Rollup merge of #139351 - EnzymeAD:autodiff-batching2, r=oli-obk

Autodiff batching2

~I will rebase it once my first PR landed.~ done.
This autodiff batch mode is more similar to scalar autodiff, since it still only takes one shadow argument.
However, that argument is supposed to be `width` times larger.

r? `@oli-obk`

Tracking:

- https://github.com/rust-lang/rust/issues/124509
This commit is contained in:
Matthias Krüger 2025-04-17 21:53:23 +02:00 committed by GitHub
commit 87a163523f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 220 additions and 28 deletions

View file

@ -123,7 +123,7 @@ impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
/// Empty string, to be used where LLVM expects an instruction name, indicating
/// that the instruction is to be left unnamed (i.e. numbered, in textual IR).
// FIXME(eddyb) pass `&CStr` directly to FFI once it's a thin pointer.
const UNNAMED: *const c_char = c"".as_ptr();
pub(crate) const UNNAMED: *const c_char = c"".as_ptr();
impl<'ll, CX: Borrow<SCx<'ll>>> BackendTypes for GenericBuilder<'_, 'll, CX> {
type Value = <GenericCx<'ll, CX> as BackendTypes>::Value;

View file

@ -10,7 +10,7 @@ use rustc_middle::bug;
use tracing::{debug, trace};
use crate::back::write::llvm_err;
use crate::builder::SBuilder;
use crate::builder::{SBuilder, UNNAMED};
use crate::context::SimpleCx;
use crate::declare::declare_simple_fn;
use crate::errors::{AutoDiffWithoutEnable, LlvmError};
@ -51,6 +51,7 @@ fn has_sret(fnc: &Value) -> bool {
// using iterators and peek()?
fn match_args_from_caller_to_enzyme<'ll>(
cx: &SimpleCx<'ll>,
builder: &SBuilder<'ll, 'll>,
width: u32,
args: &mut Vec<&'ll llvm::Value>,
inputs: &[DiffActivity],
@ -78,7 +79,9 @@ fn match_args_from_caller_to_enzyme<'ll>(
let enzyme_const = cx.create_metadata("enzyme_const".to_string()).unwrap();
let enzyme_out = cx.create_metadata("enzyme_out".to_string()).unwrap();
let enzyme_dup = cx.create_metadata("enzyme_dup".to_string()).unwrap();
let enzyme_dupv = cx.create_metadata("enzyme_dupv".to_string()).unwrap();
let enzyme_dupnoneed = cx.create_metadata("enzyme_dupnoneed".to_string()).unwrap();
let enzyme_dupnoneedv = cx.create_metadata("enzyme_dupnoneedv".to_string()).unwrap();
while activity_pos < inputs.len() {
let diff_activity = inputs[activity_pos as usize];
@ -90,13 +93,34 @@ fn match_args_from_caller_to_enzyme<'ll>(
DiffActivity::Active => (enzyme_out, false),
DiffActivity::ActiveOnly => (enzyme_out, false),
DiffActivity::Dual => (enzyme_dup, true),
DiffActivity::Dualv => (enzyme_dupv, true),
DiffActivity::DualOnly => (enzyme_dupnoneed, true),
DiffActivity::DualvOnly => (enzyme_dupnoneedv, true),
DiffActivity::Duplicated => (enzyme_dup, true),
DiffActivity::DuplicatedOnly => (enzyme_dupnoneed, true),
DiffActivity::FakeActivitySize => (enzyme_const, false),
DiffActivity::FakeActivitySize(_) => (enzyme_const, false),
};
let outer_arg = outer_args[outer_pos];
args.push(cx.get_metadata_value(activity));
if matches!(diff_activity, DiffActivity::Dualv) {
let next_outer_arg = outer_args[outer_pos + 1];
let elem_bytes_size: u64 = match inputs[activity_pos + 1] {
DiffActivity::FakeActivitySize(Some(s)) => s.into(),
_ => bug!("incorrect Dualv handling recognized."),
};
// stride: sizeof(T) * n_elems.
// n_elems is the next integer.
// Now we multiply `4 * next_outer_arg` to get the stride.
let mul = unsafe {
llvm::LLVMBuildMul(
builder.llbuilder,
cx.get_const_i64(elem_bytes_size),
next_outer_arg,
UNNAMED,
)
};
args.push(mul);
}
args.push(outer_arg);
if duplicated {
// We know that duplicated args by construction have a following argument,
@ -114,7 +138,7 @@ fn match_args_from_caller_to_enzyme<'ll>(
} else {
let next_activity = inputs[activity_pos + 1];
// We analyze the MIR types and add this dummy activity if we visit a slice.
next_activity == DiffActivity::FakeActivitySize
matches!(next_activity, DiffActivity::FakeActivitySize(_))
}
};
if slice {
@ -125,7 +149,10 @@ fn match_args_from_caller_to_enzyme<'ll>(
// int2 >= int1, which means the shadow vector is large enough to store the gradient.
assert_eq!(cx.type_kind(next_outer_ty), TypeKind::Integer);
for i in 0..(width as usize) {
let iterations =
if matches!(diff_activity, DiffActivity::Dualv) { 1 } else { width as usize };
for i in 0..iterations {
let next_outer_arg2 = outer_args[outer_pos + 2 * (i + 1)];
let next_outer_ty2 = cx.val_ty(next_outer_arg2);
assert_eq!(cx.type_kind(next_outer_ty2), TypeKind::Pointer);
@ -136,7 +163,7 @@ fn match_args_from_caller_to_enzyme<'ll>(
}
args.push(cx.get_metadata_value(enzyme_const));
args.push(next_outer_arg);
outer_pos += 2 + 2 * width as usize;
outer_pos += 2 + 2 * iterations;
activity_pos += 2;
} else {
// A duplicated pointer will have the following two outer_fn arguments:
@ -360,6 +387,7 @@ fn generate_enzyme_call<'ll>(
let outer_args: Vec<&llvm::Value> = get_params(outer_fn);
match_args_from_caller_to_enzyme(
&cx,
&builder,
attrs.width,
&mut args,
&attrs.input_activity,