Differential D62393 Diff 223845 libomptarget/deviceRTLs/nvptx/src/target_impl.h

Changeset View

Standalone View

	Show First 20 Lines • Show All 92 Lines • ▼ Show 20 Lines
	INLINE void __kmpc_impl_syncwarp(__kmpc_impl_lanemask_t Mask) {			INLINE void __kmpc_impl_syncwarp(__kmpc_impl_lanemask_t Mask) {
	#if CUDA_VERSION >= 9000			#if CUDA_VERSION >= 9000
	__syncwarp(Mask);			__syncwarp(Mask);
	#else			#else
	// In Cuda < 9.0 no need to sync threads in warps.			// In Cuda < 9.0 no need to sync threads in warps.
	#endif // CUDA_VERSION			#endif // CUDA_VERSION
	}			}

				#ifdef __LP64__
				#define PTR_CONSTRAINT "l"
				#else // __LP64__
				#define PTR_CONSTRAINT "r"
				#endif // __LP64__

				INLINE int32_t __kmpc_impl_get_parallel_level(uint8_t &ParLevel) {
				// Use volatile access in case of CUDA8 to prevent dangerous optimizations
				// performed by ptxas.
				int32_t ParLevelVal;
				#if defined(CUDA_VERSION) && CUDA_VERSION <= 8000
				asm volatile("ld.volatile.u8 %0, [%1];"
				: "=r"(ParLevelVal)
				: PTR_CONSTRAINT(&ParLevel)
				: "memory");
				#else
				ParLevelVal = ParLevel;
				#endif // CUDA_VERSION
				return ParLevelVal;
				}

				INLINE void __kmpc_impl_set_parallel_level(uint8_t &ParLevel,
				int32_t ParLevelVal) {
				#if defined(CUDA_VERSION) && CUDA_VERSION <= 8000
				asm volatile("st.volatile.u8 [%1], %0;" ::"r"(ParLevelVal),
				PTR_CONSTRAINT(&ParLevel)
				: "memory");
				#else // CUDA_VERSION
				ParLevel = ParLevelVal;
				#endif // CUDA_VERSION
				}

				#undef PTR_CONSTRAINT

	#endif			#endif