From 9f9ecd9a3887a6d25066160488b27f99edd906cf Mon Sep 17 00:00:00 2001
From: avinashkethineedi <avinash.kethineedi@amd.com>
Date: Mon, 28 Oct 2024 21:35:14 +0000
Subject: [PATCH] Remove the device-side active-set-based reduction API
 interface from rocSHMEM

[ROCm/rocshmem commit: c22048112e3800597ac049e681327ee1a7474cb6]
---
 .../rocshmem/include/roc_shmem/roc_shmem.hpp  |  4 ----
 projects/rocshmem/src/roc_shmem_gpu.cpp       | 20 -------------------
 2 files changed, 24 deletions(-)

diff --git a/projects/rocshmem/include/roc_shmem/roc_shmem.hpp b/projects/rocshmem/include/roc_shmem/roc_shmem.hpp
index c0484e8bd7..7a0b58ad6a 100644
--- a/projects/rocshmem/include/roc_shmem/roc_shmem.hpp
+++ b/projects/rocshmem/include/roc_shmem/roc_shmem.hpp
@@ -837,10 +837,6 @@ __device__ ATTR_NO_INLINE void roc_shmem_threadfence_system();
  * MACRO DECLARE SHMEM_REDUCTION APIs
  */
 #define REDUCTION_API_GEN(T, TNAME, Op_API)                                    \
-  __device__ ATTR_NO_INLINE void roc_shmem_ctx_##TNAME##_##Op_API##_wg_to_all( \
-      roc_shmem_ctx_t ctx, T *dest, const T *source, int nreduce,              \
-      int PE_start, int logPE_stride, int PE_size, T *pWrk,                    \
-      long *pSync); /* NOLINT */                                               \
   __device__ ATTR_NO_INLINE void roc_shmem_ctx_##TNAME##_##Op_API##_wg_to_all( \
       roc_shmem_ctx_t ctx, roc_shmem_team_t team, T *dest, const T *source,    \
       int nreduce);                                                            \
diff --git a/projects/rocshmem/src/roc_shmem_gpu.cpp b/projects/rocshmem/src/roc_shmem_gpu.cpp
index 582249eec9..87635a4160 100644
--- a/projects/rocshmem/src/roc_shmem_gpu.cpp
+++ b/projects/rocshmem/src/roc_shmem_gpu.cpp
@@ -429,17 +429,6 @@ __device__ void *roc_shmem_ptr(const void *dest, int pe) {
   return get_internal_ctx(ROC_SHMEM_CTX_DEFAULT)->shmem_ptr(dest, pe);
 }
 
-template <typename T, ROC_SHMEM_OP Op>
-__device__ void roc_shmem_wg_to_all(roc_shmem_ctx_t ctx, T *dest,
-                                    const T *source, int nreduce, int PE_start,
-                                    int logPE_stride, int PE_size, T *pWrk,
-                                    long *pSync) {
-  GPU_DPRINTF("Function: roc_shmem_to_all\n");
-
-  get_internal_ctx(ctx)->to_all<T, Op>(dest, source, nreduce, PE_start,
-                                       logPE_stride, PE_size, pWrk, pSync);
-}
-
 template <typename T, ROC_SHMEM_OP Op>
 __device__ void roc_shmem_wg_to_all(roc_shmem_ctx_t ctx, roc_shmem_team_t team,
                                     T *dest, const T *source, int nreduce) {
@@ -886,9 +875,6 @@ __device__ int roc_shmem_team_translate_pe(roc_shmem_team_t src_team,
  * Template generator for reductions
  */
 #define REDUCTION_GEN(T, Op)                                                 \
-  template __device__ void roc_shmem_wg_to_all<T, Op>(                       \
-      roc_shmem_ctx_t ctx, T * dest, const T *source, int nreduce,           \
-      int PE_start, int logPE_stride, int PE_size, T *pWrk, long *pSync);    \
   template __device__ void roc_shmem_wg_to_all<T, Op>(                       \
       roc_shmem_ctx_t ctx, roc_shmem_team_t team, T * dest, const T *source, \
       int nreduce);
@@ -1100,12 +1086,6 @@ __device__ int roc_shmem_team_translate_pe(roc_shmem_team_t src_team,
  **/
 
 #define REDUCTION_DEF_GEN(T, TNAME, Op_API, Op)                             \
-  __device__ void roc_shmem_ctx_##TNAME##_##Op_API##_wg_to_all(             \
-      roc_shmem_ctx_t ctx, T *dest, const T *source, int nreduce,           \
-      int PE_start, int logPE_stride, int PE_size, T *pWrk, long *pSync) {  \
-    roc_shmem_wg_to_all<T, Op>(ctx, dest, source, nreduce, PE_start,        \
-                               logPE_stride, PE_size, pWrk, pSync);         \
-  }                                                                         \
   __device__ void roc_shmem_ctx_##TNAME##_##Op_API##_wg_to_all(             \
       roc_shmem_ctx_t ctx, roc_shmem_team_t team, T *dest, const T *source, \
       int nreduce) {                                                        \