[CODEGEN] Various bugfixes that make it possible to fuse RNG in a matmul epilogue (#356)

This commit is contained in:
Philippe Tillet
2021-10-24 02:30:46 -07:00
committed by GitHub
parent 858dec8372
commit 5ce1b726dc
17 changed files with 149 additions and 60 deletions

View File

@@ -147,12 +147,14 @@ public:
void visit_store_inst(ir::store_inst*);
void visit_unmasked_store_inst(ir::unmasked_store_inst*);
void visit_masked_store_inst(ir::masked_store_inst*);
void visit_cat_inst(ir::cat_inst*);
void visit_reshape_inst(ir::reshape_inst*);
void visit_splat_inst(ir::splat_inst*);
void visit_broadcast_inst(ir::broadcast_inst*);
void visit_downcast_inst(ir::downcast_inst*);
void visit_exp_inst(ir::exp_inst*);
void visit_cos_inst(ir::cos_inst*);
void visit_umulhi_inst(ir::umulhi_inst* x);
void visit_sin_inst(ir::sin_inst*);
void visit_log_inst(ir::log_inst*);
void visit_get_program_id_inst(ir::get_program_id_inst*);

View File

@@ -137,6 +137,7 @@ public:
// Block instruction
value *create_splat(value *arg, const type::block_shapes_t &shapes);
value *create_reshape(value *arg, const type::block_shapes_t &shapes);
value *create_cat(value *lhs, value *rhs);
value *create_broadcast(value *arg, const type::block_shapes_t &shapes);
// Built-in instruction
value *create_get_program_id(unsigned axis);
@@ -153,6 +154,8 @@ public:
value *create_reduce(value *A, reduce_inst::op_t op, unsigned axis);
value *create_select(value *pred, value *if_value, value *else_value);
// Intrinsics
// These have no place in the IR, and hopefully they can be removed at some point
value *create_umulhi(value* lhs, value* rhs);
value *create_copy_to_shared(value *arg);
value *create_masked_load_async(value *arg, value *mask, value *false_value, load_inst::CACHE_MODIFIER cache);
value *create_copy_from_shared(value *arg);

View File

@@ -61,6 +61,7 @@ struct dispatch{
// casting ops
static ir::value *reshape(ir::value *input, shape_t shape, ir::builder *builder);
static ir::value *cat(ir::value *lhs, ir::value *rhs, ir::builder *builder);
static ir::value *broadcast(ir::value *input, shape_t shape, ir::builder *builder);
static std::tuple<ir::value*, ir::value*> broadcast(ir::value *lhs, ir::value* rhs, ir::builder *builder);
static ir::value *bitcast(ir::value *input, ir::type *type, ir::builder *builder);
@@ -90,6 +91,7 @@ struct dispatch{
static ir::value *sum(ir::value *input, unsigned int axis, ir::builder *builder);
// math
static ir::value *umulhi(ir::value *x, ir::value *y, ir::builder *builder);
static ir::value *exp(ir::value *x, ir::builder *builder);
static ir::value *log(ir::value *x, ir::builder *builder);
static ir::value *cos(ir::value *x, ir::builder *builder);

View File

@@ -132,6 +132,7 @@ enum value_id_t: unsigned {
// retile
INST_RESHAPE,
INST_SPLAT,
INST_CAT,
INST_BROADCAST,
INST_DOWNCAST,
// builtin
@@ -142,6 +143,7 @@ enum value_id_t: unsigned {
INST_ATOMIC_EXCH,
INST_ATOMIC_RMW,
// math
INST_UMULHI,
INST_EXP,
INST_COS,
INST_SIN,

