math.hip

Go to the documentation of this file.

/*
 Copyright (c) 2021-2025, The Neko Authors
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions
 are met:
 
   * Redistributions of source code must retain the above copyright
     notice, this list of conditions and the following disclaimer.
 
   * Redistributions in binary form must reproduce the above
     copyright notice, this list of conditions and the following
     disclaimer in the documentation and/or other materials provided
     with the distribution.
 
   * Neither the name of the authors nor the names of its
     contributors may be used to endorse or promote products derived
     from this software without specific prior written permission.
 
 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
 COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
 ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 POSSIBILITY OF SUCH DAMAGE.
*/
 
#include <hip/hip_runtime.h>
#include <device/device_config.h>
#include <device/hip/check.h>
#include "math_kernel.h"
 
extern "C" {
 
#include <math/bcknd/device/device_mpi_reduce.h>
#include <math/bcknd/device/device_mpi_op.h>
 
#ifdef HAVE_RCCL
#include <math/bcknd/device/device_nccl_reduce.h>
#include <math/bcknd/device/device_nccl_op.h>
#endif
  
  void hip_copy(void *a, void *b, int *n) {
    HIP_CHECK(hipMemcpyAsync(a, b, (*n) * sizeof(real),
                             hipMemcpyDeviceToDevice,
                             (hipStream_t) glb_cmd_queue));
  }
  void hip_copy(void *a, void *b, int *n) {…}
 
  void hip_masked_copy(void *a, void *b, void *mask, int *n, int *m) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*m)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(masked_copy_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (int *) mask, *n, *m);
 
    HIP_CHECK(hipGetLastError());
 
  }
  void hip_masked_copy(void *a, void *b, void *mask, int *n, int *m) {…}
  
 
  void hip_masked_red_copy(void *a, void *b, void *mask, int *n, int *m) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*m)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(masked_red_copy_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (int *) mask, *n, *m);
 
    HIP_CHECK(hipGetLastError());
 
  } 
  void hip_masked_red_copy(void *a, void *b, void *mask, int *n, int *m) {…}
 
  void hip_masked_atomic_reduction(void *a, void *b, void *mask, int *n, int *m) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*m)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(masked_atomic_reduction_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (int *) mask, *n, *m);
 
    HIP_CHECK(hipGetLastError());
 
  } 
  void hip_masked_atomic_reduction(void *a, void *b, void *mask, int *n, int *m) {…}
 
  void hip_cfill_mask(void* a, real* c, int* size, void* mask, int* mask_size) {
 
      const dim3 nthrds(1024, 1, 1);
      const dim3 nblcks(((*mask_size) + 1024 - 1) / 1024, 1, 1);
      
    hipLaunchKernelGGL(HIP_KERNEL_NAME(cfill_mask_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real*)a, *c, *size, (int*)mask, *mask_size);
 
    HIP_CHECK(hipGetLastError());
  }
  void hip_cfill_mask(void* a, real* c, int* size, void* mask, int* mask_size) {…}
 
  void hip_rzero(void *a, int *n) {
    HIP_CHECK(hipMemsetAsync(a, 0, (*n) * sizeof(real),
                             (hipStream_t) glb_cmd_queue));
  }
  void hip_rzero(void *a, int *n) {…}
 
  void hip_cmult(void *a, real *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(cmult_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, *c, *n);
    HIP_CHECK(hipGetLastError());
    
  }
  void hip_cmult(void *a, real *c, int *n) {…}
 
  void hip_cmult2(void *a, void *b, real *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(cmult2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a,(real *) b, *c, *n);
    HIP_CHECK(hipGetLastError());
    
  }
  void hip_cmult2(void *a, void *b, real *c, int *n) {…}
  void hip_cadd(void *a, real *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(cadd_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, *c, *n);
    HIP_CHECK(hipGetLastError()); 
  }
  void hip_cadd(void *a, real *c, int *n) {…}
 
  void hip_cadd2(void *a, void *b, real *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(cadd2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, *c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_cadd2(void *a, void *b, real *c, int *n) {…}
 
  void hip_cfill(void *a, real *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    if (*n > 0) { 
      hipLaunchKernelGGL(HIP_KERNEL_NAME(cfill_kernel<real>),
                         nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                         (real *) a, *c, *n);
      HIP_CHECK(hipGetLastError());
    }
  }
  void hip_cfill(void *a, real *c, int *n) {…}
 
