Use an input file for tf_to_cubin.

Also use a template file for the mlir definitions of the tensorflow ops. Add
another build macro which converts the template to a mlir file by replacing
the type placeholders.

PiperOrigin-RevId: 315227102
Change-Id: I75cd2109c4b087122ae7ecb6b1d0d5fe35813a25
This commit is contained in:
Adrian Kuegel 2020-06-08 00:57:53 -07:00 committed by TensorFlower Gardener
parent c7f8f9ecfb
commit 56ef8c1ac2
6 changed files with 83 additions and 40 deletions

View File

@ -47,6 +47,7 @@ bool ParseStringList(std::string string_list, std::vector<uint32_t>* result) {
} // namespace } // namespace
int main(int argc, char** argv) { int main(int argc, char** argv) {
std::string input_file = "foo.mlir";
std::string output_file = "foo.bin"; std::string output_file = "foo.bin";
int32_t architecture = 50; int32_t architecture = 50;
std::vector<uint32_t> tile_sizes; std::vector<uint32_t> tile_sizes;
@ -75,6 +76,7 @@ int main(int argc, char** argv) {
}; };
std::vector<tensorflow::Flag> flag_list = { std::vector<tensorflow::Flag> flag_list = {
tensorflow::Flag("input", &input_file, "input file"),
tensorflow::Flag("output", &output_file, "output file"), tensorflow::Flag("output", &output_file, "output file"),
tensorflow::Flag("arch", &architecture, tensorflow::Flag("arch", &architecture,
"target architecture (e.g. 50 for sm_50)"), "target architecture (e.g. 50 for sm_50)"),
@ -94,8 +96,16 @@ int main(int argc, char** argv) {
std::pair<int32_t, int32_t> compute_capability(architecture / 10, std::pair<int32_t, int32_t> compute_capability(architecture / 10,
architecture % 10); architecture % 10);
std::string tf_code;
auto read_status = tensorflow::ReadFileToString(tensorflow::Env::Default(),
input_file, &tf_code);
if (!read_status.ok()) {
LOG(ERROR) << read_status;
return 1;
}
auto cubin = tensorflow::kernel_gen::GenerateCubinForTfCode( auto cubin = tensorflow::kernel_gen::GenerateCubinForTfCode(
argv[1], compute_capability, tile_sizes, same_shape, unroll_factors); tf_code, compute_capability, tile_sizes, same_shape, unroll_factors);
if (!cubin.ok()) { if (!cubin.ok()) {
LOG(ERROR) << cubin.status(); LOG(ERROR) << cubin.status();

View File

@ -6,18 +6,8 @@ package(
licenses = ["notice"], # Apache 2.0 licenses = ["notice"], # Apache 2.0
) )
bias_add_kernel = """
func @bias_add(%arg0: tensor<?x?xf99>,
%arg1: tensor<?xf99>) -> tensor<?x?xf99> {
%0 = "tf.BiasAdd"(%arg0, %arg1) { T = "tfdtype$DT_TYPE" }
: (tensor<?x?xf99>, tensor<?xf99>) -> tensor<?x?xf99>
return %0 : tensor<?x?xf99>
}
"""
gen_kernel_library( gen_kernel_library(
name = "bias_add", name = "bias_add",
op = bias_add_kernel,
same_shape = "0,2", same_shape = "0,2",
tile_size = "16x16", tile_size = "16x16",
types = [ types = [
@ -27,17 +17,8 @@ gen_kernel_library(
], ],
) )
relu_kernel = """
func @relu(%arg0: tensor<?xf99>) -> tensor<?xf99> {
%0 = "tf.Relu"(%arg0) { T = "tfdtype$DT_TYPE" }
: (tensor<?xf99>) -> tensor<?xf99>
return %0 : tensor<?xf99>
}
"""
gen_kernel_library( gen_kernel_library(
name = "relu", name = "relu",
op = relu_kernel,
same_shape = "0,1", same_shape = "0,1",
tile_size = "256", tile_size = "256",
types = [ types = [
@ -47,17 +28,8 @@ gen_kernel_library(
], ],
) )
tanh_kernel = """
func @tanh(%arg0: tensor<?xf99>) -> tensor<?xf99> {
%0 = "tf.Tanh"(%arg0) { T = "tfdtype$DT_TYPE" }
: (tensor<?xf99>) -> tensor<?xf99>
return %0 : tensor<?xf99>
}
"""
gen_kernel_library( gen_kernel_library(
name = "tanh", name = "tanh",
op = tanh_kernel,
tile_size = "256", tile_size = "256",
types = [ types = [
"f32", "f32",

View File

@ -0,0 +1,6 @@
func @bias_add(%arg0: tensor<?x?xf99>,
%arg1: tensor<?xf99>) -> tensor<?x?xf99> {
%0 = "tf.BiasAdd"(%arg0, %arg1) { T = "tfdtype$DT_TYPE" }
: (tensor<?x?xf99>, tensor<?xf99>) -> tensor<?x?xf99>
return %0 : tensor<?x?xf99>
}