View File

@@ -520,6 +520,21 @@ public:
// retile_inst classes
//===----------------------------------------------------------------------===//
// cat
class cat_inst: public instruction {
private:
std::string repr_impl() const { return "cat"; }
cat_inst(value *x, value *y, const std::string &name, instruction *next);
public:
static instruction* create(value *lhs, value *rhs,
const std::string &name = "",
instruction *next = nullptr);
_TRITON_DEFINE_CLONE(cat_inst)
_TRITON_DEFINE_ACCEPT(cat_inst)
};
// retile
class retile_inst: public unary_inst {
@@ -654,6 +669,17 @@ public:
static instruction* create(value *ptr, value *cmp, value *val, const std::string &name = "", instruction *next = nullptr);
};
class umulhi_inst: public builtin_inst {
private:
umulhi_inst(value *lhs, value *rhs, const std::string &name = "", instruction *next = nullptr);
std::string repr_impl() const { return "umulhi"; }
_TRITON_DEFINE_CLONE(umulhi_inst)
_TRITON_DEFINE_ACCEPT(umulhi_inst)
public:
static instruction* create(value *lhs, value *rhs, const std::string &name = "", instruction *next = nullptr);
};
class exp_inst: public builtin_inst {
private:
exp_inst(value *val, const std::string &name = "", instruction *next = nullptr);
@@ -803,6 +829,7 @@ public:
// intrinsics classes
//===----------------------------------------------------------------------===//
class copy_to_shared_inst: public unary_inst{
private:
using unary_inst::unary_inst;
@@ -884,35 +911,6 @@ public:
instruction *next=nullptr);
};
//// On NVIDIA, implementation is such that
//// constant_range = nv_dynamic_program_idx + nv_static_program_idx
//// so as to enable re-association on nv_static_program_idx which is constant
//class make_range_dyn: public instruction {
//private:
// make_range_dyn(type *ty, const std::string &name, instruction *next);
// std::string repr_impl() const { return "nv_dynamic_program_idx"; }
// _TRITON_DEFINE_CLONE(make_range_dyn)
// _TRITON_DEFINE_ACCEPT(make_range_dyn)
//public:
// static make_range_dyn* create(type *ty, const std::string &name = "", instruction *next = nullptr);
//};
//class make_range_sta: public constant {
//private:
// make_range_sta(make_range *range);
//public:
// static make_range_sta *get(make_range* range);
// make_range* get_range() const;
// std::string repr() const { return "nv_static_program_idx"; }
// _TRITON_DEFINE_ACCEPT(make_range_sta)
//private:
// make_range *range_;
//};
/* constant range */
class make_range: public instruction{
make_range(type *ty, constant_int* first, constant_int* last);

View File

@@ -45,9 +45,11 @@ class masked_store_inst;
class retile_inst;
class reshape_inst;
class splat_inst;
class cat_inst;
class broadcast_inst;
class downcast_inst;
class umulhi_inst;
class exp_inst;
class cos_inst;
class sin_inst;
@@ -122,6 +124,7 @@ public:
virtual void visit_unmasked_store_inst(unmasked_store_inst*) = 0;
virtual void visit_masked_store_inst(masked_store_inst*) = 0;
virtual void visit_umulhi_inst(umulhi_inst*) = 0;
virtual void visit_exp_inst(exp_inst*) = 0;
virtual void visit_cos_inst(cos_inst*) = 0;
virtual void visit_sin_inst(sin_inst*) = 0;
@@ -129,6 +132,7 @@ public:
virtual void visit_reshape_inst(reshape_inst*) = 0;
virtual void visit_splat_inst(splat_inst*) = 0;
virtual void visit_cat_inst(cat_inst*) = 0;
virtual void visit_broadcast_inst(broadcast_inst*) = 0;
virtual void visit_downcast_inst(downcast_inst*) = 0;
@@ -150,13 +154,10 @@ public:
virtual void visit_masked_load_async_inst(masked_load_async_inst*)= 0;
virtual void visit_barrier_inst(barrier_inst*) = 0;
virtual void visit_async_wait_inst(async_wait_inst*) = 0;
// virtual void visit_make_range_dyn(make_range_dyn*) = 0;
virtual void visit_make_range(make_range*) = 0;
virtual void visit_prefetch_s_inst(prefetch_s_inst*) = 0;
virtual void visit_function(function*) = 0;
// virtual void visit_make_range_sta(make_range_sta*) = 0;
virtual void visit_undef_value(undef_value*) = 0;
virtual void visit_constant_int(constant_int*) = 0;
virtual void visit_constant_fp(constant_fp*) = 0;