https://github.com/halide/Halide
Tip revision: 5919b14eea78c52df561d738a1f7c51be7fd0fdf authored by Derek Gerstmann on 09 February 2024, 22:57:21 UTC
Formatting pass
Formatting pass
Tip revision: 5919b14
Prefetch.cpp
#include <map>
#include <string>
#include <utility>
#include "Bounds.h"
#include "ExprUsesVar.h"
#include "Function.h"
#include "IRMutator.h"
#include "IROperator.h"
#include "Prefetch.h"
#include "Scope.h"
#include "Simplify.h"
#include "Target.h"
#include "Util.h"
namespace Halide {
namespace Internal {
using std::map;
using std::set;
using std::string;
using std::vector;
/**
* The steps by which prefetch directives are injected and lowered are a bit nonobvious;
* here's the overall flow at the time this comment was written:
*
* - When the .prefetch() schedule directive is used, a PrefetchDirective is
* added to the relevant Function
* - At the start of lowering (schedule_functions()), a placeholder Prefetch IR nodes (w/ no region) are inserted
* - Various lowering passes mutate the placeholder Prefetch IR nodes as appropriate
* - After storage folding, the Prefetch IR nodes are updated with a proper region (via inject_prefetch())
* - In storage_flattening(), Prefetch IR nodes transformed to Call::prefetch intrinsics:
* - These calls are always wrapped in Evaluate IR Nodes
* - The Evaluate IR nodes are, in turn, possibly in IfTheElse (if condition != const_true())
* - After this point, no Prefetch IR nodes should remain in the IR (only prefetch intrinsics)
* - reduce_prefetch_dimension() is later called to reduce prefetch the dimensionality of the prefetch intrinsic.
*/
namespace {
// Collect the bounds of all the externally referenced buffers in a stmt.
class CollectExternalBufferBounds : public IRVisitor {
public:
map<string, Box> buffers;
using IRVisitor::visit;
void add_buffer_bounds(const string &name, const Buffer<> &image, const Parameter ¶m, int dims) {
Box b;
for (int i = 0; i < dims; ++i) {
string dim_name = std::to_string(i);
Expr buf_min_i = Variable::make(Int(32), name + ".min." + dim_name,
image, param, ReductionDomain());
Expr buf_extent_i = Variable::make(Int(32), name + ".extent." + dim_name,
image, param, ReductionDomain());
Expr buf_max_i = buf_min_i + buf_extent_i - 1;
b.push_back(Interval(buf_min_i, buf_max_i));
}
buffers.emplace(name, b);
}
void visit(const Call *op) override {
IRVisitor::visit(op);
add_buffer_bounds(op->name, op->image, op->param, (int)op->args.size());
}
void visit(const Variable *op) override {
if (op->param.defined() && op->param.is_buffer()) {
add_buffer_bounds(op->name, Buffer<>(), op->param, op->param.dimensions());
}
}
};
class InjectPrefetch : public IRMutator {
public:
InjectPrefetch(const map<string, Function> &e, const map<string, Box> &buffers)
: env(e), external_buffers(buffers) {
}
private:
const map<string, Function> &env;
const map<string, Box> &external_buffers;
Scope<Box> buffer_bounds;
using IRMutator::visit;
Box get_buffer_bounds(const string &name, int dims) {
if (buffer_bounds.contains(name)) {
const Box &b = buffer_bounds.ref(name);
internal_assert((int)b.size() == dims);
return b;
}
// It is an external buffer.
user_assert(env.find(name) == env.end())
<< "Prefetch to buffer \"" << name << "\" which has not been allocated\n";
const auto &iter = external_buffers.find(name);
internal_assert(iter != external_buffers.end());
return iter->second;
}
Stmt visit(const Realize *op) override {
Box b;
b.used = op->condition;
for (const auto &r : op->bounds) {
b.push_back(Interval(r.min, r.min + r.extent - 1));
}
ScopedBinding<Box> bind(buffer_bounds, op->name, b);
return IRMutator::visit(op);
}
Stmt visit(const Prefetch *op) override {
Stmt body = mutate(op->body);
const PrefetchDirective &p = op->prefetch;
Expr at = Variable::make(Int(32), p.at);
Expr from = Variable::make(Int(32), p.from);
// Add loop variable + prefetch offset to interval scope for box computation
Expr fetch_at = from + p.offset;
map<string, Box> boxes_rw = boxes_touched(LetStmt::make(p.from, fetch_at, body));
// TODO(psuriana): Only prefetch the newly accessed data. We
// should subtract the box accessed during previous iteration
// from the one accessed during this iteration.
