| /external/tensorflow/tensorflow/core/profiler/g3doc/ |
| D | profile_model_architecture.md | 48 * It must have `RegisterStatistics('flops')` defined in TensorFlow. tfprof
51 * It must have known "shape" information for RegisterStatistics('flops')
70 _TFProfRoot (--/17.63b flops)
71 gradients/pool_logit/xw_plus_b/MatMul_grad/MatMul (163.84k/163.84k flops)
72 gradients/pool_logit/xw_plus_b/MatMul_grad/MatMul_1 (163.84k/163.84k flops)
73 init/init_conv/Conv2D (113.25m/113.25m flops)
74 pool_logit/xw_plus_b (1.28k/165.12k flops)
75 pool_logit/xw_plus_b/MatMul (163.84k/163.84k flops)
76 unit_1_0/sub1/conv1/Conv2D (603.98m/603.98m flops)
77 unit_1_0/sub2/conv2/Conv2D (603.98m/603.98m flops)
[all …]
|
| D | command_line.md | 299 # Get run-time shape information in order to fill shapes and get flops.
|
| /external/eigen/bench/perf_monitoring/gemm/ |
| D | gemm.cpp | 43 double flops = 2. * m * n * k; in bench() local
44 long rep = std::max(1., std::min(100., up/flops) ); in bench()
45 long tries = std::max(tm0, std::min(tm1, up/flops) ); in bench()
49 return 1e-9 * rep * flops / t.best(); in bench()
|
| D | lazy_gemm.cpp | 44 double flops = 2. * m * n * k; in bench() local
45 long rep = std::max(10., std::min(10000., up/flops) ); in bench()
46 long tries = std::max(tm0, std::min(tm1, up/flops) ); in bench()
50 return 1e-9 * rep * flops / t.best(); in bench()
|
| /external/tensorflow/tensorflow/contrib/tpu/profiler/ |
| D | op_profile.proto | 58 // - utilization is flops/time
59 // - wasted potential flops is proportional to time - flops
61 double flops = 2; field
|
| /external/tensorflow/tensorflow/compiler/xla/tests/ |
| D | xla_hlo_profile_test.cc | 45 string flops; member
54 if (RE2::FullMatch(parsed_line.flops, "[0-9.TGMk]+FLOP/s")) { in HasFlops()
57 << parsed_line.flops << "'"; in HasFlops()
62 << parsed_line.flops << "'"; in HasFlops()
105 &parsed_line.usec, &parsed_line.flops, &parsed_line.trops, in ParseOneProfileOutputLine()
|
| /external/tensorflow/tensorflow/python/kernel_tests/ |
| D | matmul_op_test.py | 143 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value
145 self.assertEqual(7200, flops)
154 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value
156 self.assertEqual(7200, flops)
|
| /external/tensorflow/tensorflow/compiler/xla/ |
| D | util.cc | 237 string HumanReadableNumOps(double flops, double nanoseconds, in HumanReadableNumOps() argument
242 double nano_flops = flops / nanoseconds; in HumanReadableNumOps()
256 string HumanReadableNumFlops(double flops, double nanoseconds) { in HumanReadableNumFlops() argument
257 return HumanReadableNumOps(flops, nanoseconds, "FL"); in HumanReadableNumFlops()
|
| D | util.h | 380 string HumanReadableNumFlops(double flops, double nanoseconds);
|
| /external/tensorflow/tensorflow/compiler/xla/service/ |
| D | hlo_cost_analysis.cc | 420 double flops = 0.0; in HandleCrossReplicaSum() local
424 flops += ShapeUtil::ElementsIn(subshape); in HandleCrossReplicaSum()
427 current_properties_[kFlopsKey] = flops; in HandleCrossReplicaSum()
|
| /external/llvm/test/CodeGen/X86/ |
| D | break-false-dep.ll | 109 ; flops-6. Make sure the unused register read by vcvtsi2sdq is zeroed
|
| /external/eigen/unsupported/Eigen/ |
| D | MatrixFunctions | 455 cost is approximately \f$ 25 n^3 \f$ real flops for the real Schur
456 decomposition and \f$ 3\frac13 n^3 \f$ real flops for the remainder
|
| /external/tensorflow/tensorflow/python/framework/ |
| D | ops_test.py | 2368 flops = ops.get_stats_for_node_def(graph, node, "flops")
2369 self.assertEqual(20, flops.value)
|
| /external/llvm/lib/Target/X86/ |
| D | README.txt | 1011 From flops:
|
| /external/swiftshader/third_party/LLVM/lib/Target/X86/ |
| D | README.txt | 1174 From flops:
|
| /external/jline/src/src/test/resources/jline/example/ |
| D | english.gz | |