Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Metal] Support fast_math, preps for saturating_grid_dim #1443

Merged
merged 2 commits into from
Jul 14, 2020
Merged

[Metal] Support fast_math, preps for saturating_grid_dim #1443

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
de897b9
[Metal] Support fast_math, preps for saturating_grid_dim
k-ye Jul 9, 2020
69eefb3
grid dim
k-ye Jul 10, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion taichi/backends/metal/api.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,13 +64,14 @@ nsobj_unique_ptr<MTLBlitCommandEncoder> new_blit_command_encoder(

nsobj_unique_ptr<MTLLibrary> new_library_with_source(MTLDevice *device,
const std::string &source,
bool fast_math,
int msl_version) {
auto source_str = mac::wrap_string_as_ns_string(source);

id options = clscall("MTLCompileOptions", "alloc");
options = call(options, "init");
auto options_cleanup = wrap_as_nsobj_unique_ptr(options);
call(options, "setFastMathEnabled:", false);
call(options, "setFastMathEnabled:", fast_math);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cool! Not sure about how fast_math works. IIUC is this the same as specifying precision mediump float; in OpenGL?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

They are both for optimizing performance, but take different approaches. From what i can tell, precision mediump float reduces the bits to represent float. Metal also has a similar concept of half (16bits). Fast math, on the other hand, reduces the instructions in the computation to produce an approximated result.

if (msl_version != kMslVersionNone) {
call(options, "setLanguageVersion:", msl_version);
}
Expand Down
1 change: 1 addition & 0 deletions taichi/backends/metal/api.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,7 @@ nsobj_unique_ptr<MTLBlitCommandEncoder> new_blit_command_encoder(
// See https://developer.apple.com/documentation/metal/mtllanguageversion
nsobj_unique_ptr<MTLLibrary> new_library_with_source(MTLDevice *device,
const std::string &source,
bool fast_math,
int msl_version);

nsobj_unique_ptr<MTLFunction> new_function_with_name(MTLLibrary *library,
Expand Down
67 changes: 41 additions & 26 deletions taichi/backends/metal/kernel_manager.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,15 +84,17 @@ using InputBuffersMap = std::unordered_map<BufferEnum, MTLBuffer *>;
class CompiledMtlKernelBase {
public:
struct Params {
const KernelAttributes *kernel_attribs;
bool is_jit_evaluator;
const CompileConfig *config;
const KernelAttributes *kernel_attribs;
MTLDevice *device;
MTLFunction *mtl_func;
};

explicit CompiledMtlKernelBase(Params &params)
: is_jit_evalutor_(params.is_jit_evaluator),
kernel_attribs_(*params.kernel_attribs),
: kernel_attribs_(*params.kernel_attribs),
config_(params.config),
is_jit_evalutor_(params.is_jit_evaluator),
pipeline_state_(
new_compute_pipeline_state_with_function(params.device,
params.mtl_func)) {
Expand All @@ -113,7 +115,7 @@ class CompiledMtlKernelBase {

void launch_if_not_empty(BindBuffers buffers,
MTLCommandBuffer *command_buffer) {
const int num_threads = kernel_attribs_.num_threads;
const int num_threads = get_total_num_threads();
if (num_threads == 0) {
return;
}
Expand All @@ -130,40 +132,49 @@ class CompiledMtlKernelBase {
set_mtl_buffer(encoder.get(), b.first, /*offset=*/0, bi);
}

const int native_block_dim =
get_max_total_threads_per_threadgroup(pipeline_state_.get());

int num_threads_per_group = 0;
// Sometimes it is helpful to limit the maximum GPU block dim for the
// kernels. E.g., when you are generating iPhone shaders on a Mac.
const int prescribed_block_dim =
(std::size_t)get_current_program().config.max_block_dim;
if (prescribed_block_dim != 0) {
num_threads_per_group = std::min(native_block_dim, prescribed_block_dim);
} else {
num_threads_per_group = native_block_dim;
}
Comment on lines -136 to -145
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OFT?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not really :) These are the necessary changes for supporting saturing_grid_dim.


const int num_threads_per_group = get_num_threads_per_group(num_threads);
const int num_groups =
((num_threads + num_threads_per_group - 1) / num_threads_per_group);

const int dispatch_num_threads =
std::min(num_threads, num_threads_per_group);

if (!is_jit_evalutor_) {
ActionRecorder::get_instance().record(
"launch_kernel",
{ActionArg("kernel_name", kernel_attribs_.name),
ActionArg("num_groups", num_groups),
ActionArg("num_threads_per_group", dispatch_num_threads)});
ActionArg("num_threads_per_group", num_threads_per_group)});
}

dispatch_threadgroups(encoder.get(), num_groups, dispatch_num_threads);
dispatch_threadgroups(encoder.get(), num_groups, num_threads_per_group);
end_encoding(encoder.get());
}

