Use target_impl functions to replace more inline asm

[JonChesterfield]

Summary

Use target_impl functions to replace more inline asm
Follow on from D65836. Removes remaining asm shuffles and lanemask accessors
Also changes the types of target_impl bitwise functions to unsigned.

[jdoerfert]

LGTM

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/loop.cu

if (schedule == kmp_sched_static_chunk) {
             omptarget_nvptx_threadPrivateContext->Chunk(tid));
     }
   }
 
   ////////////////////////////////////////////////////////////////////////////////
   // Support for dispatch next
 
   INLINE static int64_t Shuffle(unsigned active, int64_t val, int leader) {
-    int lo, hi;
+    uint32_t lo, hi;
     __kmpc_impl_unpack(val, lo, hi);
     hi = __kmpc_impl_shfl_sync(active, hi, leader);
     lo = __kmpc_impl_shfl_sync(active, lo, leader);
     return __kmpc_impl_pack(lo, hi);
   }
 
   INLINE static uint64_t NextIter() {
     __kmpc_impl_lanemask_t active = __ACTIVEMASK();
-    int leader = __kmpc_impl_ffs(active) - 1;
-    int change = __kmpc_impl_popc(active);
+    uint32_t leader = __kmpc_impl_ffs(active) - 1;
+    uint32_t change = __kmpc_impl_popc(active);
     __kmpc_impl_lanemask_t lane_mask_lt = __kmpc_impl_lanemask_lt();
     unsigned int rank = __kmpc_impl_popc(active & lane_mask_lt);
     uint64_t warp_res;
     if (rank == 0) {
       warp_res = atomicAdd(
           (unsigned long long *)&omptarget_nvptx_threadPrivateContext->Cnt(),
           change);
     }

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/parallel.cu

 ////////////////////////////////////////////////////////////////////////////////
 // support for convergent simd (team of threads in a warp only)
 ////////////////////////////////////////////////////////////////////////////////
 EXTERN bool __kmpc_kernel_convergent_simd(void *buffer, uint32_t Mask,
                                           bool *IsFinal, int32_t *LaneSource,
                                           int32_t *LaneId, int32_t *NumLanes) {
   PRINT0(LD_IO, "call to __kmpc_kernel_convergent_simd\n");
   uint32_t ConvergentMask = Mask;
-  int32_t ConvergentSize = __popc(ConvergentMask);
+  int32_t ConvergentSize = __kmpc_impl_popc(ConvergentMask);
   uint32_t WorkRemaining = ConvergentMask >> (*LaneSource + 1);
-  *LaneSource += __ffs(WorkRemaining);
-  *IsFinal = __popc(WorkRemaining) == 1;
-  uint32_t lanemask_lt;
-  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(lanemask_lt));
-  *LaneId = __popc(ConvergentMask & lanemask_lt);
+  *LaneSource += __kmpc_impl_ffs(WorkRemaining);
+  *IsFinal = __kmpc_impl_popc(WorkRemaining) == 1;
+  uint32_t lanemask_lt = __kmpc_impl_lanemask_lt();
+  *LaneId = __kmpc_impl_popc(ConvergentMask & lanemask_lt);
 
   int threadId = GetLogicalThreadIdInBlock(isSPMDMode());
   int sourceThreadId = (threadId & ~(WARPSIZE - 1)) + *LaneSource;
 
   ConvergentSimdJob *job = (ConvergentSimdJob *)buffer;
   int32_t SimdLimit =
       omptarget_nvptx_threadPrivateContext->SimdLimitForNextSimd(threadId);
   job->slimForNextSimd = SimdLimit;

 ////////////////////////////////////////////////////////////////////////////////
 // support for convergent parallelism (team of threads in a warp only)
 ////////////////////////////////////////////////////////////////////////////////
 EXTERN bool __kmpc_kernel_convergent_parallel(void *buffer, uint32_t Mask,
                                               bool *IsFinal,
                                               int32_t *LaneSource) {
   PRINT0(LD_IO, "call to __kmpc_kernel_convergent_parallel\n");
   uint32_t ConvergentMask = Mask;
-  int32_t ConvergentSize = __popc(ConvergentMask);
+  int32_t ConvergentSize = __kmpc_impl_popc(ConvergentMask);
   uint32_t WorkRemaining = ConvergentMask >> (*LaneSource + 1);
-  *LaneSource += __ffs(WorkRemaining);
-  *IsFinal = __popc(WorkRemaining) == 1;
-  uint32_t lanemask_lt;
-  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(lanemask_lt));
-  uint32_t OmpId = __popc(ConvergentMask & lanemask_lt);
+  *LaneSource += __kmpc_impl_ffs(WorkRemaining);
+  *IsFinal = __kmpc_impl_popc(WorkRemaining) == 1;
+  uint32_t lanemask_lt = __kmpc_impl_lanemask_lt();
+  uint32_t OmpId = __kmpc_impl_popc(ConvergentMask & lanemask_lt);
 
   int threadId = GetLogicalThreadIdInBlock(isSPMDMode());
   int sourceThreadId = (threadId & ~(WARPSIZE - 1)) + *LaneSource;
 
