| /external/ruy/example/ |
| D | example.cc | 21 void ExampleMulFloat(ruy::Context *context) { in ExampleMulFloat()
26 ruy::Matrix<float> lhs; in ExampleMulFloat()
27 ruy::MakeSimpleLayout(2, 2, ruy::Order::kRowMajor, lhs.mutable_layout()); in ExampleMulFloat()
29 ruy::Matrix<float> rhs; in ExampleMulFloat()
30 ruy::MakeSimpleLayout(2, 2, ruy::Order::kColMajor, rhs.mutable_layout()); in ExampleMulFloat()
32 ruy::Matrix<float> dst; in ExampleMulFloat()
33 ruy::MakeSimpleLayout(2, 2, ruy::Order::kColMajor, dst.mutable_layout()); in ExampleMulFloat()
36 ruy::MulParams<float, float> mul_params; in ExampleMulFloat()
37 ruy::Mul(lhs, rhs, mul_params, context, &dst); in ExampleMulFloat()
45 void ExampleMulFloatWithBiasAddAndClamp(ruy::Context *context) { in ExampleMulFloatWithBiasAddAndClamp()
[all …]
|
| D | parametrized_example.cc | 69 ruy::Matrix<LhsType> lhs; in run()
70 ruy::Matrix<RhsType> rhs; in run()
71 ruy::Matrix<DstType> dst; in run()
75 return ruy::Order::kRowMajor; in run()
77 return ruy::Order::kColMajor; in run()
84 auto make_layout = [](int rows, int cols, int stride, ruy::Order order, in run()
85 ruy::Layout* layout) { in run()
89 int base_stride = order == ruy::Order::kRowMajor ? cols : rows; in run()
104 lhs.set_cache_policy(static_cast<ruy::CachePolicy>(params.lhs_cache_policy)); in run()
105 rhs.set_cache_policy(static_cast<ruy::CachePolicy>(params.rhs_cache_policy)); in run()
[all …]
|
| D | BUILD | 9 deps = ["//ruy"],
15 deps = ["//ruy"],
|
| D | README.md | 9 bazel build //ruy/example:example
13 ./bazel-bin/ruy/example
|
| D | CMakeLists.txt | 11 ruy
20 ruy
|
| /external/ruy/ |
| D | Android.bp | 45 "ruy/allocator.cc",
46 "ruy/apply_multiplier.cc",
47 "ruy/block_map.cc",
48 "ruy/blocking_counter.cc",
49 "ruy/context.cc",
50 "ruy/context_get_ctx.cc",
51 "ruy/cpuinfo.cc",
52 "ruy/ctx.cc",
53 "ruy/frontend.cc",
54 "ruy/have_built_path_for_avx.cc",
[all …]
|
| D | CMakeLists.txt | 17 project(ruy CXX) project
31 option(RUY_MINIMAL_BUILD "Disable ruy's tests, examples, etc. Build only ruy public libraries." ${R…
36 option(RUY_PROFILER "Enable ruy's built-in profiler (harms performance)" OFF)
43 # Skip cpuinfo if it was already generated, which can happen when ruy is
60 # ruy is a subdirectory in a wider project that is already providing
62 # in such a way that cpuinfo gets generated after ruy. In that case,
64 # target never gets defined, ruy will fail to compile.
65 # On the other hand, if ruy is the top-level project here (not part of a
70 "That typically means that the git submodules of the ruy "
71 "repository haven't been checked out. Try this in the ruy "
[all …]
|
| D | README.md | 1 # The ruy matrix multiplication library
5 ruy is a matrix multiplication library. Its focus is to cover the matrix
9 ruy supports both floating-point and 8bit-integer-quantized matrices.
13 ruy is designed to achieve high performance not just on very large sizes, as
|
| /external/tensorflow/tensorflow/lite/kernels/ |
| D | cpu_backend_gemm_ruy.h | 30 inline ruy::CachePolicy ToRuyCachePolicy(CachePolicy cache_policy) { in ToRuyCachePolicy()
33 return ruy::CachePolicy::kNeverCache; in ToRuyCachePolicy()
35 return ruy::CachePolicy::kCacheIfLargeSpeedup; in ToRuyCachePolicy()
37 return ruy::CachePolicy::kAlwaysCache; in ToRuyCachePolicy()
40 return ruy::CachePolicy::kNeverCache; in ToRuyCachePolicy()
46 ruy::Matrix<Scalar>* dst, bool use_caching = false) {
47 ruy::Order ruy_order = params.order == Order::kColMajor
48 ? ruy::Order::kColMajor
49 : ruy::Order::kRowMajor;
50 ruy::MakeSimpleLayout(params.rows, params.cols, ruy_order,
[all …]
|
| D | BUILD | 14 # Enables usage of ruy exclusively as the GEMM backend in TFLite kernels.
