY_VISIBILITY_INTERNAL BlasInterfaceImpl<DeviceType::CUDA> : Interface<DeviceType::CUDA> {
234:   // typedefs
235:   using cupmBlasHandle_t      = cupmBlasHandleWrapper<cublasHandle_t, 0>;
236:   using cupmBlasError_t       = cublasStatus_t;
237:   using cupmBlasInt_t         = int;
238:   using cupmBlasPointerMode_t = cublasPointerMode_t;

240:   // values
241:   static const auto CUPMBLAS_STATUS_SUCCESS         = CUBLAS_STATUS_SUCCESS;
242:   static const auto CUPMBLAS_STATUS_NOT_INITIALIZED = CUBLAS_STATUS_NOT_INITIALIZED;
243:   static const auto CUPMBLAS_STATUS_ALLOC_FAILED    = CUBLAS_STATUS_ALLOC_FAILED;
244:   static const auto CUPMBLAS_POINTER_MODE_HOST      = CUBLAS_POINTER_MODE_HOST;
245:   static const auto CUPMBLAS_POINTER_MODE_DEVICE    = CUBLAS_POINTER_MODE_DEVICE;
246:   static const auto CUPMBLAS_OP_T                   = CUBLAS_OP_T;
247:   static const auto CUPMBLAS_OP_N                   = CUBLAS_OP_N;
248:   static const auto CUPMBLAS_OP_C                   = CUBLAS_OP_C;
249:   static const auto CUPMBLAS_FILL_MODE_LOWER        = CUBLAS_FILL_MODE_LOWER;
250:   static const auto CUPMBLAS_FILL_MODE_UPPER        = CUBLAS_FILL_MODE_UPPER;
251:   static const auto CUPMBLAS_SIDE_LEFT              = CUBLAS_SIDE_LEFT;
252:   static const auto CUPMBLAS_SIDE_RIGHT             = CUBLAS_SIDE_RIGHT;
253:   static const auto CUPMBLAS_DIAG_NON_UNIT          = CUBLAS_DIAG_NON_UNIT;

255:   // utility functions
256:   PETSC_CUPMBLAS_ALIAS_FUNCTION(Create)
257:   PETSC_CUPMBLAS_ALIAS_FUNCTION(Destroy)
258:   PETSC_CUPMBLAS_ALIAS_FUNCTION(GetStream)
259:   PETSC_CUPMBLAS_ALIAS_FUNCTION(SetStream)
260:   PETSC_CUPMBLAS_ALIAS_FUNCTION(GetPointerMode)
261:   PETSC_CUPMBLAS_ALIAS_FUNCTION(SetPointerMode)

263:   // level 1 BLAS
264:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, axpy)
265:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, copy)
266:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, scal)
267:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, dot, PetscIfPetscDefined(USE_COMPLEX, dotc, dot))
268:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, dotu, PetscIfPetscDefined(USE_COMPLEX, dotu, dot))
269:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, swap)
270:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(MODIFIED, nrm2)
271:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(IFPTYPE, amax)
272:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(MODIFIED, asum)

274:   // level 2 BLAS
275:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gemv)
276:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trmv)
277:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trsv)
278:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gbmv)
279:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, tbmv)
280:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, tbsv)
281:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, hemv, PetscIfPetscDefined(USE_COMPLEX, hemv, symv))

283:   // level 3 BLAS
284:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gemm)
285:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trsm)

287:   // BLAS extensions
288:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, geam)
289:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, dgmm)

291:   PETSC_NODISCARD static const char *cupmBlasGetErrorName(cupmBlasError_t status) noexcept { return PetscCUBLASGetErrorName(status); }
292: };
293:   #undef PETSC_CUPMBLAS_PREFIX
294:   #undef PETSC_CUPMBLAS_PREFIX_U
295:   #undef PETSC_CUPMBLAS_FP_TYPE
296:   #undef PETSC_CUPMBLAS_FP_INPUT_TYPE
297:   #undef PETSC_CUPMBLAS_FP_RETURN_TYPE
298: #endif // PetscDefined(HAVE_CUDA)

