[artiq] Fix RPC of ndarrays from host

This commit is contained in:
David Mak 2024-08-14 15:53:27 +08:00
parent 02d93b11d1
commit a59c26aa99
2 changed files with 302 additions and 61 deletions

View File

@ -2,7 +2,7 @@ use nac3core::{
codegen::{
classes::{
ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayType,
NDArrayValue, RangeValue, UntypedArrayLikeAccessor,
NDArrayValue, ProxyType, ProxyValue, RangeValue, UntypedArrayLikeAccessor,
},
expr::{destructure_range, gen_call},
irrt::call_ndarray_calc_size,
@ -22,7 +22,7 @@ use inkwell::{
module::Linkage,
types::{BasicType, IntType},
values::{BasicValueEnum, PointerValue, StructValue},
AddressSpace, IntPredicate,
AddressSpace, IntPredicate, OptimizationLevel,
};
use pyo3::{
@ -32,6 +32,7 @@ use pyo3::{
use crate::{symbol_resolver::InnerResolver, timeline::TimeFns};
use inkwell::values::IntValue;
use itertools::Itertools;
use std::{
collections::{hash_map::DefaultHasher, HashMap},
@ -486,13 +487,10 @@ fn format_rpc_arg<'ctx>(
let buffer = ctx.builder.build_array_alloca(llvm_i8, buffer_size, "rpc.arg").unwrap();
let buffer = ArraySliceValue::from_ptr_val(buffer, buffer_size, Some("rpc.arg"));
let ppdata = generator.gen_var_alloc(ctx, llvm_arg_ty.element_type(), None).unwrap();
ctx.builder.build_store(ppdata, llvm_arg.data().base_ptr(ctx, generator)).unwrap();
call_memcpy_generic(
ctx,
buffer.base_ptr(ctx, generator),
ppdata,
llvm_arg.ptr_to_data(ctx),
llvm_pdata_sizeof,
llvm_i1.const_zero(),
);
@ -528,6 +526,298 @@ fn format_rpc_arg<'ctx>(
arg_slot
}
/// Formats an RPC return value to conform to the expected format required by NAC3.
fn format_rpc_ret<'ctx>(
generator: &mut dyn CodeGenerator,
ctx: &mut CodeGenContext<'ctx, '_>,
ret_ty: Type,
) -> Option<BasicValueEnum<'ctx>> {
// -- receive value:
// T result = {
// void *ret_ptr = alloca(sizeof(T));
// void *ptr = ret_ptr;
// loop: int size = rpc_recv(ptr);
// // Non-zero: Provide `size` bytes of extra storage for variable-length data.
// if(size) { ptr = alloca(size); goto loop; }
// else *(T*)ret_ptr
// }
let llvm_i8 = ctx.ctx.i8_type();
let llvm_i32 = ctx.ctx.i32_type();
let llvm_i8_8 = ctx.ctx.struct_type(&[llvm_i8.array_type(8).into()], false);
let llvm_pi8 = llvm_i8.ptr_type(AddressSpace::default());
let rpc_recv = ctx.module.get_function("rpc_recv").unwrap_or_else(|| {
ctx.module.add_function("rpc_recv", llvm_i32.fn_type(&[llvm_pi8.into()], false), None)
});
if ctx.unifier.unioned(ret_ty, ctx.primitives.none) {
ctx.build_call_or_invoke(rpc_recv, &[llvm_pi8.const_null().into()], "rpc_recv");
return None;
}
let prehead_bb = ctx.builder.get_insert_block().unwrap();
let current_function = prehead_bb.get_parent().unwrap();
let head_bb = ctx.ctx.append_basic_block(current_function, "rpc.head");
let alloc_bb = ctx.ctx.append_basic_block(current_function, "rpc.continue");
let tail_bb = ctx.ctx.append_basic_block(current_function, "rpc.tail");
let llvm_ret_ty = ctx.get_llvm_abi_type(generator, ret_ty);
let result = match &*ctx.unifier.get_ty_immutable(ret_ty) {
TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
let llvm_i1 = ctx.ctx.bool_type();
let llvm_usize = generator.get_size_type(ctx.ctx);
// Round `val` up to its modulo `power_of_two`
let round_up = |ctx: &mut CodeGenContext<'ctx, '_>,
val: IntValue<'ctx>,
power_of_two: IntValue<'ctx>| {
debug_assert_eq!(
val.get_type().get_bit_width(),
power_of_two.get_type().get_bit_width()
);
let llvm_val_t = val.get_type();
let max_rem = ctx
.builder
.build_int_sub(power_of_two, llvm_val_t.const_int(1, false), "")
.unwrap();
ctx.builder
.build_and(
ctx.builder.build_int_add(val, max_rem, "").unwrap(),
ctx.builder.build_not(max_rem, "").unwrap(),
"",
)
.unwrap()
};
// Setup types
let (elem_ty, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, ret_ty);
let llvm_elem_ty = ctx.get_llvm_type(generator, elem_ty);
let llvm_ret_ty = NDArrayType::new(generator, ctx.ctx, llvm_elem_ty);
// Allocate the resulting ndarray
// A condition after format_rpc_ret ensures this will not be popped this off.