// TODO(psuriana): Add a new PrefetchBoundStrategy::ShiftInwards
// that shifts the base address of the prefetched box so that
// the box is completely within the bounds.
const auto &b = boxes_rw.find(p.name);
if (b != boxes_rw.end()) {
Box prefetch_box = b->second;
// Only prefetch the region that is in bounds.
Box bounds = get_buffer_bounds(b->first, b->second.size());
internal_assert(prefetch_box.size() == bounds.size());
if (p.strategy == PrefetchBoundStrategy::Clamp) {
prefetch_box = box_intersection(prefetch_box, bounds);
} else if (p.strategy == PrefetchBoundStrategy::GuardWithIf) {
Expr predicate = prefetch_box.used.defined() ? prefetch_box.used : const_true();
for (size_t i = 0; i < bounds.size(); ++i) {
predicate = predicate && (prefetch_box[i].min >= bounds[i].min) &&
(prefetch_box[i].max <= bounds[i].max);
}
prefetch_box.used = simplify(predicate);
} else {
internal_assert(p.strategy == PrefetchBoundStrategy::NonFaulting);
// Assume the prefetch won't fault when accessing region
// outside the bounds.
}
// Construct the region to be prefetched.
Region new_bounds;
for (size_t i = 0; i < prefetch_box.size(); i++) {
Expr extent = prefetch_box[i].max - prefetch_box[i].min + 1;
new_bounds.emplace_back(simplify(prefetch_box[i].min), simplify(extent));
}
Expr condition = op->condition;
if (prefetch_box.maybe_unused()) {
condition = simplify(prefetch_box.used && condition);
}
internal_assert(!new_bounds.empty());
return Prefetch::make(op->name, op->types, new_bounds, op->prefetch, std::move(condition), std::move(body));
}
if (!body.same_as(op->body)) {
return Prefetch::make(op->name, op->types, op->bounds, op->prefetch, op->condition, std::move(body));
} else if (op->bounds.empty()) {
// Remove the Prefetch IR since it is prefetching an empty region
user_warning << "Removing prefetch of " << p.name
<< " at loop nest of " << p.at
<< " from location " << p.from
<< " + offset " << p.offset
<< ") since the prefetched area will always be empty.\n";
return body;
} else {
return op;
}
}
};
class InjectPlaceholderPrefetch : public IRMutator {
public:
InjectPlaceholderPrefetch(const map<string, Function> &e, const string &prefix,
const vector<PrefetchDirective> &prefetches)
: env(e), prefix(prefix), prefetch_list(prefetches) {
}
private:
const map<string, Function> &env;
const string &prefix;
const vector<PrefetchDirective> &prefetch_list;
std::vector<string> loop_nest;
using IRMutator::visit;
Stmt add_placeholder_prefetch(const string &at, const string &from, PrefetchDirective p, Stmt body) {
debug(5) << "...Injecting placeholder prefetch for loop " << at << " from " << from << "\n";
p.at = at;
p.from = from;
internal_assert(body.defined());
if (p.param.defined()) {
return Prefetch::make(p.name, {p.param.type()}, Region(), p, const_true(), std::move(body));
} else {
const auto &it = env.find(p.name);
internal_assert(it != env.end());
return Prefetch::make(p.name, it->second.output_types(), Region(), p, const_true(), std::move(body));
}
}
Stmt visit(const For *op) override {
loop_nest.push_back(op->name);
Stmt body = mutate(op->body);
if (!prefetch_list.empty() && starts_with(op->name, prefix)) {
// If there are multiple prefetches of the same Func or ImageParam,
// use the most recent one
set<string> seen;
for (int i = prefetch_list.size() - 1; i >= 0; --i) {
const PrefetchDirective &p = prefetch_list[i];
if (!ends_with(op->name, "." + p.at) || (seen.find(p.name) != seen.end())) {
continue;
}
seen.insert(p.name);
// We pass op->name for the prefetch 'at', so that should always be a fully-qualified loop variable name
// at this point; however, 'from' will be just the left-name of the loop var and must be qualified further.
// Look through the loop_nest list to find the most recent loop that starts with 'prefix' and ends with 'from'.