   ConvergentParallelJob *job = (ConvergentParallelJob *)buffer;
   int32_t NumThreadsClause =
       omptarget_nvptx_threadPrivateContext->NumThreadsForNextParallel(threadId);
   job->tnumForNextPar = NumThreadsClause;

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/reduction.cu

 EXTERN
 void __kmpc_nvptx_end_reduce_nowait(int32_t global_tid) {}
 
 EXTERN int32_t __kmpc_shuffle_int32(int32_t val, int16_t delta, int16_t size) {
   return __kmpc_impl_shfl_down_sync(0xFFFFFFFF, val, delta, size);
 }
 
 EXTERN int64_t __kmpc_shuffle_int64(int64_t val, int16_t delta, int16_t size) {
-   int lo, hi;
-   asm volatile("mov.b64 {%0,%1}, %2;" : "=r"(lo), "=r"(hi) : "l"(val));
+   uint32_t lo, hi;
+   __kmpc_impl_unpack(val, lo, hi);
    hi = __kmpc_impl_shfl_down_sync(0xFFFFFFFF, hi, delta, size);
    lo = __kmpc_impl_shfl_down_sync(0xFFFFFFFF, lo, delta, size);
-   asm volatile("mov.b64 %0, {%1,%2};" : "=l"(val) : "r"(lo), "r"(hi));
-   return val;
+   return __kmpc_impl_pack(lo, hi);
 }
 
 INLINE static void gpu_regular_warp_reduce(void *reduce_data,
                                            kmp_ShuffleReductFctPtr shflFct) {
   for (uint32_t mask = WARPSIZE / 2; mask > 0; mask /= 2) {
     shflFct(reduce_data, /*LaneId - not used= */ 0,
             /*Offset = */ mask, /*AlgoVersion=*/0);
   }
while (mask > 0) {
     shflFct(reduce_data, /*LaneId = */ tid, /*Offset=*/mask, /*AlgoVersion=*/1);
     curr_size = (curr_size + 1) / 2;
     mask = curr_size / 2;
   }
 }
 
 INLINE static uint32_t
 gpu_irregular_simd_reduce(void *reduce_data, kmp_ShuffleReductFctPtr shflFct) {
-  uint32_t lanemask_lt;
-  uint32_t lanemask_gt;
   uint32_t size, remote_id, physical_lane_id;
   physical_lane_id = GetThreadIdInBlock() % WARPSIZE;
-  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(lanemask_lt));
+  uint32_t lanemask_lt = __kmpc_impl_lanemask_lt();
   uint32_t Liveness = __ACTIVEMASK();
-  uint32_t logical_lane_id = __popc(Liveness & lanemask_lt) * 2;
-  asm("mov.u32 %0, %%lanemask_gt;" : "=r"(lanemask_gt));
+  uint32_t logical_lane_id = __kmpc_impl_popc(Liveness & lanemask_lt) * 2;
+  uint32_t lanemask_gt = __kmpc_impl_lanemask_gt();
   do {
     Liveness = __ACTIVEMASK();
-    remote_id = __ffs(Liveness & lanemask_gt);
-    size = __popc(Liveness);
+    remote_id = __kmpc_impl_ffs(Liveness & lanemask_gt);
+    size = __kmpc_impl_popc(Liveness);
     logical_lane_id /= 2;
     shflFct(reduce_data, /*LaneId =*/logical_lane_id,
             /*Offset=*/remote_id - 1 - physical_lane_id, /*AlgoVersion=*/2);
   } while (logical_lane_id % 2 == 0 && size > 1);
   return (logical_lane_id == 0);
 }
 
 EXTERN
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   }
   return BlockThreadId == 0;
 #else
   uint32_t Liveness = __ACTIVEMASK();
   if (Liveness == 0xffffffff) // Full warp
     gpu_regular_warp_reduce(reduce_data, shflFct);
   else if (!(Liveness & (Liveness + 1))) // Partial warp but contiguous lanes
     gpu_irregular_warp_reduce(reduce_data, shflFct,
-                              /*LaneCount=*/__popc(Liveness),
+                              /*LaneCount=*/__kmpc_impl_popc(Liveness),
                               /*LaneId=*/GetThreadIdInBlock() % WARPSIZE);
   else if (!isRuntimeUninitialized) // Dispersed lanes. Only threads in L2
                                     // parallel region may enter here; return
                                     // early.
     return gpu_irregular_simd_reduce(reduce_data, shflFct);
 
   // When we have more than [warpsize] number of threads
   // a block reduction is performed here.
for (uint32_t i = NumThreads + ThreadId; i < NumTeams; i += NumThreads)
     ldFct(reduce_data, scratchpad, i, NumTeams, /*Load and reduce*/ 1);
 