let ndarray = llvm_ret_ty.new_value(generator, ctx, Some("rpc.result"));
// Setup ndims
let ndims =
if let TypeEnum::TLiteral { values, .. } = &*ctx.unifier.get_ty_immutable(ndims) {
assert_eq!(values.len(), 1);
u64::try_from(values[0].clone()).unwrap()
} else {
unreachable!();
};
// Set `ndarray.ndims`
ndarray.store_ndims(ctx, generator, llvm_usize.const_int(ndims, false));
// Allocate `ndarray.shape` [size_t; ndims]
ndarray.create_dim_sizes(ctx, llvm_usize, ndarray.load_ndims(ctx));
/*
ndarray now:
- .ndims: initialized
- .shape: allocated but uninitialized .shape
- .data: uninitialized
*/
let llvm_usize_sizeof = ctx
.builder
.build_int_truncate_or_bit_cast(llvm_usize.size_of(), llvm_usize, "")
.unwrap();
let llvm_pdata_sizeof = ctx
.builder
.build_int_truncate_or_bit_cast(
llvm_ret_ty.element_type().size_of().unwrap(),
llvm_usize,
"",
)
.unwrap();
let llvm_elem_sizeof = ctx
.builder
.build_int_truncate_or_bit_cast(llvm_elem_ty.size_of().unwrap(), llvm_usize, "")
.unwrap();
// Allocates a buffer for the initial RPC'ed object, which is guaranteed to be
// (4 + 4 * ndims) bytes with 8-byte alignment
let sizeof_dims =
ctx.builder.build_int_mul(ndarray.load_ndims(ctx), llvm_usize_sizeof, "").unwrap();
let unaligned_buffer_size =
ctx.builder.build_int_add(sizeof_dims, llvm_pdata_sizeof, "").unwrap();
let buffer_size = round_up(ctx, unaligned_buffer_size, llvm_usize.const_int(8, false));
let stackptr = call_stacksave(ctx, None);
// Just to be absolutely sure, alloca in [i8 x 8] slices to force 8-byte alignment
let buffer = ctx
.builder
.build_array_alloca(
llvm_i8_8,
ctx.builder
.build_int_unsigned_div(buffer_size, llvm_usize.const_int(8, false), "")
.unwrap(),
"rpc.buffer",
)
.unwrap();
let buffer = ctx
.builder
.build_bitcast(buffer, llvm_pi8, "")
.map(BasicValueEnum::into_pointer_value)
.unwrap();
let buffer = ArraySliceValue::from_ptr_val(buffer, buffer_size, None);
// The first call to `rpc_recv` reads the top-level ndarray object: [pdata, shape]
//
// The returned value is the number of bytes for `ndarray.data`.
let ndarray_nbytes = ctx
.build_call_or_invoke(
rpc_recv,
&[buffer.base_ptr(ctx, generator).into()], // Reads [usize; ndims]. NOTE: We are allocated [size_t; ndims].
