TL/MLX5: clean and fix after rebase

lintrunner cleaning
openucx · Dec 18, 2024 · 43dd1d7 · 43dd1d7
1 parent b8fe903
commit 43dd1d7
Show file tree

Hide file tree

Showing 8 changed files with 194 additions and 161 deletions.
diff --git a/src/components/tl/mlx5/alltoall/alltoall.h b/src/components/tl/mlx5/alltoall/alltoall.h
@@ -58,9 +58,9 @@ typedef struct ucc_tl_mlx5_alltoall_node {
     struct mlx5dv_mkey       *team_recv_mkey;
     void                     *umr_entries_buf;
     struct ibv_mr            *umr_entries_mr;
-    int                  fanin_index;
-    int                  fanin_dist;
-    int                  fanin_max_dist;
+    int                       fanin_index;
+    int                       fanin_dist;
+    int                       fanin_max_dist;
 } ucc_tl_mlx5_alltoall_node_t;
 
 typedef struct alltoall_net_ctrl {

diff --git a/src/components/tl/mlx5/alltoall/alltoall_coll.c b/src/components/tl/mlx5/alltoall/alltoall_coll.c
diff --git a/src/components/tl/mlx5/alltoall/alltoall_mkeys.c b/src/components/tl/mlx5/alltoall/alltoall_mkeys.c
@@ -299,7 +299,7 @@ ucc_status_t ucc_tl_mlx5_populate_send_recv_mkeys(ucc_tl_mlx5_team_t *    team,
     if (ucc_tl_mlx5_get_my_ctrl(a2a, seq_index)->mkey_cache_flag &
         UCC_MLX5_NEED_SEND_MKEY_UPDATE) {
         repeat_count = nbc ? a2a->net.sbgp->group_size
-                        : UCC_TL_TEAM_SIZE(team) / req->alltoall.block_width;
+                           : UCC_TL_TEAM_SIZE(team) / req->alltoall.block_width;
         for (i = 0; i < n_mkeys; i++) {
             status = populate_strided_mkey(a2a, send_mem_access_flags,
                                            node->ops[seq_index].send_mkeys[i],
@@ -314,8 +314,9 @@ ucc_status_t ucc_tl_mlx5_populate_send_recv_mkeys(ucc_tl_mlx5_team_t *    team,
     }
     if (ucc_tl_mlx5_get_my_ctrl(a2a, seq_index)->mkey_cache_flag &
         UCC_MLX5_NEED_RECV_MKEY_UPDATE) {
-        repeat_count = nbc ? a2a->net.sbgp->group_size
-                        : UCC_TL_TEAM_SIZE(team) / req->alltoall.block_height;
+        repeat_count =
+            nbc ? a2a->net.sbgp->group_size
+                : UCC_TL_TEAM_SIZE(team) / req->alltoall.block_height;
         for (i = 0; i < n_mkeys; i++) {
             status = populate_strided_mkey(a2a, recv_mem_access_flags,
                                            node->ops[seq_index].recv_mkeys[i],
@@ -336,7 +337,7 @@ static void update_mkey_entry(ucc_tl_mlx5_alltoall_t *a2a,
 {
     ucc_tl_mlx5_alltoall_node_t  *node       = &a2a->node;
     int                           block_height = req->alltoall.block_height;
-    int                           block_width = req->alltoall.block_width;
+    int                           block_width  = req->alltoall.block_width;
     size_t                        msg_size   = req->alltoall.msg_size;
     int                           nbc  = req->alltoall.num_of_blocks_columns;
     struct ibv_mr                *buff = direction_send
@@ -349,7 +350,8 @@ static void update_mkey_entry(ucc_tl_mlx5_alltoall_t *a2a,
         mkey_entry = (umr_t *)(direction_send ? MY_SEND_UMR_DATA(req, a2a, 0)
                                               : MY_RECV_UMR_DATA(req, a2a, 0));
         mkey_entry->addr        = (uintptr_t)buff->addr;
-        mkey_entry->bytes_count = (direction_send? block_width : block_height) * msg_size;
+        mkey_entry->bytes_count =
+            (direction_send ? block_width : block_height) * msg_size;
         mkey_entry->bytes_skip  = 0;
         mkey_entry->lkey        = direction_send ? buff->lkey : buff->rkey;
     } else {

diff --git a/src/components/tl/mlx5/tl_mlx5.c b/src/components/tl/mlx5/tl_mlx5.c
@@ -28,12 +28,15 @@ static ucc_config_field_t ucc_tl_mlx5_lib_config_table[] = {
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, dm_buf_num),
      UCC_CONFIG_TYPE_ULUNITS},
 