View File

@ -27,13 +27,14 @@ def _gen_kernel_image_hdr_impl(ctx):
filename = "%s.%s.cubin" % (name, arch) filename = "%s.%s.cubin" % (name, arch)
cubin = ctx.actions.declare_file(filename) cubin = ctx.actions.declare_file(filename)
ctx.actions.run( ctx.actions.run(
inputs = [ctx.file.mlir_op],
outputs = [cubin], outputs = [cubin],
executable = ctx.executable._tool, executable = ctx.executable._tool,
arguments = same_shape + [ arguments = same_shape + [
"--tile_sizes=%s" % tile_sizes, "--tile_sizes=%s" % tile_sizes,
"--arch=%s" % arch.split("_")[1], "--arch=%s" % arch.split("_")[1],
"--input=%s" % ctx.file.mlir_op.path,
"--output=%s" % cubin.path, "--output=%s" % cubin.path,
ctx.attr.op,
], ],
mnemonic = "compile", mnemonic = "compile",
) )
@ -70,7 +71,7 @@ _gen_kernel_image_hdr_rule = rule(
implementation = _gen_kernel_image_hdr_impl, implementation = _gen_kernel_image_hdr_impl,
output_to_genfiles = True, output_to_genfiles = True,
attrs = { attrs = {
"op": attr.string(mandatory = True), "mlir_op": attr.label(mandatory = True, allow_single_file = True),
"tile_size": attr.string(mandatory = True), "tile_size": attr.string(mandatory = True),
"same_shape": attr.string(), "same_shape": attr.string(),
"out": attr.output(mandatory = True), "out": attr.output(mandatory = True),
@ -87,12 +88,12 @@ _gen_kernel_image_hdr_rule = rule(
}, },
) )
def _gen_kernel_image_hdr(name, op, tile_size, tags = [], same_shape = None): def _gen_kernel_image_hdr(name, mlir_op, tile_size, tags = [], same_shape = None):
"""Generates a C header with fatbin data from a Tensorflow op.""" """Generates a C header with fatbin data from a Tensorflow op."""
if cuda_gpu_architectures(): if cuda_gpu_architectures():
_gen_kernel_image_hdr_rule( _gen_kernel_image_hdr_rule(
name = name, name = name,
op = op, mlir_op = mlir_op,
tile_size = tile_size, tile_size = tile_size,
same_shape = same_shape, same_shape = same_shape,
out = "%s.h" % name, out = "%s.h" % name,
@ -101,17 +102,61 @@ def _gen_kernel_image_hdr(name, op, tile_size, tags = [], same_shape = None):
tags = tags, tags = tags,
) )
def gen_kernel_library(name, op, types, tile_size, tags = [], same_shape = None): def _gen_mlir_op_impl(ctx):
if cuda_gpu_architectures():
type_to_dtype = { type_to_dtype = {
"f16": "DT_HALF", "f16": "DT_HALF",
"f32": "DT_FLOAT", "f32": "DT_FLOAT",
"f64": "DT_DOUBLE", "f64": "DT_DOUBLE",
} }
ctx.actions.run_shell(
inputs = [ctx.file.template],
outputs = [ctx.outputs.out],
command = "cat %s | sed s/f99/%s/g | sed s/DT_DTYPE/%s/g > %s" % (
ctx.file.template.path,
ctx.attr.type,
type_to_dtype[ctx.attr.type],
ctx.outputs.out.path,
),
)
_gen_mlir_op_rule = rule(
implementation = _gen_mlir_op_impl,
output_to_genfiles = True,
attrs = {
"template": attr.label(mandatory = True, allow_single_file = True),
"type": attr.string(mandatory = True),
"out": attr.output(mandatory = True),
},
)
def _gen_mlir_op(name, type):
_gen_mlir_op_rule(
name = "generate_{name}_{type}_mlir".format(name = name, type = type),
template = "{name}.mlir.tmpl".format(name = name),
type = type,
out = "{name}_{type}.mlir".format(name = name, type = type),
)
def gen_kernel_library(name, types, tile_size, tags = [], same_shape = None):
""" Generate a library with kernels for a specific tensorflow op.
Args:
name: The name of the tensorflow op.
types: The types ("f16", "f32", "f64") for which a kernel should be generated.
tile_size: The tiling specification, e.g. "16x16".
tags: The tags which should be added to the library.
same_shape: The information about which shapes are the same, e.g. "0,1".
"""
if cuda_gpu_architectures():
for type in types: for type in types:
_gen_mlir_op(
name = name,
type = type,
)
_gen_kernel_image_hdr( _gen_kernel_image_hdr(
name = "{name}_{type}_kernel".format(name = name, type = type), name = "{name}_{type}_kernel".format(name = name, type = type),
op = op.replace("f99", type).replace("DT_TYPE", type_to_dtype[type]), mlir_op = "{name}_{type}.mlir".format(name = name, type = type),
tile_size = tile_size, tile_size = tile_size,
tags = tags, tags = tags,
same_shape = same_shape, same_shape = same_shape,

View File

@ -0,0 +1,5 @@
func @relu(%arg0: tensor<?xf99>) -> tensor<?xf99> {
%0 = "tf.Relu"(%arg0) { T = "tfdtype$DT_TYPE" }
: (tensor<?xf99>) -> tensor<?xf99>
return %0 : tensor<?xf99>
}

View File

@ -0,0 +1,5 @@
func @tanh(%arg0: tensor<?xf99>) -> tensor<?xf99> {
%0 = "tf.Tanh"(%arg0) { T = "tfdtype$DT_DTYPE" }
: (tensor<?xf99>) -> tensor<?xf99>
return %0 : tensor<?xf99>
}