"rpc.size.next",
)
.map(BasicValueEnum::into_int_value)
.unwrap();
// debug_assert(ndarray_nbytes > 0)
if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
ctx.make_assert(
generator,
ctx.builder
.build_int_compare(
IntPredicate::UGT,
ndarray_nbytes,
ndarray_nbytes.get_type().const_zero(),
"",
)
.unwrap(),
"0:AssertionError",
"Unexpected RPC termination for ndarray - Expected data buffer next",
[None, None, None],
ctx.current_loc,
);
}
// Copy shape from the buffer to `ndarray.shape`.
let pbuffer_dims =
unsafe { buffer.ptr_offset_unchecked(ctx, generator, &llvm_pdata_sizeof, None) };
call_memcpy_generic(
ctx,
ndarray.dim_sizes().base_ptr(ctx, generator),
pbuffer_dims,
sizeof_dims,
llvm_i1.const_zero(),
);
// Restore stack from before allocation of buffer
call_stackrestore(ctx, stackptr);
// Allocate `ndarray.data`.
// `ndarray.shape` must be initialized beforehand in this implementation
// (for ndarray.create_data() to know how many elements to allocate)
let num_elements =
call_ndarray_calc_size(generator, ctx, &ndarray.dim_sizes(), (None, None));
// debug_assert(nelems * sizeof(T) >= ndarray_nbytes)
if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
let sizeof_data =
ctx.builder.build_int_mul(num_elements, llvm_elem_sizeof, "").unwrap();
ctx.make_assert(
generator,
ctx.builder.build_int_compare(IntPredicate::UGE,
sizeof_data,
ndarray_nbytes,
"",
).unwrap(),
"0:AssertionError",
"Unexpected allocation size request for ndarray data - Expected up to {0} bytes, got {1} bytes",
[Some(sizeof_data), Some(ndarray_nbytes), None],
ctx.current_loc,
);
}
ndarray.create_data(ctx, llvm_elem_ty, num_elements);
let ndarray_data = ndarray.data().base_ptr(ctx, generator);
let ndarray_data_i8 =
ctx.builder.build_pointer_cast(ndarray_data, llvm_pi8, "").unwrap();
// NOTE: Currently on `prehead_bb`
ctx.builder.build_unconditional_branch(head_bb).unwrap();
// Inserting into `head_bb`. Do `rpc_recv` for `data` recursively.
ctx.builder.position_at_end(head_bb);
let phi = ctx.builder.build_phi(llvm_pi8, "rpc.ptr").unwrap();
phi.add_incoming(&[(&ndarray_data_i8, prehead_bb)]);
let alloc_size = ctx
.build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
.map(BasicValueEnum::into_int_value)
.unwrap();
let is_done = ctx
.builder
.build_int_compare(IntPredicate::EQ, llvm_i32.const_zero(), alloc_size, "rpc.done")
.unwrap();
ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
ctx.builder.position_at_end(alloc_bb);
// Align the allocation to sizeof(T)
let alloc_size = round_up(ctx, alloc_size, llvm_elem_sizeof);
let alloc_ptr = ctx
.builder
.build_array_alloca(
llvm_elem_ty,
ctx.builder.build_int_unsigned_div(alloc_size, llvm_elem_sizeof, "").unwrap(),
"rpc.alloc",
)
.unwrap();
let alloc_ptr =
ctx.builder.build_pointer_cast(alloc_ptr, llvm_pi8, "rpc.alloc.ptr").unwrap();
phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
ctx.builder.build_unconditional_branch(head_bb).unwrap();
ctx.builder.position_at_end(tail_bb);
ndarray.as_base_value().into()
}
_ => {
let slot = ctx.builder.build_alloca(llvm_ret_ty, "rpc.ret.slot").unwrap();
let slotgen = ctx.builder.build_bitcast(slot, llvm_pi8, "rpc.ret.ptr").unwrap();
ctx.builder.build_unconditional_branch(head_bb).unwrap();
ctx.builder.position_at_end(head_bb);
let phi = ctx.builder.build_phi(llvm_pi8, "rpc.ptr").unwrap();
phi.add_incoming(&[(&slotgen, prehead_bb)]);
let alloc_size = ctx
.build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
.unwrap()
.into_int_value();
let is_done = ctx
.builder
.build_int_compare(IntPredicate::EQ, llvm_i32.const_zero(), alloc_size, "rpc.done")
.unwrap();
ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
ctx.builder.position_at_end(alloc_bb);
let alloc_ptr =
ctx.builder.build_array_alloca(llvm_pi8, alloc_size, "rpc.alloc").unwrap();
let alloc_ptr =
ctx.builder.build_bitcast(alloc_ptr, llvm_pi8, "rpc.alloc.ptr").unwrap();
phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
ctx.builder.build_unconditional_branch(head_bb).unwrap();
ctx.builder.position_at_end(tail_bb);
ctx.builder.build_load(slot, "rpc.result").unwrap()
}
};
Some(result)
}
fn rpc_codegen_callback_fn<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: Option<(Type, ValueEnum<'ctx>)>,
@ -663,63 +953,14 @@ fn rpc_codegen_callback_fn<'ctx>(
// reclaim stack space used by arguments
call_stackrestore(ctx, stackptr);
// -- receive value:
// T result = {
// void *ret_ptr = alloca(sizeof(T));
// void *ptr = ret_ptr;
// loop: int size = rpc_recv(ptr);
// // Non-zero: Provide `size` bytes of extra storage for variable-length data.