300: #if PetscDefined(HAVE_HIP)
301:   #define PETSC_CUPMBLAS_PREFIX         hipblas
302:   #define PETSC_CUPMBLAS_PREFIX_U       HIPBLAS
303:   #define PETSC_CUPMBLAS_FP_TYPE        PETSC_CUPMBLAS_FP_TYPE_U
304:   #define PETSC_CUPMBLAS_FP_INPUT_TYPE  PETSC_CUPMBLAS_FP_INPUT_TYPE_U
305:   #define PETSC_CUPMBLAS_FP_RETURN_TYPE PETSC_CUPMBLAS_FP_RETURN_TYPE_L
306: template <>
307: struct PETSC_SINGLE_LIBRARY_VISIBILITY_INTERNAL BlasInterfaceImpl<DeviceType::HIP> : Interface<DeviceType::HIP> {
308:   // typedefs
309:   using cupmBlasHandle_t      = cupmBlasHandleWrapper<hipblasHandle_t, 0>;
310:   using cupmBlasError_t       = hipblasStatus_t;
311:   using cupmBlasInt_t         = int; // rocblas will have its own
312:   using cupmBlasPointerMode_t = hipblasPointerMode_t;

314:   // values
315:   static const auto CUPMBLAS_STATUS_SUCCESS         = HIPBLAS_STATUS_SUCCESS;
316:   static const auto CUPMBLAS_STATUS_NOT_INITIALIZED = HIPBLAS_STATUS_NOT_INITIALIZED;
317:   static const auto CUPMBLAS_STATUS_ALLOC_FAILED    = HIPBLAS_STATUS_ALLOC_FAILED;
318:   static const auto CUPMBLAS_POINTER_MODE_HOST      = HIPBLAS_POINTER_MODE_HOST;
319:   static const auto CUPMBLAS_POINTER_MODE_DEVICE    = HIPBLAS_POINTER_MODE_DEVICE;
320:   static const auto CUPMBLAS_OP_T                   = HIPBLAS_OP_T;
321:   static const auto CUPMBLAS_OP_N                   = HIPBLAS_OP_N;
322:   static const auto CUPMBLAS_OP_C                   = HIPBLAS_OP_C;
323:   static const auto CUPMBLAS_FILL_MODE_LOWER        = HIPBLAS_FILL_MODE_LOWER;
324:   static const auto CUPMBLAS_FILL_MODE_UPPER        = HIPBLAS_FILL_MODE_UPPER;
325:   static const auto CUPMBLAS_SIDE_LEFT              = HIPBLAS_SIDE_LEFT;
326:   static const auto CUPMBLAS_SIDE_RIGHT             = HIPBLAS_SIDE_RIGHT;
327:   static const auto CUPMBLAS_DIAG_NON_UNIT          = HIPBLAS_DIAG_NON_UNIT;

329:   // utility functions
330:   PETSC_CUPMBLAS_ALIAS_FUNCTION(Create)
331:   PETSC_CUPMBLAS_ALIAS_FUNCTION(Destroy)
332:   PETSC_CUPMBLAS_ALIAS_FUNCTION(GetStream)
333:   PETSC_CUPMBLAS_ALIAS_FUNCTION(SetStream)
334:   PETSC_CUPMBLAS_ALIAS_FUNCTION(GetPointerMode)
335:   PETSC_CUPMBLAS_ALIAS_FUNCTION(SetPointerMode)

337:   // level 1 BLAS
338:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, axpy)
339:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, copy)
340:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, scal)
341:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, dot, PetscIfPetscDefined(USE_COMPLEX, dotc, dot))
342:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, dotu, PetscIfPetscDefined(USE_COMPLEX, dotu, dot))
343:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, swap)
344:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(MODIFIED, nrm2)
345:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(IFPTYPE, amax)
346:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(MODIFIED, asum)

348:   // level 2 BLAS
349:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gemv)
350:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trmv)
351:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trsv)
352:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gbmv)
353:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, tbmv)
354:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, tbsv)
355:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION_EXACT(STANDARD, hemv, PetscIfPetscDefined(USE_COMPLEX, hemv, symv))

