| /external/tensorflow/tensorflow/compiler/xla/service/ |
| D | hlo_cost_analysis_test.cc | 156 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
190 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
226 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
256 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
280 EXPECT_EQ(analysis.bytes_accessed(), 80); in TEST_F()
312 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
392 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F()
420 EXPECT_EQ(analysis.bytes_accessed(), sizeof(float) * (10 * 20 + 1 + 10)); in TEST_F()
444 EXPECT_EQ(analysis.bytes_accessed(), sizeof(float) * (10 * 20 + 1 + 2 * 4)); in TEST_F()
479 EXPECT_EQ(analysis.bytes_accessed(), sizeof(float) * (10 * 20 * 2 + 2 * 3)); in TEST_F()
[all …]
|
| D | human_readable_profile_builder.cc | 51 if (op.cycles > 0 && op.bytes_accessed >= 0) { in ToString()
53 HumanReadableNumBytes(op.bytes_accessed / CyclesToSeconds(op.cycles)), in ToString()
55 double bpc = static_cast<double>(op.bytes_accessed) / op.cycles; in ToString()
56 if (op.bytes_accessed > op.cycles) { in ToString()
119 total_bytes += std::max(op.bytes_accessed, int64{0}); in ToString()
211 entry.metric = static_cast<double>(op.bytes_accessed) / (1 << 20); in ToString()
|
| D | human_readable_profile_builder.h | 50 int64 transcendental_count, int64 bytes_accessed, in AddOp() argument
54 bytes_accessed, optimal_seconds}); in AddOp()
68 int64 bytes_accessed; // -1 if unknown member
|
| D | hlo_cost_analysis.cc | 57 float bytes_accessed = GetShapeSize(hlo->shape()); in Preprocess() local
61 bytes_accessed += GetShapeSize(operand->shape()); in Preprocess()
64 current_properties_[kBytesAccessedKey] = bytes_accessed; in Preprocess()
711 float bytes_accessed = GetShapeSize(hlo->shape()); in HandleTriangularSolve() local
713 bytes_accessed += GetShapeSize(hlo->operand(0)->shape()) / 2.0f; in HandleTriangularSolve()
715 bytes_accessed += GetShapeSize(hlo->operand(1)->shape()); in HandleTriangularSolve()
717 current_properties_[kBytesAccessedKey] = bytes_accessed; in HandleTriangularSolve()
730 float bytes_accessed = GetShapeSize(hlo->operand(0)->shape()) / 2.0f; in HandleCholesky() local
732 bytes_accessed += GetShapeSize(hlo->operand(0)->shape()) / 2.0f; in HandleCholesky()
734 current_properties_[kBytesAccessedKey] = bytes_accessed; in HandleCholesky()
[all …]
|
| D | hlo_profile_printer.cc | 62 instruction_info.bytes_accessed(), in PrintHloProfile()
|
| D | hlo_cost_analysis.h | 150 float bytes_accessed() const;
161 int64 bytes_accessed(const HloInstruction& hlo) const;
|
| D | hlo_profile_printer_data.proto | 34 float bytes_accessed = 6; field
|
| D | hlo_execution_profile.cc | 106 instruction_info->set_bytes_accessed(cost_analysis.bytes_accessed(*hlo)); in CreateHloProfilePrinterData()
|
| D | memory_space_assignment.cc | 261 float bytes_accessed = cost_analysis_.bytes_accessed(instruction); in GetInstructionElapsedDueToMemory() local
263 bytes_accessed / in GetInstructionElapsedDueToMemory()
272 bytes_accessed -= operand_bytes_accessed; in GetInstructionElapsedDueToMemory()
275 bytes_accessed / in GetInstructionElapsedDueToMemory()
285 bytes_accessed -= output_bytes_accessed; in GetInstructionElapsedDueToMemory()
288 bytes_accessed / in GetInstructionElapsedDueToMemory()
|
| /external/tensorflow/tensorflow/core/profiler/convert/ |
| D | op_metrics_db_combiner.cc | 71 dst->set_bytes_accessed(src.bytes_accessed() + dst->bytes_accessed()); in CombineOpMetrics()
100 src_memory_accessed.bytes_accessed() + in CombineMemoryAccessedBreakdown()
101 dst_memory_accessed->bytes_accessed()); in CombineMemoryAccessedBreakdown()
|
| D | op_metrics_to_record.h | 91 SafeDivide(metrics.bytes_accessed(), PicosToNanos(metrics.time_ps()))); in SetRooflineMetrics()
93 SafeDivide(metrics.flops(), metrics.bytes_accessed())); in SetRooflineMetrics()
94 record->set_bound_by((metrics.bytes_accessed() != 0) in SetRooflineMetrics()
|
| D | xplane_to_op_metrics_db.cc | 248 /*children_time_ps=*/0, costs.flops, costs.bytes_accessed); in ConvertDeviceTraceXPlaneToOpMetricsDb()
|
| /external/tensorflow/tensorflow/core/profiler/utils/ |
| D | op_utils.cc | 70 uint64 time_ps, uint64 children_time_ps, int64 flops, int64 bytes_accessed, in EnterOp() argument
86 op_metrics->set_bytes_accessed(op_metrics->bytes_accessed() + in EnterOp()
87 bytes_accessed * occurrences); in EnterOp()
|
| D | op_metrics_db_utils.cc | 61 tf_op_metrics->set_bytes_accessed(tf_op_metrics->bytes_accessed() + in UpdateTfOpMetricsWithDeviceOpMetrics()
62 device_op_metrics.bytes_accessed()); in UpdateTfOpMetricsWithDeviceOpMetrics()
|
| D | cost_utils.h | 44 uint64 bytes_accessed = 0LL; member
|
| D | op_utils.h | 73 uint64 children_time_ps, int64 flops, int64 bytes_accessed,
|
| /external/tensorflow/tensorflow/compiler/xla/service/cpu/ |
| D | parallel_task_assignment.cc | 68 const int64 bytes_accessed = in GetParallelTaskCount() local
69 std::max(int64{1}, cost_analysis_->bytes_accessed(*instruction)); in GetParallelTaskCount()
72 static_cast<float>(bytes_accessed); in GetParallelTaskCount()
94 10 * cost_analysis_->bytes_accessed(*instruction); in GetParallelTaskCount()
|
| /external/tensorflow/tensorflow/core/profiler/protobuf/ |
| D | op_metrics.proto | 54 uint64 bytes_accessed = 5; field
65 uint64 bytes_accessed = 3; field
|
| /external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
| D | gpu_compiler.cc | 890 cost_analysis.bytes_accessed()); in RunBackend()
|