ntt123's picture
add sparse matmul
21f3d42
# Low-level computation code, including generic and architecture-specific
# variants.
licenses(["notice"])
cc_library(
name = "gru_gates",
srcs = [
"ar_inputs.h",
"gru_gates_arm.h",
"gru_gates_avx_fixed.h",
"gru_gates_generic.h",
],
hdrs = ["gru_gates.h"],
visibility = [
"//visibility:public",
],
deps = [
":matmul",
"//sparse_matmul/numerics:fast_transcendentals",
"//sparse_matmul/numerics:types",
"//sparse_matmul/vector:cache_aligned_vector",
],
)
cc_library(
name = "kernels",
srcs = [
"kernels_arm.h",
"kernels_avx.h",
],
hdrs = [
"kernels_generic.h",
],
visibility = [
"//sparse_matmul:__subpackages__",
],
deps = [
"//sparse_matmul/numerics:fast_transcendentals",
"//sparse_matmul/numerics:types",
],
)
cc_library(
name = "matmul",
srcs = [
"matmul_fixed_avx2.cc",
"matmul_fixed_avx2.h",
"matmul_generic.cc",
"matmul_generic.h",
],
hdrs = [
"matmul.h",
],
visibility = [
"//sparse_matmul:__subpackages__",
],
deps = [
"//sparse_matmul/numerics:types",
"@com_google_absl//absl/time",
],
)
cc_library(
name = "thread_bounds",
srcs = ["thread_bounds.cc"],
hdrs = ["thread_bounds.h"],
visibility = [
"//sparse_matmul:__subpackages__",
],
deps = [
"@com_google_glog//:glog",
],
)
cc_test(
name = "gru_gates_test",
size = "small",
srcs = [
"gru_gates_test.cc",
],
deps = [
":gru_gates",
"@com_google_absl//absl/memory",
"@com_google_absl//absl/types:span",
"@com_google_googletest//:gtest_main",
],
)