// Note that it is not good enough to just prepend use 'prefix + from', as there may be splits involved, e.g.,
// prefix = g.s0, from = xo, but the var we seek is actually g.s0.x.xo (because 'g' was split at x).
string from_var;
for (int j = (int)loop_nest.size() - 1; j >= 0; --j) {
if (starts_with(loop_nest[j], prefix) && ends_with(loop_nest[j], "." + p.from)) {
from_var = loop_nest[j];
debug(5) << "Prefetch from " << p.from << " -> from_var " << from_var << "\n";
break;
}
}
if (from_var.empty()) {
user_error << "Prefetch 'from' variable '" << p.from << "' could not be found in an active loop. (Are the 'at' and 'from' variables swapped?)";
}
body = add_placeholder_prefetch(op->name, from_var, p, std::move(body));
}
}
Stmt stmt;
if (!body.same_as(op->body)) {
stmt = For::make(op->name, op->min, op->extent, op->for_type, op->partition_policy, op->device_api, std::move(body));
} else {
stmt = op;
}
internal_assert(loop_nest.back() == op->name);
loop_nest.pop_back();
return stmt;
}
};
// Reduce the prefetch dimension if bigger than 'max_dim'. It keeps the 'max_dim'
// innermost dimensions and replaces the rests with for-loops.
class ReducePrefetchDimension : public IRMutator {
using IRMutator::visit;
const size_t max_dim;
Stmt visit(const Evaluate *op) override {
Stmt stmt = IRMutator::visit(op);
op = stmt.as<Evaluate>();
internal_assert(op);
const Call *prefetch = Call::as_intrinsic(op->value, {Call::prefetch});
// TODO(psuriana): Ideally, we want to keep the loop size minimal to
// minimize the number of prefetch calls. We probably want to lift
// the dimensions with larger strides and keep the smaller ones in
// the prefetch call.
const size_t max_arg_size = 2 + 2 * max_dim; // Prefetch: {base, offset, extent0, stride0, extent1, stride1, ...}
if (prefetch && (prefetch->args.size() > max_arg_size)) {
const Expr &base_address = prefetch->args[0];
const Expr &base_offset = prefetch->args[1];
const Variable *base = base_address.as<Variable>();
internal_assert(base && base->type.is_handle());
vector<string> index_names;
Expr new_offset = base_offset;
for (size_t i = max_arg_size; i < prefetch->args.size(); i += 2) {
// const Expr &extent = prefetch->args[i + 0]; // unused
const Expr &stride = prefetch->args[i + 1];
string index_name = "prefetch_reduce_" + base->name + "." + std::to_string((i - 1) / 2);
index_names.push_back(index_name);
new_offset += Variable::make(Int(32), index_name) * stride;
}
vector<Expr> args = {base, new_offset};
for (size_t i = 2; i < max_arg_size; ++i) {
args.push_back(prefetch->args[i]);
}
stmt = Evaluate::make(Call::make(prefetch->type, Call::prefetch, args, Call::Intrinsic));
for (size_t i = 0; i < index_names.size(); ++i) {
stmt = For::make(index_names[i], 0, prefetch->args[(i + max_dim) * 2 + 2],
ForType::Serial, Partition::Auto, DeviceAPI::None, stmt);
}
debug(5) << "\nReduce prefetch to " << max_dim << " dim:\n"
<< "Before:\n"
<< Expr(prefetch) << "\nAfter:\n"
<< stmt << "\n";
}
return stmt;
}
public:
ReducePrefetchDimension(size_t dim)
: max_dim(dim) {
}
};
// If the prefetched data is larger than 'max_byte_size', we need to tile the
// prefetch. This will split the prefetch call into multiple calls by adding
// an outer for-loop around the prefetch.
class SplitPrefetch : public IRMutator {
using IRMutator::visit;
Expr max_byte_size;
Stmt visit(const Evaluate *op) override {
Stmt stmt = IRMutator::visit(op);
op = stmt.as<Evaluate>();
internal_assert(op);
if (const Call *prefetch = Call::as_intrinsic(op->value, {Call::prefetch})) {
const Expr &base_address = prefetch->args[0];
const Expr &base_offset = prefetch->args[1];
const Variable *base = base_address.as<Variable>();
internal_assert(base && base->type.is_handle());
int elem_size = prefetch->type.bytes();
vector<string> index_names;
vector<Expr> extents;
Expr new_offset = base_offset;
for (size_t i = 2; i < prefetch->args.size(); i += 2) {
Expr extent = prefetch->args[i];
Expr stride = prefetch->args[i + 1];
Expr stride_bytes = stride * elem_size;
string index_name = "prefetch_split_" + base->name + "." + std::to_string((i - 1) / 2);
index_names.push_back(index_name);
Expr is_negative_stride = (stride < 0);
Expr outer_var = Variable::make(Int(32), index_name);
Expr outer_extent;
if (can_prove(max_byte_size < stride_bytes) || can_prove(max_byte_size < -stride_bytes)) {
// If 'max_byte_size' is smaller than the absolute value of the
// stride bytes, we can only prefetch one element per iteration.
outer_extent = extent;
new_offset += outer_var * stride_bytes;
} else {
// Otherwise, we just prefetch 'max_byte_size' per iteration.
