Searched refs:bytes_accessed (Results 1 – 10 of 10) sorted by relevance
/external/tensorflow/tensorflow/compiler/xla/service/ |
D | hlo_cost_analysis_test.cc | 157 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 186 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 215 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 238 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 256 EXPECT_EQ(analysis.bytes_accessed(), 80); in TEST_F() 284 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 313 EXPECT_EQ(analysis.bytes_accessed(), in TEST_F() 489 constexpr int64 bytes_accessed = sizeof(float) * 4 * 2 * 2; in TEST_F() local 490 static_assert(bytes_accessed == 64, ""); in TEST_F() 491 EXPECT_EQ(fusion_analysis.bytes_accessed(), bytes_accessed); 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 | 52 float bytes_accessed = GetShapeSize(hlo->shape()); in Preprocess() local 54 bytes_accessed += GetShapeSize(operand->shape()); in Preprocess() 56 current_properties_[kBytesAccessedKey] = bytes_accessed; in Preprocess() 551 float bytes_accessed = GetShapeSize(hlo->operand(0)->shape()) / 2.0f; in HandleTriangularSolve() local 552 bytes_accessed += GetShapeSize(hlo->operand(1)->shape()); in HandleTriangularSolve() 553 current_properties_[kBytesAccessedKey] = bytes_accessed; in HandleTriangularSolve() 565 float bytes_accessed = GetShapeSize(hlo->operand(0)->shape()) / 2.0f; in HandleCholesky() local 566 current_properties_[kBytesAccessedKey] = bytes_accessed; in HandleCholesky() 753 float HloCostAnalysis::bytes_accessed() const { in bytes_accessed() function in xla::HloCostAnalysis 769 int64 HloCostAnalysis::bytes_accessed(const HloInstruction& hlo) const { in bytes_accessed() function in xla::HloCostAnalysis
|
D | hlo_cost_analysis.h | 135 float bytes_accessed() const; 142 int64 bytes_accessed(const HloInstruction& hlo) const;
|
D | hlo_profile_printer.cc | 62 instruction_info.bytes_accessed(), in PrintHloProfile()
|
D | hlo_execution_profile.cc | 105 instruction_info->set_bytes_accessed(cost_analysis.bytes_accessed(*hlo)); in CreateHloProfilePrinterData()
|
D | hlo_profile_printer_data.proto | 34 float bytes_accessed = 6; field
|
/external/tensorflow/tensorflow/compiler/xla/service/cpu/ |
D | parallel_task_assignment.cc | 67 const int64 bytes_accessed = in GetParallelTaskCount() local 68 std::max(int64{1}, cost_analysis_->bytes_accessed(*instruction)); in GetParallelTaskCount() 71 static_cast<float>(bytes_accessed); in GetParallelTaskCount() 92 10 * cost_analysis_->bytes_accessed(*instruction); in GetParallelTaskCount()
|
/external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
D | nvptx_compiler.cc | 784 cost_analysis.bytes_accessed()); in RunBackend()
|