// if(size) { ptr = alloca(size); goto loop; }
// else *(T*)ret_ptr
// }
let rpc_recv = ctx.module.get_function("rpc_recv").unwrap_or_else(|| {
ctx.module.add_function("rpc_recv", int32.fn_type(&[ptr_type.into()], false), None)
});
let result = format_rpc_ret(generator, ctx, fun.0.ret);
if ctx.unifier.unioned(fun.0.ret, ctx.primitives.none) {
ctx.build_call_or_invoke(rpc_recv, &[ptr_type.const_null().into()], "rpc_recv");
return Ok(None);
}
let prehead_bb = ctx.builder.get_insert_block().unwrap();
let current_function = prehead_bb.get_parent().unwrap();
let head_bb = ctx.ctx.append_basic_block(current_function, "rpc.head");
let alloc_bb = ctx.ctx.append_basic_block(current_function, "rpc.continue");
let tail_bb = ctx.ctx.append_basic_block(current_function, "rpc.tail");
let ret_ty = ctx.get_llvm_abi_type(generator, fun.0.ret);
let need_load = !ret_ty.is_pointer_type();
let slot = ctx.builder.build_alloca(ret_ty, "rpc.ret.slot").unwrap();
let slotgen = ctx.builder.build_bitcast(slot, ptr_type, "rpc.ret.ptr").unwrap();
ctx.builder.build_unconditional_branch(head_bb).unwrap();
ctx.builder.position_at_end(head_bb);
let phi = ctx.builder.build_phi(ptr_type, "rpc.ptr").unwrap();
phi.add_incoming(&[(&slotgen, prehead_bb)]);
let alloc_size = ctx
.build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
.unwrap()
.into_int_value();
let is_done = ctx
.builder
.build_int_compare(inkwell::IntPredicate::EQ, int32.const_zero(), alloc_size, "rpc.done")
.unwrap();
ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
ctx.builder.position_at_end(alloc_bb);
let alloc_ptr = ctx.builder.build_array_alloca(ptr_type, alloc_size, "rpc.alloc").unwrap();
let alloc_ptr = ctx.builder.build_bitcast(alloc_ptr, ptr_type, "rpc.alloc.ptr").unwrap();
phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
ctx.builder.build_unconditional_branch(head_bb).unwrap();
ctx.builder.position_at_end(tail_bb);
let result = ctx.builder.build_load(slot, "rpc.result").unwrap();
if need_load {
if !result.is_some_and(|res| res.get_type().is_pointer_type()) {
// An RPC returning an NDArray would not touch here.
call_stackrestore(ctx, stackptr);
}
Ok(Some(result))
Ok(result)
}
pub fn attributes_writeback(

View File

@ -1404,7 +1404,7 @@ impl<'ctx> NDArrayValue<'ctx> {
/// Returns the double-indirection pointer to the `data` array, as if by calling `getelementptr`
/// on the field.
fn ptr_to_data(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
pub fn ptr_to_data(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
let llvm_i32 = ctx.ctx.i32_type();
let var_name = self.name.map(|v| format!("{v}.data.addr")).unwrap_or_default();