Expr abs_stride_bytes = Call::make(stride_bytes.type(), Call::abs, {stride_bytes}, Call::PureIntrinsic);
outer_extent = simplify((extent * abs_stride_bytes + max_byte_size - 1) / max_byte_size);
new_offset += outer_var * simplify(select(is_negative_stride, -max_byte_size, max_byte_size));
}
extents.push_back(outer_extent);
}
Expr new_extent = 1;
Expr new_stride = simplify(max_byte_size / elem_size);
vector<Expr> args = {base, std::move(new_offset), std::move(new_extent), std::move(new_stride)};
stmt = Evaluate::make(Call::make(prefetch->type, Call::prefetch, args, Call::Intrinsic));
for (size_t i = 0; i < index_names.size(); ++i) {
stmt = For::make(index_names[i], 0, extents[i],
ForType::Serial, Partition::Auto, DeviceAPI::None, stmt);
}
debug(5) << "\nSplit prefetch to max of " << max_byte_size << " bytes:\n"
<< "Before:\n"
<< Expr(prefetch) << "\nAfter:\n"
<< stmt << "\n";
}
return stmt;
}
public:
SplitPrefetch(Expr bytes)
: max_byte_size(std::move(bytes)) {
}
};
template<typename Fn>
void traverse_block(const Stmt &s, Fn &&f) {
const Block *b = s.as<Block>();
if (!b) {
f(s);
} else {
traverse_block(b->first, f);
traverse_block(b->rest, f);
}
}
class HoistPrefetches : public IRMutator {
using IRMutator::visit;
Stmt visit(const Block *op) override {
Stmt s = op;
Stmt prefetches, body;
traverse_block(s, [this, &prefetches, &body](const Stmt &s_in) {
Stmt s = IRMutator::mutate(s_in);
const Evaluate *eval = s.as<Evaluate>();
if (eval && Call::as_intrinsic(eval->value, {Call::prefetch})) {
prefetches = prefetches.defined() ? Block::make(prefetches, s) : s;
} else {
body = body.defined() ? Block::make(body, s) : s;
}
});
if (prefetches.defined()) {
if (body.defined()) {
return Block::make(prefetches, body);
} else {
return prefetches;
}
} else {
return body;
}
}
};
} // anonymous namespace
Stmt inject_placeholder_prefetch(const Stmt &s, const map<string, Function> &env,
const string &prefix,
const vector<PrefetchDirective> &prefetches) {
Stmt stmt = InjectPlaceholderPrefetch(env, prefix, prefetches).mutate(s);
return stmt;
}
Stmt inject_prefetch(const Stmt &s, const map<string, Function> &env) {
CollectExternalBufferBounds finder;
s.accept(&finder);
return InjectPrefetch(env, finder.buffers).mutate(s);
}
Stmt reduce_prefetch_dimension(Stmt stmt, const Target &t) {
size_t max_dim = 0;
Expr max_byte_size;
// Hexagon's prefetch takes in a range of address and can be maximum of
// two dimension. Other architectures generate one prefetch per cache line.
if (t.has_feature(Target::HVX)) {
max_dim = 2;
} else if (t.arch == Target::ARM) {
// ARM's cache line size can be 32 or 64 bytes and it can switch the
// size at runtime. To be safe, we just use 32 bytes.
max_dim = 1;
max_byte_size = 32;
} else {
max_dim = 1;
max_byte_size = 64;
}
internal_assert(max_dim > 0);
stmt = ReducePrefetchDimension(max_dim).mutate(stmt);
if (max_byte_size.defined()) {
// If the max byte size is specified, we may need to tile
// the prefetch
stmt = SplitPrefetch(max_byte_size).mutate(stmt);
}
return stmt;
}
Stmt hoist_prefetches(const Stmt &s) {
return HoistPrefetches().mutate(s);
}
} // namespace Internal
} // namespace Halide