357:   // level 3 BLAS
358:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, gemm)
359:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, trsm)

361:   // BLAS extensions
362:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, geam)
363:   PETSC_CUPMBLAS_ALIAS_BLAS_FUNCTION(STANDARD, dgmm)

365:   PETSC_NODISCARD static const char *cupmBlasGetErrorName(cupmBlasError_t status) noexcept { return PetscHIPBLASGetErrorName(status); }
366: };
367:   #undef PETSC_CUPMBLAS_PREFIX
368:   #undef PETSC_CUPMBLAS_PREFIX_U
369:   #undef PETSC_CUPMBLAS_FP_TYPE
370:   #undef PETSC_CUPMBLAS_FP_INPUT_TYPE
371:   #undef PETSC_CUPMBLAS_FP_RETURN_TYPE
372: #endif // PetscDefined(HAVE_HIP)

374: #define PETSC_CUPMBLAS_IMPL_CLASS_HEADER(T) \
375:   PETSC_CUPM_INHERIT_INTERFACE_TYPEDEFS_USING(T); \
376:   /* introspection */ \
377:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasGetErrorName; \
378:   /* types */ \
379:   using cupmBlasHandle_t      = typename ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasHandle_t; \
380:   using cupmBlasError_t       = typename ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasError_t; \
381:   using cupmBlasInt_t         = typename ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasInt_t; \
382:   using cupmBlasPointerMode_t = typename ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasPointerMode_t; \
383:   /* values */ \
384:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_STATUS_SUCCESS; \
385:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_STATUS_NOT_INITIALIZED; \
386:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_STATUS_ALLOC_FAILED; \
387:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_POINTER_MODE_HOST; \
388:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_POINTER_MODE_DEVICE; \
389:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_OP_T; \
390:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_OP_N; \
391:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_OP_C; \
392:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_FILL_MODE_LOWER; \
393:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_FILL_MODE_UPPER; \
394:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_SIDE_LEFT; \
395:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_SIDE_RIGHT; \
396:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::CUPMBLAS_DIAG_NON_UNIT; \
397:   /* utility functions */ \
398:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasCreate; \
399:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasDestroy; \
400:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasGetStream; \
401:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasSetStream; \
402:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasGetPointerMode; \
403:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasSetPointerMode; \
404:   /* level 1 BLAS */ \
405:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXaxpy; \
406:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXcopy; \
407:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXscal; \
408:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXdot; \
409:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXdotu; \
410:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXswap; \
411:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXnrm2; \
412:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXamax; \
413:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXasum; \
414:   /* level 2 BLAS */ \
415:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXgemv; \
416:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXtrmv; \
417:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXtrsv; \
418:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXgbmv; \
419:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXtbmv; \
420:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXtbsv; \
421:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXhemv; \
422:   /* level 3 BLAS */ \
423:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXgemm; \
424:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXtrsm; \
425:   /* BLAS extensions */ \
426:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXgeam; \
427:   using ::Petsc::device::cupm::impl::BlasInterfaceImpl<T>::cupmBlasXdgmm

429: // The actual interface class
430: template <DeviceType T>
431: struct BlasInterface : BlasInterfaceImpl<T> {
432:   PETSC_CUPMBLAS_IMPL_CLASS_HEADER(T);

434:   PETSC_NODISCARD static constexpr const char *cupmBlasName() noexcept { return T == DeviceType::CUDA ? "cuBLAS" : "hipBLAS"; }

436:   static PetscErrorCode PetscCUPMBlasSetPointerModeFromPointer(cupmBlasHandle_t handle, const void *ptr) noexcept
437:   {
438:     auto mtype = PETSC_MEMTYPE_HOST;

440:     PetscFunctionBegin;
441:     PetscCall(PetscCUPMGetMemType(ptr, &mtype));
442:     PetscCallCUPMBLAS(cupmBlasSetPointerMode(handle, PetscMemTypeDevice(mtype) ? CUPMBLAS_POINTER_MODE_DEVICE : CUPMBLAS_POINTER_MODE_HOST));
443:     PetscFunctionReturn(PETSC_SUCCESS);
444:   }

