/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 …]
|
/external/pffft/ |
D | bench_pffft.c | 454 double show_output(const char *name, int N, int cplx, float flops, float t0, float t1, int max_iter… 456 float mflops = flops/1e6/(t1 - t0 + 1e-16); 458 if (flops != -1) 463 if (flops != -1) { 521 double te, t0, t1, tstop, flops, Tfastest; 582 …flops = (max_iter*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); /* see http://www.fftw.org/speed… 584 tmeas[TYPE_MFLOPS][ALGO_FFTPACK] = flops/1e6/(t1 - t0 + 1e-16); 586 …tmeas[TYPE_DUR_NS][ALGO_FFTPACK] = show_output("FFTPack", N, cplx, flops, t0, t1, max_iter, tableF… 627 …flops = (max_iter*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); /* see http://www.fftw.org/speed… 629 tmeas[TYPE_MFLOPS][ALGO_VECLIB] = flops/1e6/(t1 - t0 + 1e-16); [all …]
|
/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 | 51 auto flops = in SpatialConvolution() local 53 state.SetItemsProcessed(flops * state.iterations()); in SpatialConvolution() 78 auto flops = in SpatialConvolutionBackwardInput() local 80 state.SetItemsProcessed(flops * state.iterations()); in SpatialConvolutionBackwardInput() 105 auto flops = in SpatialConvolutionBackwardKernel() local 107 state.SetItemsProcessed(flops * state.iterations()); in SpatialConvolutionBackwardKernel() 269 auto flops = num_computed_elements * in CuboidConvolution() local 271 state.SetItemsProcessed(flops * state.iterations()); in CuboidConvolution() 297 auto flops = num_computed_elements * in CuboidConvolutionBackwardInput() local 299 state.SetItemsProcessed(flops * state.iterations()); in CuboidConvolutionBackwardInput() [all …]
|
/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 85 op_metrics->set_flops(op_metrics->flops() + flops * occurrences); in EnterOp()
|
D | op_metrics_db_utils.cc | 59 tf_op_metrics->set_flops(tf_op_metrics->flops() + in UpdateTfOpMetricsWithDeviceOpMetrics() 60 device_op_metrics.flops()); in UpdateTfOpMetricsWithDeviceOpMetrics()
|
D | cost_utils.h | 43 uint64 flops = 0LL; member
|
D | op_utils.h | 73 uint64 children_time_ps, int64 flops, int64 bytes_accessed,
|
/external/tensorflow/tensorflow/core/profiler/convert/ |
D | op_metrics_to_record.h | 89 SafeDivide(metrics.flops(), PicosToNanos(metrics.time_ps()))); in SetRooflineMetrics() 93 SafeDivide(metrics.flops(), metrics.bytes_accessed())); in SetRooflineMetrics() 99 : ((metrics.flops() != 0) ? "Compute" : "Unknown")); in SetRooflineMetrics()
|
D | op_metrics_db_combiner.cc | 70 dst->set_flops(src.flops() + dst->flops()); in CombineOpMetrics()
|
D | xplane_to_op_metrics_db.cc | 248 /*children_time_ps=*/0, costs.flops, costs.bytes_accessed); in ConvertDeviceTraceXPlaneToOpMetricsDb()
|
/external/tensorflow/tensorflow/core/profiler/protobuf/ |
D | op_profile.proto | 65 // - utilization is flops/time 66 // - wasted potential flops is proportional to time - flops 68 double flops = 2; field
|
D | op_metrics.proto | 52 uint64 flops = 2; field
|
/external/llvm-project/mlir/lib/ExecutionEngine/ |
D | RunnerUtils.cpp | 89 extern "C" void print_flops(double flops) { in print_flops() argument 90 fprintf(stderr, "%lf GFLOPS\n", flops / 1.0E9); in print_flops()
|
/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 | 157 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value 159 self.assertEqual(7200, flops) 168 flops = ops.get_stats_for_node_def(g, op.node_def, "flops").value 170 self.assertEqual(7200, flops)
|
/external/tensorflow/tensorflow/compiler/xla/ |
D | util.cc | 166 string HumanReadableNumOps(double flops, double nanoseconds, in HumanReadableNumOps() argument 171 double nano_flops = flops / nanoseconds; in HumanReadableNumOps() 185 string HumanReadableNumFlops(double flops, double nanoseconds) { in HumanReadableNumFlops() argument 186 return HumanReadableNumOps(flops, nanoseconds, "FL"); in HumanReadableNumFlops()
|
/external/virglrenderer/ci/previous_results/gl_host_softpipe/deqp_gl31/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/virglrenderer/ci/previous_results/es_host_softpipe/deqp_gl30/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/virglrenderer/ci/previous_results/gl_host_softpipe/deqp_gl30/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/virglrenderer/ci/previous_results/es_host_softpipe/deqp_gl31/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/virglrenderer/ci/previous_results/gl_host_softpipe/deqp_gl32/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/virglrenderer/ci/previous_results/es_host_softpipe/deqp_gl32/ |
D | ignore_tests.txt | 1 # Flip-flops between Pass and Fail
|
/external/llvm-project/mlir/test/mlir-cpu-runner/ |
D | sgemm_naive_codegen.mlir | 42 %flops = divf %num_flops_f, %t : f64 43 call @print_flops(%flops) : (f64) -> ()
|