lower_map_reduce.h

Go to the documentation of this file.

#pragma once
 
#include <mim/def.h>
#include <mim/phase.h>
 
namespace mim::plug::btensor::phase {
 
/// Lowers the buffer-world operations (`btensor.map_reduce_post`, `btensor.broadcast`, `btensor.pad`,
/// `btensor.concat`, `btensor.gather`, `btensor.scatter`) into `affine.For` loop nests over
/// `buffer.read` / `buffer.write` / `buffer.alloc`, threading `mem.M`.
/// These are the buffer-world counterparts of the corresponding `tensor.*` ops; the `tensor` plugin's
/// bufferization (`tensor.lower_to_mem`) maps the SSA tensor ops onto them.
/// Also lowers `buffer.lit` into a fill loop, so a large constant/splat tensor becomes a loop rather
/// than a monolithic `mem.store` of a giant literal array (which the LLVM backend cannot digest).
class LowerMapReduce : public RWPhase {
public:
    using RWPhase::RWPhase;
 
private:
    const Def* rewrite_imm_App(const App*) override;
    const Def* lower_map_reduce_post(const App*);
    const Def* lower_broadcast(const App*);
    const Def* lower_pad(const App*);
    const Def* lower_concat(const App*);
    const Def* lower_buffer_lit(const App*);
    const Def* lower_gather(const App*);
    const Def* lower_scatter(const App*);
};
 
} // namespace mim::plug::btensor::phase