  void hip_add2(void *a, void *b, int *n) {
    
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add2(void *a, void *b, int *n) {…}
  
  void hip_add4(void *a, void *b, void *c, void *d, int *n) {
    
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add4_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, (real *) d, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add4(void *a, void *b, void *c, void *d, int *n) {…}
 
  void hip_add2s1(void *a, void *b, real *c1, int *n) {
    
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add2s1_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b,
                       *c1, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add2s1(void *a, void *b, real *c1, int *n) {…}
 
  void hip_add2s2(void *a, void *b, real *c1, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add2s2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b,
                       *c1, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add2s2(void *a, void *b, real *c1, int *n) {…}
 
  void hip_addsqr2s2(void *a, void *b, real *c1, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(addsqr2s2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b,
                       *c1, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_addsqr2s2(void *a, void *b, real *c1, int *n) {…}
  
  void hip_add3s2(void *a, void *b, void *c, real *c1, real *c2, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add3s2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c,
                       *c1, *c2, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add3s2(void *a, void *b, void *c, real *c1, real *c2, int *n) {…}
  
  void hip_invcol1(void *a, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(invcol1_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_invcol1(void *a, int *n) {…}
 
  void hip_invcol2(void *a, void *b, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(invcol2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_invcol2(void *a, void *b, int *n) {…}
  
  void hip_col2(void *a, void *b, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(col2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_col2(void *a, void *b, int *n) {…}
  
  void hip_col3(void *a, void *b, void *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(col3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_col3(void *a, void *b, void *c, int *n) {…}
 
  void hip_subcol3(void *a, void *b, void *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(subcol3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_subcol3(void *a, void *b, void *c, int *n) {…}
  
  void hip_sub2(void *a, void *b, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(sub2_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_sub2(void *a, void *b, int *n) {…}
 
  void hip_sub3(void *a, void *b, void *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(sub3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_sub3(void *a, void *b, void *c, int *n) {…}
 
  void hip_addcol3(void *a, void *b, void *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(addcol3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_addcol3(void *a, void *b, void *c, int *n) {…}
 
  void hip_addcol4(void *a, void *b, void *c, void *d, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(addcol4_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, (real *) d, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_addcol4(void *a, void *b, void *c, void *d, int *n) {…}
 
  void hip_vdot3(void *dot, void *u1, void *u2, void *u3,
                 void *v1, void *v2, void *v3, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(vdot3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) dot, (real *) u1, (real *) u2, (real *) u3,
                       (real *) v1, (real *) v2, (real *) v3, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_vdot3(void *dot, void *u1, void *u2, void *u3, {…}
 
  void hip_vcross(void *u1, void *u2, void *u3,
                 void *v1, void *v2, void *v3,
                 void *w1, void *w2, void *w3, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(vcross_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) u1, (real *) u2, (real *) u3,
                       (real *) v1, (real *) v2, (real *) v3, 
                       (real *) w1, (real *) w2, (real *) w3, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_vcross(void *u1, void *u2, void *u3, {…}
 
 
  /*
   * Reduction buffer
   */
  int red_s = 0;
  real * bufred = NULL;
  real * bufred_d = NULL;
  
  real hip_vlsc3(void *u, void *v, void *w, int *n) {
        
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    const int nb = ((*n) + 1024 - 1)/ 1024;
    const hipStream_t stream = (hipStream_t) glb_cmd_queue;
    
    if ( nb > red_s){
      red_s = nb;
      if (bufred != NULL) {
        HIP_CHECK(hipHostFree(bufred));
        HIP_CHECK(hipFree(bufred_d));        
      }
      HIP_CHECK(hipHostMalloc(&bufred,nb*sizeof(real),hipHostMallocDefault));
      HIP_CHECK(hipMalloc(&bufred_d, nb*sizeof(real)));
    }
     
    hipLaunchKernelGGL(HIP_KERNEL_NAME(glsc3_kernel<real>),
                       nblcks, nthrds, 0, stream,
                       (real *) u, (real *) v,
                       (real *) w, bufred_d, *n);
    HIP_CHECK(hipGetLastError());
    hipLaunchKernelGGL(HIP_KERNEL_NAME(reduce_kernel<real>),
                       1, 1024, 0, stream, bufred_d, nb);
    HIP_CHECK(hipGetLastError());
 
