Skip to content

Vulkan: accept a SymInt delegate arg as a plain Int EValue #21848

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

Open
msluszniak wants to merge 1 commit into
base: main
Choose a base branch
from
+46 −6
Open

Vulkan: accept a SymInt delegate arg as a plain Int EValue #21848

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
52 changes: 46 additions & 6 deletions backends/vulkan/runtime/VulkanBackend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -748,12 +748,26 @@ class VulkanBackend final : public ::executorch::runtime::BackendInterface {
args[i]->toTensor().numel(),
equivalent_scalar_type(args[i]->toTensor().scalar_type()));
} else if (compute_graph->val_is_symint(iref)) {
VK_CHECK_COND(
args[i]->isTensor(),
"Cannot handle symint arg to graph that is not derived from a "
"scalar tensor at the moment.");
bool was_updated = maybe_update_scalar_tensor(
compute_graph, iref, args[i]->toTensor());
// A SymInt input may arrive either as a scalar tensor (the convention
// when the value was lifted from one) or as a plain Int EValue, which
// is what the emitter produces when the Method's schema declares the
// argument as SymInt. Accept both.
bool was_updated = false;
if (args[i]->isTensor()) {
was_updated = maybe_update_scalar_tensor(
compute_graph, iref, args[i]->toTensor());
} else if (args[i]->isInt()) {
const int32_t cur_val = compute_graph->read_symint(iref);
const int32_t new_val = static_cast<int32_t>(args[i]->toInt());
if (new_val != cur_val) {
compute_graph->set_symint(iref, new_val);
was_updated = true;
}
} else {
VK_THROW(
"Could not handle symint arg: caller's EValue is neither "
"a scalar tensor nor an Int.");
}
// Since symint inputs may impact tensor's sizes, trigger a resize if
// any symbolic integer shapes are updated.
should_propagate_resize = should_propagate_resize || was_updated;
Expand Down Expand Up @@ -830,6 +844,32 @@ class VulkanBackend final : public ::executorch::runtime::BackendInterface {
args[o]->toTensor().mutable_data_ptr(),
args[o]->toTensor().numel(),
equivalent_scalar_type(args[o]->toTensor().scalar_type()));
} else if (compute_graph->val_is_symint(oref)) {
// Mirror of the input path: write the graph's current SymInt value
// into the caller's slot, which may be either a scalar tensor or a
// plain Int EValue.
const int32_t out_val = compute_graph->read_symint(oref);
if (args[o]->isTensor()) {
executorch::aten::Tensor& dst = args[o]->toTensor();
const auto dtype = dst.scalar_type();
if (dtype == executorch::aten::ScalarType::Int) {
*dst.mutable_data_ptr<int32_t>() = out_val;
} else if (dtype == executorch::aten::ScalarType::Long) {
*dst.mutable_data_ptr<int64_t>() = static_cast<int64_t>(out_val);
} else {
VK_THROW(
"Could not handle symint output with destination tensor dtype ",
static_cast<int>(dtype));
}
} else if (args[o]->isInt()) {
// Overwrite the EValue's int payload in place. EValue stores its int
// as int64_t, so a SymInt fits without truncation.
*args[o] = EValue(static_cast<int64_t>(out_val));
} else {
VK_THROW(
"Could not handle symint output: caller's EValue is neither "
"a scalar tensor nor an Int.");
}
}
// TensorRef values represent constant tensors which will not have been
// modified by the graph execution. Therefore, if a constant tensor is
Expand Down
Loading