/aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/gpu/tests/ |
H A D | hlo_to_llvm_ir.cc | 67 tensorflow::se::CudaComputeCapability cuda_compute_capability; in CompileAndPrintLlvmIr() local 68 cuda_compute_capability.major = sm / 10; in CompileAndPrintLlvmIr() 69 cuda_compute_capability.minor = sm % 10; in CompileAndPrintLlvmIr() 83 cuda_compute_capability, rocm_compute_capability, in CompileAndPrintLlvmIr() 93 llvm_module.get(), cuda_compute_capability, in CompileAndPrintLlvmIr()
|
H A D | gpu_atomic_test.cc | 119 .cuda_compute_capability() in TEST_F()
|
H A D | mlir_gpu_test_base.cc | 62 stream_exec->GetDeviceDescription().cuda_compute_capability(), in CompileMlirModule()
|
/aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
H A D | nvptx_compiler.cc | 85 stream_exec->GetDeviceDescription().cuda_compute_capability()); in OptimizeHloConvolutionCanonicalization() 87 stream_exec->GetDeviceDescription().cuda_compute_capability()); in OptimizeHloConvolutionCanonicalization() 126 if (stream_exec->GetDeviceDescription().cuda_compute_capability().IsAtLeast( in OptimizeHloPostLayoutAssignment() 131 if (stream_exec->GetDeviceDescription().cuda_compute_capability().IsAtLeast( in OptimizeHloPostLayoutAssignment() 153 if (!stream_exec->GetDeviceDescription().cuda_compute_capability().IsAtLeast( in OptimizeHloPostLayoutAssignment() 324 return stream_exec->GetDeviceDescription().cuda_compute_capability(); in GetGpuVersion()
|
H A D | ir_emitter_context.h | 42 se::CudaComputeCapability cuda_compute_capability, in IrEmitterContext() argument 49 cuda_compute_capability_(cuda_compute_capability), in IrEmitterContext() 64 se::CudaComputeCapability cuda_compute_capability() const { in cuda_compute_capability() function
|
H A D | gpu_compiler.cc | 263 .cuda_compute_capability() in IsConvBF16Supported() 406 .cuda_compute_capability() in OptimizeHloModule() 755 stream_exec->GetDeviceDescription().cuda_compute_capability()); in OptimizeHloPostLayoutAssignment() 899 options.cuda_compute_capability = in LowerToJitRt() 900 stream_exec->GetDeviceDescription().cuda_compute_capability(); in LowerToJitRt() 970 se::CudaComputeCapability cuda_compute_capability, in CompileModuleToLlvmIrImpl() argument 1001 absl::StrCat("sm_", cuda_compute_capability.ToString(), in CompileModuleToLlvmIrImpl() 1037 gpu_device_info, cuda_compute_capability, rocm_compute_capability, in CompileModuleToLlvmIrImpl() 1361 stream_exec->GetDeviceDescription().cuda_compute_capability(), in RunBackend() 1457 stream_exec->GetDeviceDescription().cuda_compute_capability(), in CompileAheadOfTime() [all …]
|
H A D | tree_reduction_rewriter.h | 79 se::CudaComputeCapability cuda_compute_capability) in GpuTreeReductionRewriter() argument 80 : cuda_compute_capability_(cuda_compute_capability) {} in GpuTreeReductionRewriter()
|
H A D | tree_reduction_rewriter.cc | 49 se::CudaComputeCapability cuda_compute_capability) in ReductionRewriterVisitor() argument 50 : cuda_compute_capability_(cuda_compute_capability) {} in ReductionRewriterVisitor()
|
H A D | ir_emission_utils.h | 101 se::CudaComputeCapability cuda_compute_capability);
|
H A D | gpu_compiler.h | 178 se::CudaComputeCapability cuda_compute_capability,
|
/aosp_15_r20/external/tensorflow/tensorflow/tools/tensorflow_builder/config_detector/ |
H A D | BUILD | 20 ":cuda_compute_capability", 28 name = "cuda_compute_capability", 29 srcs = ["data/cuda_compute_capability.py"],
|
H A D | config_detector.py | 71 from tensorflow.tools.tensorflow_builder.config_detector.data import cuda_compute_capability 249 gpu_dict = cuda_compute_capability.retrieve_from_golden() 388 cuda_compute_capa = cuda_compute_capability.retrieve_from_web() 390 cuda_compute_capa = cuda_compute_capability.retrieve_from_golden()
|
/aosp_15_r20/external/tensorflow/tensorflow/compiler/mlir/tfrt/transforms/lmhlo_to_gpu/ |
H A D | lmhlo_to_gpu_binary.h | 45 options.cuda_compute_capability = {5, 2}; in DefaultGpuBinaryOptions() 53 stream_executor::CudaComputeCapability cuda_compute_capability; member
