onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation
/*******************************************************************************
* Copyright 2022 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#include "gpu/intel/jit/pass/pass.hpp"
#include "gemmstone/../../dsl/ir/pass/trace.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace jit {

class external_var_visitor_t : public scope_visitor_t {
public:
    void _visit(const var_t &obj) override {
        if (!is_expr_defined(obj)) external_vars.insert(obj);
    }

    object_eq_set_t<expr_t> external_vars;
};

stmt_t inject_external_var_let(const stmt_t &_stmt, ir_context_t &ir_ctx) {
    ir::trace_start();
    auto stmt = _stmt;
    external_var_visitor_t v;
    v.visit(stmt);

    std::vector<expr_t> external_vars(
            v.external_vars.begin(), v.external_vars.end());
    std::sort(external_vars.begin(), external_vars.end(),
            [&](const expr_t &a, const expr_t &b) {
        return a.as<var_t>().name < b.as<var_t>().name;
    });
    for (auto &var : external_vars)
        stmt = let_t::make(var, {}, stmt);

    ir::trace_pass("inject_external_var_let", stmt, ir_ctx);
    return stmt;
}

class store_splitter_t : public ir_mutator_t {
public:
    store_splitter_t(const dsl::hw_t &hw) : hw_(hw) {}

    object_t _mutate(const store_t &obj) override {
        int elems = obj.value.type().elems();
        int elem_size = obj.value.type().base().size();
        int stride = (obj.has_default_stride() ? 1 : obj.stride / elem_size);
        int store_size = elem_size * stride * elems;
        const auto grf_size = hw_.grf_size();
        if (store_size <= 2 * grf_size) return ir_mutator_t::_mutate(obj);

        int step = 2 * grf_size / (stride * elem_size);
        stmt_t new_stmt;
        for (int i = 0; i < elems; i += step) {
            int cur_elems = std::min(step, elems - i);
            gpu_assert(math::is_pow2(cur_elems));
            int off = i * stride * elem_size;
            auto store = store_t::make(obj.buf, obj.off + off,
                    split_expr(obj.value, i, i + cur_elems), obj.stride);
            new_stmt = new_stmt.append(store);
        }
        return std::move(new_stmt);
    }

private:
    static expr_t split_expr(const expr_t &e, int beg, int end) {
        auto *shuffle = e.as_ptr<shuffle_t>();
        if (shuffle) return shuffle_t::make(shuffle, beg, end);

        auto *binary = e.as_ptr<binary_op_t>();
        if (binary) {
            auto a = split_expr(binary->a, beg, end);
            auto b = split_expr(binary->b, beg, end);
            return binary_op_t::make(binary->op_kind, a, b);
        }
        auto *load = e.as_ptr<load_t>();
        if (load) {
            int stride = load->stride;
            if (load->has_default_stride()) stride = load->type.base().size();
            return load_t::make(load->type.with_elems(end - beg), load->buf,
                    load->off + beg * stride, load->stride);
        }
        gpu_error_not_expected();
        return expr_t();
    }

    dsl::hw_t hw_;
};

stmt_t split_wide_stores(const stmt_t &s, ir_context_t &ir_ctx) {
    ir::trace_start();
    auto ret = store_splitter_t(ir_ctx.hw()).mutate(s);
    ir::trace_pass("split_wide_stores", ret, ir_ctx);
    return ret;
}

class if_condition_fixer_t : public ir_mutator_t {
public:
    if_condition_fixer_t(int simd_size)
        : simd_size_(simd_size), in_cond_(false) {}

    object_t _mutate(const if_t &obj) override {
        auto _new_obj = ir_mutator_t::_mutate(obj);
        auto &new_obj = _new_obj.as<if_t>();
        flag_setter_t in_cond(&in_cond_, true);
        auto cond = mutate(new_obj.cond);
        return if_t::make(cond, new_obj.body, new_obj.else_body);
    }

    object_t _mutate(const binary_op_t &obj) override {
        if (!in_cond_) return obj;
        auto broadcast = [&](const expr_t &operand) {
            object_t ret;
            if (is_cmp_op(obj.op_kind) && obj.type.elems() == 1)
                ret = shuffle_t::make_broadcast(operand, simd_size_);
            else
                ret = mutate(operand);
            return ret;
        };
        auto a = broadcast(obj.a);
        auto b = broadcast(obj.b);
        return binary_op_t::make(obj.op_kind, a, b);
    }

private:
    struct flag_setter_t {
        flag_setter_t(bool *flag, bool value) : flag(flag), old(*flag) {
            *flag = value;
        }
        ~flag_setter_t() { *flag = old; }

        bool *flag;
        bool old;
    };

    int simd_size_;
    bool in_cond_;
};

stmt_t fixup_if_conditions(const stmt_t &s, ir_context_t &ir_ctx) {
    ir::trace_start();
    auto ret = if_condition_fixer_t(ir_ctx.options().simd()).mutate(s);
    ir::trace_pass("fixup_if_conditions", ret, ir_ctx);
    return ret;
}

class int64_expr_optimizer_t : public ir_mutator_t {
public:
#define HANDLE_IR_OBJECT(type) \
    object_t _mutate(const type &obj) override { return mutate_expr(obj); }

    HANDLE_EXPR_IR_OBJECTS()

#undef HANDLE_IR_OBJECT

private:
    template <typename T>
    object_t mutate_expr(const T &obj) {
        auto new_obj = ir_mutator_t::_mutate(obj);
        if (auto *binary = new_obj.template as_ptr<binary_op_t>()) {
            if (binary->op_kind == op_kind_t::_add) {
                new_obj = simplify_64_bit_add(new_obj);
            }
        }
        return new_obj;
    }
};

stmt_t optimize_int64_exprs(const stmt_t &s, ir_context_t &ir_ctx) {
    ir::trace_start();
    auto ret = int64_expr_optimizer_t().mutate(s);
    ir::trace_pass("optimize_int64_exprs", ret, ir_ctx);
    return ret;
}

} // namespace jit
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl