UCT/GDA: Use wqe_idx for rc_gda progress

src/tools/perf/cuda/ucp_cuda_kernel.cu

@@ -228,7 +228,7 @@ ucp_perf_cuda_send_sync(ucp_perf_cuda_params &params, ucx_perf_counter_t idx,
                         ucp_device_request_t &req)
 {
     ucs_status_t status = ucp_perf_cuda_send_nbx<level, cmd>(params, idx, req);
-    if (status != UCS_OK) {
+    if (UCS_STATUS_IS_ERR(status)) {
         return status;
     }
 
@@ -262,7 +262,7 @@ ucp_perf_cuda_put_multi_bw_kernel(ucx_perf_cuda_context &ctx,
 
         ucp_device_request_t &req = request_mgr.get_request();
         status = ucp_perf_cuda_send_nbx<level, cmd>(params, idx, req);
-        if (status != UCS_OK) {
+        if (UCS_STATUS_IS_ERR(status)) {
             ucs_device_error("send failed: %d", status);
             goto out;
         }

src/ucp/api/device/ucp_device_impl.h

@@ -24,6 +24,7 @@
  */
 typedef struct ucp_device_request {
     uct_device_completion_t comp;
+    ucs_status_t            status;
     uct_device_ep_h         device_ep;
 } ucp_device_request_t;
 
@@ -51,9 +52,6 @@ UCS_F_DEVICE void ucp_device_request_init(uct_device_ep_t *device_ep,
     if (req != nullptr) {
         comp           = &req->comp;
         req->device_ep = device_ep;
-        uct_device_completion_init(comp);
-        /* TODO: Handle multiple device posts with same req? */
-        ++comp->count;
     } else {
         comp = nullptr;
     }
@@ -63,16 +61,20 @@ UCS_F_DEVICE void ucp_device_request_init(uct_device_ep_t *device_ep,
 /**
  * Macro for device put operations with retry logic
  */
-#define UCP_DEVICE_SEND_BLOCKING(_level, _uct_device_ep_send, _device_ep, ...) \
+#define UCP_DEVICE_SEND_BLOCKING(_level, _uct_device_ep_send, _device_ep, \
+                                 _req, ...) \
     ({ \
         ucs_status_t _status; \
         do { \
             _status = _uct_device_ep_send<_level>(_device_ep, __VA_ARGS__); \
             if (_status != UCS_ERR_NO_RESOURCE) { \
                 break; \
             } \
-            _status = uct_device_ep_progress<_level>(_device_ep); \
-        } while (!UCS_STATUS_IS_ERR(_status)); \
+            uct_device_ep_progress<_level>(_device_ep); \
+        } while (1); \
+        if (_req != nullptr) { \
+            _req->status = _status; \
+        } \
         _status; \
     })
 
@@ -148,8 +150,8 @@ UCS_F_DEVICE ucs_status_t ucp_device_put_single(
     }
 
     return UCP_DEVICE_SEND_BLOCKING(level, uct_device_ep_put_single, device_ep,
-                                    uct_elem, address, remote_address, length,
-                                    flags, comp);
+                                    req, uct_elem, address, remote_address,
+                                    length, flags, comp);
 }
 
 
@@ -199,8 +201,8 @@ UCS_F_DEVICE ucs_status_t ucp_device_counter_inc(
     }
 
     return UCP_DEVICE_SEND_BLOCKING(level, uct_device_ep_atomic_add, device_ep,
-                                    uct_elem, inc_value, remote_address, flags,
-                                    comp);
+                                    req, uct_elem, inc_value, remote_address,
+                                    flags, comp);
 }
 
 
@@ -263,8 +265,9 @@ UCS_F_DEVICE ucs_status_t ucp_device_put_multi(
     }
 
     return UCP_DEVICE_SEND_BLOCKING(level, uct_device_ep_put_multi, device_ep,
-                                    uct_mem_list, mem_list_h->mem_list_length,
-                                    addresses, remote_addresses, lengths,
+                                    req, uct_mem_list,
+                                    mem_list_h->mem_list_length, addresses,
+                                    remote_addresses, lengths,
                                     counter_inc_value, counter_remote_address,
                                     flags, comp);
 }
@@ -338,10 +341,11 @@ UCS_F_DEVICE ucs_status_t ucp_device_put_multi_partial(
     }
 
     return UCP_DEVICE_SEND_BLOCKING(level, uct_device_ep_put_multi_partial,
-                                    device_ep, uct_mem_list, mem_list_indices,
-                                    mem_list_count, addresses, remote_addresses,
-                                    lengths, counter_index, counter_inc_value,
-                                    counter_remote_address, flags, comp);
+                                    device_ep, req, uct_mem_list,
+                                    mem_list_indices, mem_list_count, addresses,
+                                    remote_addresses, lengths, counter_index,
+                                    counter_inc_value, counter_remote_address,
+                                    flags, comp);
 }
 