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d, sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
 
    return bufred[0];
  }
  real hip_vlsc3(void *u, void *v, void *w, int *n) {…}
  
 
  real hip_glsc3(void *a, void *b, void *c, int *n) {
        
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    const int nb = ((*n) + 1024 - 1)/ 1024;
    const hipStream_t stream = (hipStream_t) glb_cmd_queue;
    
    if ( nb > red_s){
      red_s = nb;
      if (bufred != NULL) {
        HIP_CHECK(hipHostFree(bufred));
        HIP_CHECK(hipFree(bufred_d));        
      }
      HIP_CHECK(hipHostMalloc(&bufred,nb*sizeof(real),hipHostMallocDefault));
      HIP_CHECK(hipMalloc(&bufred_d, nb*sizeof(real)));
    }
     
    hipLaunchKernelGGL(HIP_KERNEL_NAME(glsc3_kernel<real>),
                       nblcks, nthrds, 0, stream,
                       (real *) a, (real *) b,
                       (real *) c, bufred_d, *n);
    HIP_CHECK(hipGetLastError());
    hipLaunchKernelGGL(HIP_KERNEL_NAME(reduce_kernel<real>),
                       1, 1024, 0, stream, bufred_d, nb);
    HIP_CHECK(hipGetLastError());
 
#ifdef HAVE_RCCL
    device_nccl_allreduce(bufred_d, bufred_d, 1, sizeof(real),
                          DEVICE_NCCL_SUM, stream);
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d, sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#elif HAVE_DEVICE_MPI
    HIP_CHECK(hipStreamSynchronize(stream));
    device_mpi_allreduce(bufred_d, bufred, 1, sizeof(real), DEVICE_MPI_SUM);
#else
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d, sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#endif
    return bufred[0];
  }
  real hip_glsc3(void *a, void *b, void *c, int *n) {…}
 
  void hip_glsc3_many(real *h, void * w, void *v,void *mult, int *j, int *n){ 
    int pow2 = 1;
    while(pow2 < (*j)){
      pow2 = 2*pow2;
    }
    const int nt = 1024/pow2;   
    const dim3 nthrds(pow2, nt, 1);
    const dim3 nblcks(((*n)+nt - 1)/nt, 1, 1);
    const dim3 nthrds_red(1024,1,1);
    const dim3 nblcks_red( (*j),1,1);
    const int nb = ((*n) + nt - 1)/nt;
    const hipStream_t stream = (hipStream_t) glb_cmd_queue;
    
    if((*j)*nb>red_s){
      red_s = (*j)*nb;
      if (bufred != NULL) {
        HIP_CHECK(hipHostFree(bufred));
        HIP_CHECK(hipFree(bufred_d));
      }      
      HIP_CHECK(hipHostMalloc(&bufred,(*j)*nb*sizeof(real),hipHostMallocDefault));
      HIP_CHECK(hipMalloc(&bufred_d, (*j)*nb*sizeof(real)));
    }
    hipLaunchKernelGGL(HIP_KERNEL_NAME(glsc3_many_kernel<real>),
                       nblcks, nthrds, 0, stream,
                       (const real *) w, (const real **) v,
                       (const real *)mult, bufred_d, *j, *n);
    HIP_CHECK(hipGetLastError());
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(glsc3_reduce_kernel<real>),
                       nblcks_red, nthrds_red, 0, stream,
                       bufred_d, nb, *j);
    HIP_CHECK(hipGetLastError());
 
#ifdef HAVE_RCCL
    device_nccl_allreduce(bufred_d, bufred_d, (*j), sizeof(real),
                          DEVICE_NCCL_SUM, stream);
    HIP_CHECK(hipMemcpyAsync(h, bufred_d, (*j)* sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#elif HAVE_DEVICE_MPI
    HIP_CHECK(hipStreamSynchronize(stream));
    device_mpi_allreduce(bufred_d, h, (*j), sizeof(real), DEVICE_MPI_SUM);
#else
    HIP_CHECK(hipMemcpyAsync(h, bufred_d, (*j)* sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#endif    
  }
  void hip_glsc3_many(real *h, void * w, void *v,void *mult, int *j, int *n) {…}
 
  void hip_add2s2_many(void *x, void **p, void *alpha, int *j, int *n) {
        
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add2s2_many_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) x, (const real **) p, (real *) alpha, *j, *n);
    HIP_CHECK(hipGetLastError());
 
