- Sort Score
- Result 10 results
- Languages All
Results 1 - 10 of 12 for pipelining (0.14 sec)
-
tensorflow/compiler/mlir/tensorflow/transforms/sparsecore/sparsecore_passes.h
// For architectures that support accelerated embedding lookups, this pass will // rewrite the graph to use pipelining for better device utilization. std::unique_ptr<OperationPass<ModuleOp>> CreateEmbeddingSequencingPass(); // This is a strictly sequential and formally correct fallback option for the // embedding pipelining pass intended for debugging during pipelining // development. std::unique_ptr<OperationPass<ModuleOp>> CreateEmbeddingPipeliningPass();
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Mar 28 23:42:09 UTC 2024 - 2.1K bytes - Viewed (0) -
tensorflow/compiler/mlir/tensorflow/transforms/sparsecore/sparsecore_passes.td
include "mlir/Pass/PassBase.td" def EmbeddingPipeliningPass : Pass<"tf-embedding-pipelining", "mlir::ModuleOp"> { let summary = "Rewrite graph for embedding pipelining"; let constructor = "TFDevice::CreateEmbeddingPipeliningPass()"; let description = [{ For architectures that support accelerated embedding lookups, this pass will rewrite the graph to use pipelining for better device utilization. }]; }
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Mar 28 23:42:09 UTC 2024 - 3.9K bytes - Viewed (0) -
tensorflow/compiler/mlir/tensorflow/tests/embedding_pipelining.mlir
// RUN: tf-opt %s -split-input-file -verify-diagnostics -tf-embedding-pipelining | FILECHECK_OPTS="" FileCheck %s // This test verifies the handling of TPU replicated inputs and outputs as well as the extraction of the four main functions. module { func.func @main() { %cst_main = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Mon Oct 30 06:52:55 UTC 2023 - 33.1K bytes - Viewed (0) -
tensorflow/compiler/mlir/tensorflow/transforms/sparsecore/embedding_pipelining.cc
// Enable automated pipelining pass unless: // 1. The user disables it via flag, or // 2. The graph contains TF.Summary ops. Graphs like this typically only run // for a single step which doesn't work in pipelining. if (tensorflow::GetBuildXlaOpsPassFlags() ->tf_xla_disable_full_embedding_pipelining) { LOG(INFO) << "Embedding pipelining disabled via flag."; return false; }
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Apr 25 16:01:03 UTC 2024 - 92.9K bytes - Viewed (0) -
tensorflow/compiler/mlir/tf2xla/internal/clustering_bridge_passes.cc
// Only one of EmbeddingSequencing and EmbeddingPipelining will actually // run and the logic is in EmbeddingPipeliningPass. If the pipelining pass // runs, embedding attributes are stripped and the sequencing pass will have // no effect. If the pipelining pass doesn't run, embedding attributes are // preserved and the sequencing rewrite will trigger. pm.addPass(mlir::TFDevice::CreateEmbeddingPipeliningPass());
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Tue Apr 30 16:09:14 UTC 2024 - 11.2K bytes - Viewed (0) -
tensorflow/compiler/jit/cluster_scoping_pass.cc
/*edge_filter=*/nullptr); } // This preserves the parallelism between pipeline stages. For example, below // is a typical pattern of input pipelining in Tensorflow and this heuristic // ensures Node_X and Node_Y are put into different clusters. Without the // heuristic, they may be put into the same cluster and it can introduce
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Feb 22 08:47:20 UTC 2024 - 5.7K bytes - Viewed (0) -
tensorflow/compiler/mlir/tensorflow/transforms/sparsecore/embedding_sequencing.cc
// separate functions for proper sequencing of TF2 TPU Embedding (see // tpu_embedding_v3.py). This pass is a precursor for pipelining (see // embedding_pipelining.cc) and DOES NOT permit parallel execution across SC and // TC. This pass is a temporary fallback to use while developing full pipelining // capabilities. // // Ops are broken up into: // 1. SC forward pass // 2. TC forward/backward pass // 3. SC backward pass
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Apr 25 16:01:03 UTC 2024 - 39.4K bytes - Viewed (0) -
tensorflow/compiler/jit/flags.h
// guarantee that tests are run on XLA and not on TF's CPU implementation. bool tf_xla_disable_constant_folding; // Disables full embedding pipelining when true. Instead, strict SparseCore // TensorCore sequencing will be used. bool tf_xla_disable_full_embedding_pipelining; // Force the WhileOps in embedding_pipelining and embedding_sequencing to use
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Wed Apr 17 18:52:57 UTC 2024 - 14.5K bytes - Viewed (0) -
tensorflow/compiler/jit/flags.cc
"compilation."), Flag("tf_xla_disable_full_embedding_pipelining", &build_ops_flags->tf_xla_disable_full_embedding_pipelining, "If true then disables full embedding pipelining and instead use " "strict SparseCore / TensorCore sequencing."), Flag("tf_xla_embedding_parallel_iterations", &build_ops_flags->tf_xla_embedding_parallel_iterations,
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Wed Apr 17 18:52:57 UTC 2024 - 24.5K bytes - Viewed (0) -
tensorflow/compiler/mlir/tfrt/tests/mlrt/async_while.mlir
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Mon Oct 30 06:52:55 UTC 2023 - 22.2K bytes - Viewed (0)