 
@@ -409,19 +413,14 @@ UCS_F_DEVICE void ucp_device_counter_write(void *counter_ptr, uint64_t value)
 template<ucs_device_level_t level = UCS_DEVICE_LEVEL_THREAD>
 UCS_F_DEVICE ucs_status_t ucp_device_progress_req(ucp_device_request_t *req)
 {
-    ucs_status_t status;
-
-    if (ucs_likely(req->comp.count == 0)) {
-        return req->comp.status;
-    }
-
-    status = uct_device_ep_progress<level>(req->device_ep);
-    if (status != UCS_OK) {
-        return status;
+    if (ucs_likely(req->status != UCS_INPROGRESS)) {
+        return req->status;
     }
 
-    return (ucs_likely(req->comp.count == 0)) ? req->comp.status :
-                                                UCS_INPROGRESS;
+    uct_device_ep_progress<level>(req->device_ep);
+    req->status = uct_device_ep_check_completion<level>(req->device_ep,
+                                                        &req->comp);
+    return req->status;
 }
 
 #endif /* UCP_DEVICE_IMPL_H */

src/uct/api/device/uct_device_impl.h

@@ -15,6 +15,11 @@
 
 #include <uct/ib/mlx5/gdaki/gdaki.cuh>
 
+union uct_device_completion {
+    uct_rc_gda_completion_t   rc_gda;
+    uct_cuda_ipc_completion_t cuda_ipc;
+};
+
 
 /**
  * @ingroup UCT_DEVICE
@@ -242,34 +247,37 @@ UCS_F_DEVICE ucs_status_t uct_device_ep_put_multi_partial(
  * @brief Progress all operations on device endpoint @a device_ep.
  *
  * @param [in]  device_ep       Device endpoint to be used for the operation.
- *
- * @return UCS_OK           - Some operation was completed.
- * @return UCS_INPROGRESS   - No progress on the endpoint.
- * @return Error code as defined by @ref ucs_status_t
  */
 template<ucs_device_level_t level>
-UCS_F_DEVICE ucs_status_t uct_device_ep_progress(uct_device_ep_h device_ep)
+UCS_F_DEVICE void uct_device_ep_progress(uct_device_ep_h device_ep)
 {
     if (device_ep->uct_tl_id == UCT_DEVICE_TL_RC_MLX5_GDA) {
-        return uct_rc_mlx5_gda_ep_progress<level>(device_ep);
-    } else if (device_ep->uct_tl_id == UCT_DEVICE_TL_CUDA_IPC) {
-        return UCS_OK;
+        uct_rc_mlx5_gda_ep_progress<level>(device_ep);
     }
-
-    return UCS_ERR_UNSUPPORTED;
 }
 
 
 /**
  * @ingroup UCT_DEVICE
- * @brief Initialize a device completion object.
+ * @brief Check whether opetation executed on device endpoint @a device_ep was
+ * completed.
+ *
+ * @param [in]  device_ep       Device endpoint to be used for the operation.
+ * @param [in]  comp            Completion object tracking operation progress.
  *
- * @param [out] comp  Device completion object to initialize.
+ * @return UCS_OK           - Some operation was completed.
+ * @return UCS_INPROGRESS   - No progress on the endpoint.
+ * @return Error code as defined by @ref ucs_status_t
  */
-UCS_F_DEVICE void uct_device_completion_init(uct_device_completion_t *comp)
+template<ucs_device_level_t level>
+UCS_F_DEVICE ucs_status_t uct_device_ep_check_completion(
+        uct_device_ep_h device_ep, uct_device_completion_t *comp)
 {
-    comp->count  = 0;
-    comp->status = UCS_OK;
+    if (device_ep->uct_tl_id == UCT_DEVICE_TL_RC_MLX5_GDA) {
+        return uct_rc_mlx5_gda_ep_check_completion<level>(device_ep, comp);
+    }
+
+    return UCS_ERR_UNSUPPORTED;
 }
 