   // Reduce across warps to the warp master.
   uint32_t Liveness = __ACTIVEMASK();
   if (Liveness == 0xffffffff) // Full warp
     gpu_regular_warp_reduce(reduce_data, shflFct);
   else // Partial warp but contiguous lanes
     gpu_irregular_warp_reduce(reduce_data, shflFct,
-                              /*LaneCount=*/__popc(Liveness),
+                              /*LaneCount=*/__kmpc_impl_popc(Liveness),
                               /*LaneId=*/ThreadId % WARPSIZE);
 
   // When we have more than [warpsize] number of threads
   // a block reduction is performed here.
   uint32_t ActiveThreads = NumTeams < NumThreads ? NumTeams : NumThreads;
   if (ActiveThreads > WARPSIZE) {
     uint32_t WarpsNeeded = (ActiveThreads + WARPSIZE - 1) / WARPSIZE;
     // Gather all the reduced values from each warp

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/supporti.h

 INLINE int IsTeamMaster(int ompThreadId) { return (ompThreadId == 0); }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Parallel level
 
 INLINE void IncParallelLevel(bool ActiveParallel) {
   unsigned Active = __ACTIVEMASK();
   __kmpc_impl_syncwarp(Active);
-  unsigned LaneMaskLt;
-  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(LaneMaskLt));
-  unsigned Rank = __popc(Active & LaneMaskLt);
+  unsigned LaneMaskLt = __kmpc_impl_lanemask_lt();
+  unsigned Rank = __kmpc_impl_popc(Active & LaneMaskLt);
   if (Rank == 0) {
     parallelLevel[GetWarpId()] +=
         (1 + (ActiveParallel ? OMP_ACTIVE_PARALLEL_LEVEL : 0));
     __threadfence();
   }
   __kmpc_impl_syncwarp(Active);
 }
 
 INLINE void DecParallelLevel(bool ActiveParallel) {
   unsigned Active = __ACTIVEMASK();
   __kmpc_impl_syncwarp(Active);
-  unsigned LaneMaskLt;
-  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(LaneMaskLt));
-  unsigned Rank = __popc(Active & LaneMaskLt);
+  unsigned LaneMaskLt = __kmpc_impl_lanemask_lt();
+  unsigned Rank = __kmpc_impl_popc(Active & LaneMaskLt);
   if (Rank == 0) {
     parallelLevel[GetWarpId()] -=
         (1 + (ActiveParallel ? OMP_ACTIVE_PARALLEL_LEVEL : 0));
     __threadfence();
   }
   __kmpc_impl_syncwarp(Active);
 }
 

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/target_impl.h

 //===----------------------------------------------------------------------===//
 #ifndef _TARGET_IMPL_H_
 #define _TARGET_IMPL_H_
 
 #include <stdint.h>
 
 #include "option.h"
 
-INLINE void __kmpc_impl_unpack(int64_t val, int32_t &lo, int32_t &hi) {
+INLINE void __kmpc_impl_unpack(uint64_t val, uint32_t &lo, uint32_t &hi) {
   asm volatile("mov.b64 {%0,%1}, %2;" : "=r"(lo), "=r"(hi) : "l"(val));
 }
 
-INLINE int64_t __kmpc_impl_pack(int32_t lo, int32_t hi) {
-  int64_t val;
+INLINE uint64_t __kmpc_impl_pack(uint32_t lo, uint32_t hi) {
+  uint64_t val;
   asm volatile("mov.b64 %0, {%1,%2};" : "=l"(val) : "r"(lo), "r"(hi));
   return val;
 }
 
 typedef uint32_t __kmpc_impl_lanemask_t;
 
 INLINE __kmpc_impl_lanemask_t __kmpc_impl_lanemask_lt() {
   __kmpc_impl_lanemask_t res;
   asm("mov.u32 %0, %%lanemask_lt;" : "=r"(res));
   return res;
 }
 
-INLINE int __kmpc_impl_ffs(uint32_t x) { return __ffs(x); }
+INLINE __kmpc_impl_lanemask_t __kmpc_impl_lanemask_gt() {
+  __kmpc_impl_lanemask_t res;
+  asm("mov.u32 %0, %%lanemask_gt;" : "=r"(res));
+  return res;
+}
+
+INLINE uint32_t __kmpc_impl_ffs(uint32_t x) { return __ffs(x); }
 
-INLINE int __kmpc_impl_popc(uint32_t x) { return __popc(x); }
+INLINE uint32_t __kmpc_impl_popc(uint32_t x) { return __popc(x); }
 
 #ifndef CUDA_VERSION
 #error CUDA_VERSION macro is undefined, something wrong with cuda.
 #endif
 
 // In Cuda 9.0, the *_sync() version takes an extra argument 'mask'.
 
 INLINE int32_t __kmpc_impl_shfl_sync(__kmpc_impl_lanemask_t Mask, int32_t Var,