446:   static PetscErrorCode checkCupmBlasIntCast(PetscInt x) noexcept
447:   {
448:     PetscFunctionBegin;
449:     PetscCheck((std::is_same<PetscInt, cupmBlasInt_t>::value) || (x <= std::numeric_limits<cupmBlasInt_t>::max()), PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "%" PetscInt_FMT " is too big for %s, which may be restricted to 32-bit integers", x, cupmBlasName());
450:     PetscCheck(x >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Passing negative integer (%" PetscInt_FMT ") to %s routine", x, cupmBlasName());
451:     PetscFunctionReturn(PETSC_SUCCESS);
452:   }

454:   static PetscErrorCode PetscCUPMBlasIntCast(PetscInt x, cupmBlasInt_t *y) noexcept
455:   {
456:     PetscFunctionBegin;
457:     *y = static_cast<cupmBlasInt_t>(x);
458:     PetscCall(checkCupmBlasIntCast(x));
459:     PetscFunctionReturn(PETSC_SUCCESS);
460:   }

462:   class CUPMBlasPointerModeGuard {
463:   public:
464:     CUPMBlasPointerModeGuard(const cupmBlasHandle_t &handle, cupmBlasPointerMode_t mode) noexcept : handle_{handle}
465:     {
466:       PetscFunctionBegin;
467:       PetscCallCUPMBLASAbort(PETSC_COMM_SELF, cupmBlasGetPointerMode(handle, &this->saved_));
468:       PetscCallCUPMBLASAbort(PETSC_COMM_SELF, cupmBlasSetPointerMode(handle, mode));
469:       PetscFunctionReturnVoid();
470:     }

472:     CUPMBlasPointerModeGuard(const cupmBlasHandle_t &handle, PetscMemType mtype) noexcept : CUPMBlasPointerModeGuard{handle, PetscMemTypeDevice(mtype) ? CUPMBLAS_POINTER_MODE_DEVICE : CUPMBLAS_POINTER_MODE_HOST} { }

474:     ~CUPMBlasPointerModeGuard() noexcept
475:     {
476:       PetscFunctionBegin;
477:       PetscCallCUPMBLASAbort(PETSC_COMM_SELF, cupmBlasSetPointerMode(this->handle_, this->saved_));
478:       PetscFunctionReturnVoid();
479:     }

481:   private:
482:     cupmBlasHandle_t      handle_;
483:     cupmBlasPointerMode_t saved_;
484:   };
485: };

487: #define PETSC_CUPMBLAS_INHERIT_INTERFACE_TYPEDEFS_USING(T) \
488:   PETSC_CUPMBLAS_IMPL_CLASS_HEADER(T); \
489:   using ::Petsc::device::cupm::impl::BlasInterface<T>::cupmBlasName; \
490:   using ::Petsc::device::cupm::impl::BlasInterface<T>::PetscCUPMBlasSetPointerModeFromPointer; \
491:   using ::Petsc::device::cupm::impl::BlasInterface<T>::checkCupmBlasIntCast; \
492:   using ::Petsc::device::cupm::impl::BlasInterface<T>::PetscCUPMBlasIntCast; \
493:   using CUPMBlasPointerModeGuard = typename ::Petsc::device::cupm::impl::BlasInterface<T>::CUPMBlasPointerModeGuard

495: #if PetscDefined(HAVE_CUDA)
496: extern template struct PETSC_SINGLE_LIBRARY_VISIBILITY_INTERNAL BlasInterface<DeviceType::CUDA>;
497: #endif

499: #if PetscDefined(HAVE_HIP)
500: extern template struct PETSC_SINGLE_LIBRARY_VISIBILITY_INTERNAL BlasInterface<DeviceType::HIP>;
501: #endif

503: } // namespace impl

505: } // namespace cupm

507: } // namespace device

509: } // namespace Petsc