-    {"FORCE_REGULAR", "y", "Force the regular case where the block dimensions "
-    "divide ppn. Requires BLOCK_SIZE=0",
-     ucc_offsetof(ucc_tl_mlx5_lib_config_t, force_regular), UCC_CONFIG_TYPE_BOOL},
+    {"FORCE_REGULAR", "y",
+     "Force the regular case where the block dimensions "
+     "divide ppn. Requires BLOCK_SIZE=0",
+     ucc_offsetof(ucc_tl_mlx5_lib_config_t, force_regular),
+     UCC_CONFIG_TYPE_BOOL},
 
     {"FORCE_LONGER", "y", "Force the blocks to have more height than width",
-     ucc_offsetof(ucc_tl_mlx5_lib_config_t, force_longer), UCC_CONFIG_TYPE_BOOL},
+     ucc_offsetof(ucc_tl_mlx5_lib_config_t, force_longer),
+     UCC_CONFIG_TYPE_BOOL},
 
     {"FORCE_WIDER", "n", "Force the blocks to have more width than height",
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, force_wider), UCC_CONFIG_TYPE_BOOL},
@@ -118,18 +121,21 @@ static ucc_config_field_t ucc_tl_mlx5_lib_config_table[] = {
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, fanin_kn_radix),
      UCC_CONFIG_TYPE_UINT},
 
-    {"SEND_BATCH_SIZE", "1", "number of blocks that are transposed "
-    "on the NIC before being sent as a batch to a remote peer",
+    {"SEND_BATCH_SIZE", "1",
+     "number of blocks that are transposed "
+     "on the NIC before being sent as a batch to a remote peer",
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, block_batch_size),
      UCC_CONFIG_TYPE_UINT},
 
-    {"NBR_SERIALIZED_BATCHES", "1", "number of block batches "
-    "(within the set of blocks to be sent to a given remote peer) "
+    {"NBR_SERIALIZED_BATCHES", "1",
+     "number of block batches "
+     "(within the set of blocks to be sent to a given remote peer) "
      "serialized on the same device memory chunk",
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, nbr_serialized_batches),
      UCC_CONFIG_TYPE_UINT},
 
-    {"NBR_BATCHES_PER_PASSAGE", "32", "",
+    {"NBR_BATCHES_PER_PASSAGE", "32",
+     "number of batches of blocks sent to one remote node before enqueing",
      ucc_offsetof(ucc_tl_mlx5_lib_config_t, nbr_batches_per_passage),
      UCC_CONFIG_TYPE_UINT},
 
@@ -155,7 +161,8 @@ static ucc_config_field_t ucc_tl_mlx5_context_config_table[] = {
      ucc_offsetof(ucc_tl_mlx5_context_config_t, mcast_ctx_conf.ib_dev_name),
      UCC_CONFIG_TYPE_STRING},
     {"FANIN_NPOLLS", "1000",
-     "Number of shared memory polling before returning UCC_INPROGRESS during internode FANIN",
+     "Number of shared memory polling before returning UCC_INPROGRESS during "
+     "internode FANIN",
      ucc_offsetof(ucc_tl_mlx5_context_config_t, npolls), UCC_CONFIG_TYPE_UINT},
 
     {NULL}};

diff --git a/src/components/tl/mlx5/tl_mlx5.h b/src/components/tl/mlx5/tl_mlx5.h
@@ -65,8 +65,8 @@ typedef struct ucc_tl_mlx5_lib_config {
     int                                     nbr_batches_per_passage;
     int                                     block_batch_size;
     int                                     force_regular;
-    int                      force_longer;
-    int                      force_wider;
+    int                                     force_longer;
+    int                                     force_wider;
 } ucc_tl_mlx5_lib_config_t;
 
 typedef struct ucc_tl_mlx5_context_config {
@@ -103,7 +103,7 @@ typedef struct ucc_tl_mlx5_task ucc_tl_mlx5_task_t;
 typedef struct ucc_tl_mlx5_schedule ucc_tl_mlx5_schedule_t;
 typedef struct ucc_tl_mlx5_dm_chunk_t {
     uintptr_t addr; // 0 based offset from the beginning of
-                      // memic_mr (obtained with ibv_reg_dm_mr)
+                    // memic_mr (obtained with ibv_reg_dm_mr)
     ucc_tl_mlx5_schedule_t *task;
     int                     posted_sends;
     int                     posted_all;

diff --git a/src/components/tl/mlx5/tl_mlx5_dm.c b/src/components/tl/mlx5/tl_mlx5_dm.c
@@ -77,21 +77,16 @@ static void ucc_tl_mlx5_dm_chunk_init(ucc_mpool_t *mp,        //NOLINT
     ucc_tl_mlx5_dm_chunk_t *c    = (ucc_tl_mlx5_dm_chunk_t *)obj;
     ucc_tl_mlx5_team_t     *team =
         ucc_container_of(mp, ucc_tl_mlx5_team_t, dm_pool);
-    c->addr         = (uintptr_t)PTR_OFFSET(
-                                     (UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_host)?
-                                     team->dm_ptr : NULL,
-                                  team->dm_offset);
-    team->dm_offset =  team->dm_offset +
-                            UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_buf_size
-                            * UCC_TL_MLX5_TEAM_LIB(team)->cfg.block_batch_size;
-    c->posted_sends = 0;
-    c->posted_all=0;
-    c->completed_sends      = 0;
-}
 