const bool is_jit_evalutor_;
int get_total_num_threads() const {
int num_threads = kernel_attribs_.num_threads;
// TODO(k-ye): Surface |saturating_grid_dim| to ti.init() once #1396 is in.
// const int prescribed_grid_dim = config_->saturating_grid_dim;
// if (prescribed_grid_dim > 0) {
// num_threads = std::min(num_threads, prescribed_grid_dim);
// }
return num_threads;
}

int get_num_threads_per_group(int total_num_threads) const {
int num_threads_per_group =
get_max_total_threads_per_threadgroup(pipeline_state_.get());
// Sometimes it is helpful to limit the maximum GPU block dim for the
// kernels. E.g., when you are generating iPhone shaders on a Mac.
const int prescribed_block_dim = config_->max_block_dim;
if (prescribed_block_dim > 0) {
num_threads_per_group =
std::min(num_threads_per_group, prescribed_block_dim);
}
// Cap by |total_num_threads| in case this is a very small kernel.
return std::min(num_threads_per_group, total_num_threads);
}

KernelAttributes kernel_attribs_;
const CompileConfig *const config_;
const bool is_jit_evalutor_;
nsobj_unique_ptr<MTLComputePipelineState> pipeline_state_;
};

Expand Down Expand Up @@ -253,15 +264,16 @@ class CompiledTaichiKernel {
MTLDevice *device;
MemoryPool *mem_pool;
KernelProfilerBase *profiler;
const CompileConfig *compile_config;
};

CompiledTaichiKernel(Params params)
: ti_kernel_attribs(*params.ti_kernel_attribs),
ctx_attribs(*params.ctx_attribs) {
auto *const device = params.device;
auto kernel_lib = new_library_with_source(
device, params.mtl_source_code,
infer_msl_version(ti_kernel_attribs.used_features));
device, params.mtl_source_code, params.compile_config->fast_math,
infer_msl_version(params.ti_kernel_attribs->used_features));
if (kernel_lib == nullptr) {
TI_ERROR("Failed to compile Metal kernel! Generated code:\n\n{}",
params.mtl_source_code);
Expand All @@ -286,6 +298,7 @@ class CompiledTaichiKernel {
RuntimeListOpsMtlKernel::Params kparams;
kparams.kernel_attribs = &ka;
kparams.is_jit_evaluator = ti_kernel_attribs.is_jit_evaluator;
kparams.config = params.compile_config;
kparams.device = device;
kparams.mtl_func = mtl_func.get();
kparams.mem_pool = params.mem_pool;
Expand All @@ -295,6 +308,7 @@ class CompiledTaichiKernel {
UserMtlKernel::Params kparams;
kparams.kernel_attribs = &ka;
kparams.is_jit_evaluator = ti_kernel_attribs.is_jit_evaluator;
kparams.config = params.compile_config;
kparams.device = device;
kparams.mtl_func = mtl_func.get();
kernel = std::make_unique<UserMtlKernel>(kparams);
Expand Down Expand Up @@ -574,6 +588,7 @@ class KernelManager::Impl {
params.device = device_.get();
params.mem_pool = mem_pool_;
params.profiler = profiler_;
params.compile_config = config_;
compiled_taichi_kernels_[taichi_kernel_name] =
std::make_unique<CompiledTaichiKernel>(params);
TI_DEBUG("Registered Taichi kernel <{}>", taichi_kernel_name);
Expand Down
4 changes: 2 additions & 2 deletions tests/python/test_element_wise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@ def _c_mod(a, b):

@pytest.mark.parametrize('lhs_is_mat,rhs_is_mat', [(True, True), (True, False),
(False, True)])
@ti.all_archs
@ti.all_archs_with(fast_math=False)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does pytest.approx with larger rel= and abs= solve your issue?

Copy link
Member Author

@k-ye k-ye Jul 10, 2020 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nope, it failed at //, which translates to floor(y / z). For y[1][1] and z[1][1], they are both 3, so I guess that the division becomes something like 0.99999, which gets floored to 0.0 instead of 1.0. I just tested by setting y at a slightly higher value, e.g. 3.0001, and it worked.. But i think such kind of rounding error is expected when enabling fast math, so it's not a big deal? Maybe we should also disable fast math in tests by default.

def test_binary_f(lhs_is_mat, rhs_is_mat):
x = ti.Matrix(3, 2, ti.f32, 16)
if lhs_is_mat:
Expand Down Expand Up @@ -145,7 +145,7 @@ def func():


@pytest.mark.parametrize('rhs_is_mat', [True, False])
@ti.all_archs
@ti.all_archs_with(fast_math=False)
def test_writeback_binary_f(rhs_is_mat):
x = ti.Matrix(3, 2, ti.f32, 9)
y = ti.Matrix(3, 2, ti.f32, ())
Expand Down