/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/utility/workgroup_synchronization.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/utility/workgroup_synchronization.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/utility/workgroup_synchronization.hpp Source File

Go to the documentation of this file.

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
 #include "ck/host_utility/hip_check_error.hpp"
  
 namespace ck {
  
 // Initialization flag of Barrier object, can be any value except for zero
 static constexpr int BarrierInitFlag = 0x7856;
  
 // 1) only the first thread-block in the synchronizaton group is supposed to call this function. It
 // is the responsibility of the user to ensure the two integer values in p_control_bits are zeros
 // before calling gms_init().
 // 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
 // repetitious initialization of p_control_bits buffer is required
 static __device__ void gms_init(int NumWarps, int* p_control_bits)
 {
     union
     {
         int two32[2];
         unsigned long one64;
     } regs;
  
     regs.two32[0] = BarrierInitFlag;
     regs.two32[1] = NumWarps;
  
     if(threadIdx.x == 0)
         atomicCAS(reinterpret_cast<unsigned long*>(p_control_bits), 0, regs.one64);
 };
  
 // all the workgroups in the synchronization group is supposed to call this function
 static __device__ void gms_barrier(int* p_control_bits)
 {
     constexpr int mask = WarpSize - 1;
  
     if((threadIdx.x & mask) == 0)
     {
         // ensure the barrier object is initialized
         do
         {
             const int r0 = __atomic_load_n(&p_control_bits[0], __ATOMIC_RELAXED);
  
             if(r0 == BarrierInitFlag)
                 break;
  
         } while(true);
  
         // go ahead toward the barrier line
         atomicSub(&p_control_bits[1], 1);
  
         // wait until all warps have arrived
         do
         {
             const int r1 = __atomic_load_n(&p_control_bits[1], __ATOMIC_RELAXED);
  
             if(r1 == 0)
                 break;
  
         } while(true);
     };
 };
  
 // 1) Only the first thread-block in the synchronizaton group is supposed to call this function.
 // 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
 // repetitious initialization of p_control_bits buffer is required
 static __device__ void gms_reset(int* p_control_bits)
 {
     // reset the barrier object
     if(threadIdx.x == 0)
         (void)atomicCAS(&p_control_bits[0], BarrierInitFlag, 0);
 };
  
 } // namespace ck