 #endif

src/uct/api/device/uct_device_types.h

@@ -42,10 +42,7 @@ typedef struct uct_device_ep {
 
 
 /* Completion object for device operations */
-typedef struct uct_device_completion {
-    uint32_t     count;  /* How many operations are pending */
-    ucs_status_t status; /* Status of the operation */
-} uct_device_completion_t;
+typedef union uct_device_completion uct_device_completion_t;
 
 
 /* Base structure for all device memory elements */

src/uct/cuda/cuda_ipc/cuda_ipc.cuh

@@ -304,8 +304,6 @@ uct_cuda_ipc_ep_put_single(uct_device_ep_h device_ep,
     mapped_rem_addr = uct_cuda_ipc_map_remote(cuda_ipc_mem_element, remote_address);
     uct_cuda_ipc_copy_level<level>(mapped_rem_addr, address, length);
     uct_cuda_ipc_level_sync<level>();
-    --comp->count;
-
     return UCS_OK;
 }
 
@@ -339,10 +337,6 @@ uct_cuda_ipc_ep_put_multi(uct_device_ep_h device_ep,
     }
 
     uct_cuda_ipc_level_sync<level>();
-    if (lane_id == 0) {
-        --comp->count;
-    }
-
     return UCS_OK;
 }
 
@@ -376,10 +370,6 @@ uct_cuda_ipc_ep_put_multi_partial(uct_device_ep_h device_ep,
     }
 
     uct_cuda_ipc_level_sync<level>();
-    if (lane_id == 0) {
-        --comp->count;
-    }
-
     return UCS_OK;
 }
 
@@ -403,10 +393,6 @@ uct_cuda_ipc_ep_atomic_add(uct_device_ep_h device_ep,
     }
 
     uct_cuda_ipc_level_sync<level>();
-    if (lane_id == 0) {
-        --comp->count;
-    }
-
     return UCS_OK;
 }
 

src/uct/cuda/cuda_ipc/cuda_ipc_device.h

@@ -12,4 +12,8 @@ typedef struct {
     ptrdiff_t mapped_offset;
 } uct_cuda_ipc_device_mem_element_t;
 
+
+typedef struct {
+} uct_cuda_ipc_completion_t;
+
 #endif

src/uct/ib/mlx5/gdaki/gdaki.c

@@ -92,10 +92,8 @@ static UCS_CLASS_INIT_FUNC(uct_rc_gdaki_ep_t, const uct_ep_params_t *params)
     uct_ib_mlx5_wq_calc_sizes(&qp_attr);
 
     cq_attr.flags      |= UCT_IB_MLX5_CQ_IGNORE_OVERRUN;
-    cq_attr.umem_offset = ucs_align_up_pow2(
-            sizeof(uct_rc_gdaki_dev_ep_t) +
-                    qp_attr.max_tx * sizeof(uct_rc_gdaki_op_t),
-            ucs_get_page_size());
+    cq_attr.umem_offset = ucs_align_up_pow2(sizeof(uct_rc_gdaki_dev_ep_t),
+                                            ucs_get_page_size());
 
     qp_attr.mmio_mode     = UCT_IB_MLX5_MMIO_MODE_DB;
     qp_attr.super.srq_num = 0;
@@ -109,9 +107,9 @@ static UCS_CLASS_INIT_FUNC(uct_rc_gdaki_ep_t, const uct_ep_params_t *params)
     dev_ep_size = qp_attr.umem_offset + qp_attr.len;
     /*
      * dev_ep layout:
-     * +---------------------+-------+---------+---------+
-     * | counters, dbr       | ops   | cq buff | wq buff |
-     * +---------------------+-------+---------+---------+
+     * +---------------------+---------+---------+
+     * | counters, dbr       | cq buff | wq buff |
+     * +---------------------+---------+---------+
      */
     status      = uct_rc_gdaki_alloc(dev_ep_size, ucs_get_page_size(),
                                      (void**)&self->ep_gpu, &self->ep_raw);