-    c->offset       = (ptrdiff_t)team->dm_offset;
-    team->dm_offset = PTR_OFFSET(team->dm_offset,
-                                 UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_buf_size);
+    c->addr = (uintptr_t)PTR_OFFSET(
+        (UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_host) ? team->dm_ptr : NULL,
+        team->dm_offset);
+    c->posted_sends    = 0;
+    c->posted_all      = 0;
+    c->completed_sends = 0;
+    team->dm_offset =
+        team->dm_offset + UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_buf_size *
+                              UCC_TL_MLX5_TEAM_LIB(team)->cfg.block_batch_size;
 }
 
 static ucc_mpool_ops_t ucc_tl_mlx5_dm_ops = {
@@ -230,17 +225,19 @@ ucc_status_t ucc_tl_mlx5_dm_init(ucc_tl_mlx5_team_t *team)
     }
 
     status = ucc_tl_mlx5_dm_alloc_reg(
-        ctx->shared_ctx, ctx->shared_pd, cfg->dm_host, cfg->dm_buf_size * cfg->block_batch_size,
-        &cfg->dm_buf_num, &team->dm_ptr, &team->dm_mr, UCC_TL_TEAM_LIB(team));
+        ctx->shared_ctx, ctx->shared_pd, cfg->dm_host,
+        cfg->dm_buf_size * cfg->block_batch_size, &cfg->dm_buf_num,
+        &team->dm_ptr, &team->dm_mr, UCC_TL_TEAM_LIB(team));
     if (status != UCC_OK) {
         goto err_dm_alloc;
     }
     team->dm_offset = 0;
-    // TODO: fix case dm_host=true
-    status = ucc_mpool_init(&team->dm_pool, 0, sizeof(ucc_tl_mlx5_dm_chunk_t),
-                            0, UCC_CACHE_LINE_SIZE, 1,
-                            cfg->dm_buf_num, &ucc_tl_mlx5_dm_ops,
-                            ctx->super.super.ucc_context->thread_mode, "mlx5 dm pool");
+    // TODO: fix/check the case dm_host=true
+    ucc_assert(!cfg->dm_host);
+    status = ucc_mpool_init(
+        &team->dm_pool, 0, sizeof(ucc_tl_mlx5_dm_chunk_t), 0,
+        UCC_CACHE_LINE_SIZE, 1, cfg->dm_buf_num, &ucc_tl_mlx5_dm_ops,
+        ctx->super.super.ucc_context->thread_mode, "mlx5 dm pool");
     if (status != UCC_OK) {
         tl_debug(UCC_TL_TEAM_LIB(team), "failed to init dm pool");
         goto err_mpool_init;

diff --git a/src/components/tl/mlx5/tl_mlx5_wqe.c b/src/components/tl/mlx5/tl_mlx5_wqe.c
@@ -39,7 +39,8 @@ static inline uint8_t get_umr_mr_flags(uint32_t acc)
 
 typedef struct transpose_seg {
     __be32 element_size; /* 8 bit value */
-    __be16 num_cols;     /* 7 bit value */  //TODO: from PRM we should have the rows first and then the colls... is this a bug ?
+    //From PRM we should have the rows first and then the colls. This is probably a naming error
+    __be16 num_cols;     /* 7 bit value */
     __be16 num_rows;     /* 7 bit value */
     __be64 padding;
 } transpose_seg_t;

diff --git a/test/mpi/buffer.cc b/test/mpi/buffer.cc
@@ -182,13 +182,6 @@ ucc_status_t compare_buffers(void *_rst, void *expected, size_t count,
     } else {
         status = memcmp(rst, expected, count*ucc_dt_size(dt)) ?
             UCC_ERR_NO_MESSAGE : UCC_OK;
-        // uint8_t* a = (uint8_t*)rst;
-        // uint8_t* b = (uint8_t*)expected;
-        // for (int i=0; i<count*ucc_dt_size(dt); i++ ){
-        //     if (a[i] != b[i]) {
-        //         printf("!?!?!?!?!? FAILUUUUUURE at i=%d, a=%hhn, b=%hhn, range=%ld\n", i, a, b, count*ucc_dt_size(dt));
-        //     }
-        // }
     }
 
     if (UCC_MEMORY_TYPE_HOST != mt && UCC_MEMORY_TYPE_CUDA_MANAGED != mt) {