15 # This will cause TFLite to build with ruy only, providing a smaller binary.
22 # Disables usage of ruy as the exclusive GEMM backend in TFLite kernels.
23 # TFLite will be built with ruy and other GEMM libraries. Ruy will not be
357 # For now this unconditionally depends on both ruy and gemmlowp.
360 "@ruy//ruy:context",
386 # For now this unconditionally depends on both ruy and gemmlowp.
390 "@ruy//ruy:context",
391 "@ruy//ruy:thread_pool",
430 # Depend on ruy regardless of `tflite_with_ruy`. See the comment in
[all …]
|
| D | cpu_backend_context.h | 40 ruy::Context* ruy_context() const { return ruy_context_.get(); } in ruy_context()
97 const std::unique_ptr<ruy::Context> ruy_context_;
|
| /external/ruy/ruy/ |
| D | prepacked_cache_test.cc | 28 namespace ruy { namespace
230 ruy::Context context; in TestCachePolicies()
231 ruy::Ctx* ctx = get_ctx(&context); in TestCachePolicies()
239 ruy::Matrix<float> lhs; in TestCachePolicies()
240 ruy::MakeSimpleLayout(2, 2, ruy::Order::kRowMajor, lhs.mutable_layout()); in TestCachePolicies()
242 ruy::Matrix<float> rhs; in TestCachePolicies()
243 ruy::MakeSimpleLayout(2, 1, ruy::Order::kColMajor, rhs.mutable_layout()); in TestCachePolicies()
245 ruy::Matrix<float> dst; in TestCachePolicies()
246 ruy::MakeSimpleLayout(2, 1, ruy::Order::kColMajor, dst.mutable_layout()); in TestCachePolicies()
249 ruy::MulParams<float, float> mul_params; in TestCachePolicies()
[all …]
|
| D | perchannel_buffers_reallocation_test.cc | 9 namespace ruy { namespace
52 ruy::Matrix<LhsScalar> lhs; in TestPerChannelBuffersReallocation()
53 ruy::MakeSimpleLayout(kMatrixSize, kMatrixSize, ruy::Order::kRowMajor, in TestPerChannelBuffersReallocation()
57 ruy::Matrix<RhsScalar> rhs; in TestPerChannelBuffersReallocation()
58 ruy::MakeSimpleLayout(kMatrixSize, kMatrixSize, ruy::Order::kColMajor, in TestPerChannelBuffersReallocation()
63 ruy::Matrix<DstScalar> dst; in TestPerChannelBuffersReallocation()
64 ruy::MakeSimpleLayout(kMatrixSize, kMatrixSize, ruy::Order::kColMajor, in TestPerChannelBuffersReallocation()
68 ruy::Context context; in TestPerChannelBuffersReallocation()
75 ruy::Mul<kPath>(lhs, rhs, mul_params, &context, &dst); in TestPerChannelBuffersReallocation()
|
| D | build_defs.bzl | 4 # Returns warnings flags to use for all ruy code.
29 # Returns flags to use to enable NEON if applicable, for all ruy code.
34 "//ruy:arm32_assuming_neon": [
41 # Returns optimization flags to use for all ruy code.
49 "//ruy:do_not_want_O3": [],
53 # Returns compiler flags to use for all ruy code.
59 "//ruy:x86_64_and_not_msvc": ["-mavx"],
66 "//ruy:x86_64_and_not_msvc": ["-mavx2", "-mfma"],
76 …"//ruy:x86_64_and_not_msvc": ["$(STACK_FRAME_UNLIMITED)", "-mavx512f", "-mavx512vl", "-mavx512cd",…
|
| D | BUILD | 230 "//ruy/profiler:instrumentation",
257 ":ruy",
311 "//ruy/profiler:instrumentation",
530 "//ruy/profiler:instrumentation",
548 "//ruy/profiler:instrumentation",
572 "//ruy/profiler:instrumentation",
594 "//ruy/profiler:instrumentation",
616 "//ruy/profiler:instrumentation",
637 "//ruy/profiler:instrumentation",
674 "//ruy/profiler:instrumentation",
[all …]
|
| D | blocking_counter.cc | 21 namespace ruy { namespace
46 ruy::Wait(condition, spin_duration, &count_cond_, &count_mutex_); in Wait()
|
| /external/tensorflow/third_party/ruy/ |
| D | workspace.bzl | 1 """Loads the ruy library, used by TensorFlow Lite."""
