| /external/tensorflow/tensorflow/core/profiler/g3doc/ |
| D | profile_model_architecture.md | 48 * It must have `RegisterStatistics('flops')` defined in TensorFlow. tfprof
52 * It must have known "shape" information for RegisterStatistics('flops') to
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/core/kernels/ |
| D | eigen_benchmark_cpu_test.cc | 53 auto flops = in SpatialConvolution() local
55 ::tensorflow::testing::ItemsProcessed(flops * iters); in SpatialConvolution()
81 auto flops = in SpatialConvolutionBackwardInput() local
83 ::tensorflow::testing::ItemsProcessed(flops * iters); in SpatialConvolutionBackwardInput()
109 auto flops = in SpatialConvolutionBackwardKernel() local
111 ::tensorflow::testing::ItemsProcessed(flops * iters); in SpatialConvolutionBackwardKernel()
275 auto flops = num_computed_elements * in CuboidConvolution() local
277 ::tensorflow::testing::ItemsProcessed(flops * iters); in CuboidConvolution()
304 auto flops = num_computed_elements * in CuboidConvolutionBackwardInput() local
306 ::tensorflow::testing::ItemsProcessed(flops * iters); in CuboidConvolutionBackwardInput()
[all …]
|
| /external/tensorflow/tensorflow/core/profiler/ |
| D | op_profile.proto | 64 // - utilization is flops/time
65 // - wasted potential flops is proportional to time - flops
67 double flops = 2; field
|
| /external/tensorflow/tensorflow/compiler/xla/ |
| D | util.cc | 151 string HumanReadableNumOps(double flops, double nanoseconds, in HumanReadableNumOps() argument
156 double nano_flops = flops / nanoseconds; in HumanReadableNumOps()
170 string HumanReadableNumFlops(double flops, double nanoseconds) { in HumanReadableNumFlops() argument
171 return HumanReadableNumOps(flops, nanoseconds, "FL"); in HumanReadableNumFlops()
|
| D | util.h | 437 string HumanReadableNumFlops(double flops, double nanoseconds);
|
| /external/tensorflow/tensorflow/compiler/xla/tests/ |
| D | xla_hlo_profile_test.cc | 49 string flops; member
58 if (RE2::FullMatch(parsed_line.flops, "[0-9.TGMk]+FLOP/s")) { in HasFlops()
61 << parsed_line.flops << "'"; in HasFlops()
66 << parsed_line.flops << "'"; in HasFlops()
110 &parsed_line.usec, &parsed_line.flops, &parsed_line.trops,
|
| /external/tensorflow/tensorflow/python/kernel_tests/ |
| D | matmul_op_test.py | 155 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value
157 self.assertEqual(7200, flops)
166 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value
168 self.assertEqual(7200, flops)
|
| /external/tensorflow/tensorflow/compiler/xla/service/ |
| D | hlo_cost_analysis.cc | 582 double flops = 0.0; in HandleAllReduce() local
586 flops += ShapeUtil::ElementsIn(subshape); in HandleAllReduce()
589 current_properties_[kFlopsKey] = flops; in HandleAllReduce()
|
| /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/swiftshader/third_party/llvm-7.0/llvm/test/CodeGen/X86/ |
| D | break-false-dep.ll | 110 ; flops-6. Make sure the unused register read by vcvtsi2sdq is zeroed
|
| /external/tensorflow/tensorflow/python/framework/ |
| D | ops_test.py | 2669 flops = ops.get_stats_for_node_def(graph, node, "flops")
2670 self.assertEqual(20, flops.value)
|
| /external/skqp/src/compute/skc/platforms/cl_12/kernels/ |
| D | render.cl | 817 // evaluation might save some flops.
|
| /external/llvm/lib/Target/X86/ |
| D | README.txt | 1011 From flops:
|
| /external/swiftshader/third_party/llvm-7.0/llvm/lib/Target/X86/ |
| D | README.txt | 1011 From flops:
|
| /external/skia/src/compute/skc/platforms/cl_12/kernels/ |
| D | render.cl | 817 // evaluation might save some 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 | |
| /external/cldr/tools/java/org/unicode/cldr/util/data/transforms/ |
| D | internal_raw_IPA.txt | 60974 flops %32310 flɑps
|
| D | internal_raw_IPA-old.txt | 72240 flops %21123 flɑps
|