  }
  void hip_add2s2_many(void *x, void **p, void *alpha, int *j, int *n) {…}
 
  void hip_add3(void *a, void *b, void *c, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
 
    hipLaunchKernelGGL(HIP_KERNEL_NAME(add3_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, (real *) b, (real *) c, *n);
    HIP_CHECK(hipGetLastError());
  }
  void hip_add3(void *a, void *b, void *c, int *n) {…}
 
  real hip_glsc2(void *a, void *b, int *n) {
        
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    const int nb = ((*n) + 1024 - 1)/ 1024;
    const hipStream_t stream = (hipStream_t) glb_cmd_queue;
        
    if ( nb > red_s){
      red_s = nb;
      if (bufred != NULL) {
        HIP_CHECK(hipHostFree(bufred));
        HIP_CHECK(hipFree(bufred_d));        
      }
      HIP_CHECK(hipHostMalloc(&bufred,nb*sizeof(real),hipHostMallocDefault));
      HIP_CHECK(hipMalloc(&bufred_d, nb*sizeof(real)));
    }
     
    hipLaunchKernelGGL(HIP_KERNEL_NAME(glsc2_kernel<real>),
                       nblcks, nthrds, 0, stream,
                       (real *) a, (real *) b, bufred_d, *n);
    HIP_CHECK(hipGetLastError());
    hipLaunchKernelGGL(HIP_KERNEL_NAME(reduce_kernel<real>),
                       1, 1024, 0, stream, bufred_d, nb);
    HIP_CHECK(hipGetLastError());
 
#ifdef HAVE_RCCL
    device_nccl_allreduce(bufred_d, bufred_d, 1, sizeof(real),
                          DEVICE_NCCL_SUM, stream);
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d,sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#elif HAVE_DEVICE_MPI
    HIP_CHECK(hipStreamSynchronize(stream));
    device_mpi_allreduce(bufred_d, bufred, 1, sizeof(real), DEVICE_MPI_SUM);
#else
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d, sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#endif
    return bufred[0];
  }
  real hip_glsc2(void *a, void *b, int *n) {…}
 
  real hip_glsum(void *a, int *n) {
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    const int nb = ((*n) + 1024 - 1)/ 1024;
    const hipStream_t stream = (hipStream_t) glb_cmd_queue;
    
    if ( nb > red_s){
      red_s = nb;
      if (bufred != NULL) {
        HIP_CHECK(hipHostFree(bufred));
        HIP_CHECK(hipFree(bufred_d));        
      }
      HIP_CHECK(hipHostMalloc(&bufred,nb*sizeof(real),hipHostMallocDefault));
      HIP_CHECK(hipMalloc(&bufred_d, nb*sizeof(real)));
    }
    if( *n > 0) { 
      hipLaunchKernelGGL(HIP_KERNEL_NAME(glsum_kernel<real>),
                       nblcks, nthrds, 0, stream,
                       (real *) a, bufred_d, *n);
      HIP_CHECK(hipGetLastError());
      hipLaunchKernelGGL(HIP_KERNEL_NAME(reduce_kernel<real>),
                       1, 1024, 0, stream, bufred_d, nb);
      HIP_CHECK(hipGetLastError());
    }
 
#ifdef HAVE_RCCL
    device_nccl_allreduce(bufred_d, bufred_d, 1, sizeof(real),
                          DEVICE_NCCL_SUM, stream);
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d,sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#elif HAVE_DEVICE_MPI
    HIP_CHECK(hipStreamSynchronize(stream));
    device_mpi_allreduce(bufred_d, bufred, 1, sizeof(real), DEVICE_MPI_SUM);
#else
    HIP_CHECK(hipMemcpyAsync(bufred, bufred_d,sizeof(real),
                             hipMemcpyDeviceToHost, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
#endif    
    return bufred[0];
  }
  real hip_glsum(void *a, int *n) {…}
 
  void hip_absval(void *a, int *n) {
 
    const dim3 nthrds(1024, 1, 1);
    const dim3 nblcks(((*n)+1024 - 1)/ 1024, 1, 1);
    
    hipLaunchKernelGGL(HIP_KERNEL_NAME(absval_kernel<real>),
                       nblcks, nthrds, 0, (hipStream_t) glb_cmd_queue,
                       (real *) a, *n);
    HIP_CHECK(hipGetLastError());
    
  }
  void hip_absval(void *a, int *n) {…}
}