|
H A D | kernel_ops_pattern.cc | 235 gpu_options.cuda_compute_capability, gpu_options.rocm_compute_capability, in Emit() 319 gpu_options.cuda_compute_capability, in Match()
|
/aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/stream_executor/ |
H A D | device_description.cc | 92 result["CUDA Compute Capability"] = cuda_compute_capability().ToString(); in ToMap() 109 CudaComputeCapability DeviceDescription::cuda_compute_capability() const { in cuda_compute_capability() function in stream_executor::DeviceDescription
|
H A D | stream_executor_pimpl.cc | 283 GetDeviceDescription().cuda_compute_capability(), out_algorithms); in GetConvolveAlgorithms() 286 GetDeviceDescription().cuda_compute_capability(), out_algorithms); in GetConvolveAlgorithms() 289 GetDeviceDescription().cuda_compute_capability(), out_algorithms); in GetConvolveAlgorithms()
|
H A D | dnn.cc | 111 CudaComputeCapability cuda_compute_capability, in GetConvolveAlgorithms() argument 192 CudaComputeCapability cuda_compute_capability, in GetConvolveBackwardDataAlgorithms() argument 198 CudaComputeCapability cuda_compute_capability, in GetConvolveBackwardFilterAlgorithms() argument
|
H A D | device_description.h | 271 CudaComputeCapability cuda_compute_capability() const;
|
H A D | dnn.h | 1386 CudaComputeCapability cuda_compute_capability, 1497 CudaComputeCapability cuda_compute_capability, 1503 CudaComputeCapability cuda_compute_capability,
|
/aosp_15_r20/external/tensorflow/tensorflow/core/common_runtime/gpu/ |
H A D | gpu_device.cc | 1102 se->GetDeviceDescription().cuda_compute_capability(); in SingleVirtualDeviceMemoryLimit() 1263 (*details)["compute_capability"] = desc->cuda_compute_capability().ToString(); in GetDeviceDetails() 1550 ", compute capability: ", desc.cuda_compute_capability().ToString()); in GetShortDeviceDescription() 1905 << description->cuda_compute_capability().ToString() in GetValidDeviceIds() 1975 if (desc->cuda_compute_capability() < min_supported_capability) { in GetValidDeviceIds() 1980 << desc->cuda_compute_capability().ToString() in GetValidDeviceIds() 2020 auto device_capability = desc->cuda_compute_capability(); in GetValidDeviceIds()
|
/aosp_15_r20/external/tensorflow/tensorflow/core/grappler/ |
H A D | devices.cc | 53 (*desc)->cuda_compute_capability().IsAtLeast( in GetNumAvailableGPUs()
|
/aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/stream_executor/cuda/ |
H A D | cuda_dnn.h | 217 CudaComputeCapability cuda_compute_capability, 272 CudaComputeCapability cuda_compute_capability, 276 CudaComputeCapability cuda_compute_capability,
|
H A D | cuda_dnn.cc | 912 CudaComputeCapability cuda_compute_capability) { in TensorOpMathAvailable() argument 913 return cuda_compute_capability.IsAtLeast(7); in TensorOpMathAvailable() 4750 auto cuda_compute_capability = stream->GetCudaComputeCapability(); in GetConvolveRunners() local 4758 got_algos = GetConvolveAlgorithms(cuda_compute_capability, &algorithms); in GetConvolveRunners() 4761 got_algos = GetConvolveBackwardFilterAlgorithms(cuda_compute_capability, in GetConvolveRunners() 4765 got_algos = GetConvolveBackwardDataAlgorithms(cuda_compute_capability, in GetConvolveRunners() 5219 auto cuda_compute_capability = stream->GetCudaComputeCapability(); in GetFusedConvolveRunners() local 5220 if (!GetConvolveAlgorithms(cuda_compute_capability, &algorithms)) { in GetFusedConvolveRunners() 5275 CudaComputeCapability cuda_compute_capability, in GetConvolveAlgorithms() argument 5280 TensorOpMathAvailable(cuda_compute_capability); in GetConvolveAlgorithms() [all …]
|
/aosp_15_r20/external/tensorflow/tensorflow/stream_executor/rocm/ |
H A D | rocm_dnn.h | 237 CudaComputeCapability cuda_compute_capability, 276 CudaComputeCapability cuda_compute_capability, 280 CudaComputeCapability cuda_compute_capability,
|
/aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/ |
H A D | platform_util.cc | 131 se::CudaComputeCapability cc = description.cuda_compute_capability(); in IsDeviceSupported()
|