7 name = "ruy",
9 strip_prefix = "ruy-4790797d11a81f96baf24f3731fd3ca44c2c5f8b",
11 …"https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/ruy/archive/4790797d11a81f…
12 "https://github.com/google/ruy/archive/4790797d11a81f96baf24f3731fd3ca44c2c5f8b.zip",
14 build_file = "//third_party/ruy:BUILD",
|
| /external/tensorflow/tensorflow/lite/tools/cmake/modules/ |
| D | ruy.cmake | 16 if(TARGET ruy OR ruy_POPULATED)
23 ruy
24 GIT_REPOSITORY https://github.com/google/ruy
28 SOURCE_DIR "${CMAKE_BINARY_DIR}/ruy"
30 OverridableFetchContent_GetProperties(ruy)
32 OverridableFetchContent_Populate(ruy)
38 "${CMAKE_CURRENT_LIST_DIR}/ruy"
|
| /external/XNNPACK/bench/ |
| D | qu8-gemm.cc | 232 static ruy::Context context; in RuyBenchmark()
235 ruy::Matrix<uint8_t> ruy_a; in RuyBenchmark()
236 ruy::MakeSimpleLayout(nc, kc, ruy::Order::kRowMajor, ruy_a.mutable_layout()); in RuyBenchmark()
238 ruy::Matrix<uint8_t> ruy_b; in RuyBenchmark()
239 ruy::MakeSimpleLayout(kc, mc, ruy::Order::kColMajor, ruy_b.mutable_layout()); in RuyBenchmark()
242 ruy::Matrix<uint8_t> ruy_c; in RuyBenchmark()
243 ruy::MakeSimpleLayout(nc, mc, ruy::Order::kColMajor, ruy_c.mutable_layout()); in RuyBenchmark()
246 ruy::MulParams<int32_t, uint8_t> mul_params; in RuyBenchmark()
261 ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); in RuyBenchmark()
281 ruy::Mul(ruy_a, ruy_b, mul_params, &context, &ruy_c); in RuyBenchmark()
|
| /external/tensorflow/tensorflow/lite/tools/cmake/modules/ruy/ |
| D | CMakeLists.txt | 18 project(ruy CXX) project
30 file(GLOB RUY_SOURCES "${RUY_SOURCE_DIR}/ruy/*.*")
36 add_library(ruy ${RUY_SOURCES}) target
37 target_include_directories(ruy PUBLIC "${RUY_SOURCE_DIR}")
|
| /external/ruy/ruy/profiler/ |
| D | test_instrumented_library.cc | 23 ruy::profiler::ScopeLabel function_label( in MergeSortRecurse()
33 ruy::profiler::ScopeLabel merging_sorted_halves_label( in MergeSortRecurse()
56 ruy::profiler::ScopeLabel function_label("MergeSort (size=%d)", size); in MergeSort()
|
| D | README.md | 5 The present directory is the "ruy profiler". As a time profiler, it allows to
33 #include "ruy/profiler/instrumentation.h"
38 ruy::profiler::ScopeLabel function_label("SomeFunction");
47 ruy::profiler::ScopeLabel extra_work_label("Some more work");
61 ruy::profiler::ScopeLabel function_label("SomeFunction (size=%d)", size);
83 ruy::profiler::ScopeProfile profile;
95 //tensorflow/lite/experimental/ruy/profiler:test
|
| /external/tensorflow/tensorflow/lite/kernels/internal/optimized/sparse_ops/ |
| D | fully_connected.h | 36 ruy::profiler::ScopeLabel label("FullyConnected"); in FullyConnectedSparseWeight()
37 ruy::profiler::ScopeLabel inner_label("Random Sparse"); in FullyConnectedSparseWeight()
83 ruy::profiler::ScopeLabel label("FullyConnected"); in FullyConnectedSparseWeight1x4Impl()
84 ruy::profiler::ScopeLabel inner_label("1x4 Block Sparse"); in FullyConnectedSparseWeight1x4Impl()
104 ruy::profiler::ScopeLabel activation_label("activation function"); in FullyConnectedSparseWeight1x4Impl()
|
| /external/tensorflow/tensorflow/lite/tools/make/ |
| D | Makefile | 39 -I$(MAKEFILE_DIR)/downloads/ruy \
143 $(wildcard tensorflow/lite/tools/make/downloads/ruy/ruy/*.cc)
219 CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/instrumentation.cc
220 CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/profiler.cc
221 CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/treeview.cc
|
| /external/tensorflow/tensorflow/lite/kernels/internal/ |
| D | BUILD | 275 "@ruy//ruy/profiler:instrumentation",
326 "@ruy//ruy/profiler:instrumentation",
546 "@ruy//ruy/profiler:instrumentation",
646 "@ruy//ruy/profiler:instrumentation",
741 "@ruy//ruy",
765 "@ruy//ruy/profiler:instrumentation",
920 "@ruy//ruy:context",
|