onednn-src 0.1.13

/*******************************************************************************
* Copyright 2020 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 "xpu/stream_profiler.hpp"

#include "gpu/intel/compute/types_interop.hpp"
#include "gpu/intel/engine.hpp"
#include "gpu/intel/stream.hpp"

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

status_t stream_t::zero_pad(const memory_t *memory, const exec_ctx_t &ctx) {
    memory_desc_wrapper mdw(memory->md());

    if (mdw.format_kind() != format_kind::blocked) return status::unimplemented;

    if (mdw.nelems(false) == mdw.nelems(true)) return status::success;

    if (!has_zero_pad_primitive()) return impl::stream_t::zero_pad(memory, ctx);

    // Kernel only compiled to support data types of length 1, 2, 4 or 8 currently
    if (!utils::one_of(mdw.data_type_size(), 1u, 2u, 4u, 8u))
        return status::unimplemented;

    const blocking_desc_t blocking_desc = mdw.blocking_desc();

    const int max_step_nelems = ZERO_PAD_MAX_STEP_SIZE;
    size_t step_nelems = 1;
    for (int i = 0; i < blocking_desc.inner_nblks; i++) {
        step_nelems *= blocking_desc.inner_blks[i];
    }

    assert(step_nelems <= max_step_nelems);
    if (step_nelems > max_step_nelems)
        return impl::stream_t::zero_pad(memory, ctx);

    impl::engine_t *engine = this->engine();

    primitive_t *zero_pad_primitive;
    const resource_mapper_t *mapper;
    CHECK(utils::downcast<engine_t *>(engine)->get_zero_pad_primitive(
            zero_pad_primitive, mapper));

    exec_args_t zero_pad_args;
    memory_arg_t arg = {const_cast<memory_t *>(memory), true};
    zero_pad_args[DNNL_ARG_SRC] = arg;
    exec_ctx_t zero_pad_ctx(this, std::move(zero_pad_args));
    zero_pad_ctx.set_resource_mapper(mapper);

    // Verbose is implemented separately here since fake primitive descriptor
    // contains only primitive_kind in internal op_desc, but no md. Such design
    // was chosen to avoid re-creation of zeropad primitive in case it lives as
    // a regular one in cache and may be evicted from there. It means that md
    // is available only with incoming memory at execution point here, that's
    // why separate logic is written apart from a common place.
    // XXX: re-consider, once zeropad appears in other places in the library.
    if (get_verbose(verbose_t::exec_profile)) {
        CHECK(this->wait());
        double start_ms = get_msec();
        CHECK(zero_pad_primitive->execute(zero_pad_ctx));
        status_t status = this->wait();
        double duration_ms = get_msec() - start_ms;
        std::stringstream info;
        info << "gpu,zero_pad," << zero_pad_primitive->pd()->name() << ",undef,"
             << md2fmt_str("data", memory->md(), format_kind::undef) << ",,,"
             << md2dim_str(memory->md());
        VPROF(start_ms, primitive, exec, VERBOSE_profile, info.str().c_str(),
                duration_ms);

        return status;
    } else {
        return zero_pad_primitive->execute(zero_pad_ctx);
    }
}

status_t stream_t::notify_profiling_complete() const {
    return profiler().notify_profiling_complete();
}

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