onednn-src 0.1.13

/*******************************************************************************
* Copyright 2021 Intel Corporation
* Copyright 2021-2024 FUJITSU LIMITED
* Copyright 2025 Arm Ltd. and affiliates
*
* 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 "common/c_types_map.hpp"
#include "common/dnnl_thread.hpp"
#include "common/memory_tracking.hpp"

#include "common/bfloat16.hpp"

#include "cpu/aarch64/jit_uni_dw_convolution.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace aarch64 {

using namespace dnnl::impl::status;
using namespace dnnl::impl::memory_tracking::names;
using namespace dnnl::impl::utils;

template <cpu_isa_t isa, data_type_t src_type, data_type_t dst_type>
void jit_uni_dw_convolution_fwd_t<isa, src_type, dst_type>::execute_forward(
        const exec_ctx_t &ctx) const {
    auto src = CTX_IN_MEM(const data_t *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const data_t *, DNNL_ARG_WEIGHTS);
    auto dst = CTX_OUT_MEM(dst_data_t *, DNNL_ARG_DST);

    const memory_desc_wrapper src_d(pd()->src_md());
    const memory_desc_wrapper dst_d(pd()->dst_md());
    const memory_desc_wrapper weights_d(pd()->weights_md(0));
    const memory_desc_wrapper bias_d(pd()->weights_md(1));

    const auto &jcp = pd()->jcp_;

    f32_data_t *bias = nullptr;
    if (pd()->desc()->bias_desc.data_type == data_type::bf16) {
        auto bias_in = CTX_IN_MEM(const bf16_data_t *, DNNL_ARG_BIAS);
        bias = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                key_conv_bias_bf16_convert_wsp);
        cvt_bfloat16_to_float(bias, bias_in, jcp.oc_without_padding);
        utils::array_set(bias + jcp.oc_without_padding, 0.f,
                jcp.oc - jcp.oc_without_padding);
    } else {
        auto bias_in = CTX_IN_MEM(const f32_data_t *, DNNL_ARG_BIAS);
        if (pd()->wants_padded_bias()) {
            auto padded_bias
                    = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                            key_conv_padded_bias);
            utils::array_copy(padded_bias, bias_in, jcp.oc_without_padding);
            utils::array_set(padded_bias + jcp.oc_without_padding, 0.f,
                    jcp.oc - jcp.oc_without_padding);
            bias = padded_bias;
        } else
            bias = const_cast<float *>(bias_in);
    }

    const int dil_h = jcp.dilate_h + 1;
    const int str_h = jcp.stride_h;
    const int ch_step = jcp.nb_ch_blocking;
    const int ow = 0;
    const int iw = 0;
    const int kw = 0;
    const int chb_work = utils::div_up(jcp.nb_ch, ch_step);
    const auto is_src_layout_nxc = jcp.src_tag == format_tag::nhwc;
    const auto is_dst_layout_nxc = jcp.dst_tag == format_tag::nhwc;

    const int work_amount = jcp.mb * chb_work * jcp.oh;
    const auto nthr = jcp.nthr;

    parallel(nthr, [&](const int ithr, const int nthr) {
        int start {0}, end {0};
        balance211(work_amount, nthr, ithr, start, end);

        int n {0}, chb {0}, oh {0};
        if (jcp.loop_order == loop_ngcw)
            utils::nd_iterator_init(
                    start, n, jcp.mb, chb, chb_work, oh, jcp.oh);
        else if (jcp.loop_order == loop_nhwcg)
            utils::nd_iterator_init(
                    start, n, jcp.mb, oh, jcp.oh, chb, chb_work);
        else
            assert(!"unsupported loop order");

        auto iwork = start;
        while (iwork < end) {

            int ch = chb * ch_step;

            const int i_t_overflow
                    = nstl::max(0, (int)(jcp.t_pad - oh * str_h));
            const int i_b_overflow
                    = nstl::max(jcp.ih,
                              (int)(oh * str_h + (jcp.kh - 1) * dil_h
                                      - jcp.t_pad + 1))
                    - jcp.ih;

            const int ih
                    = nstl::max((int)(oh * str_h - jcp.t_pad
                                        + div_up(i_t_overflow, dil_h) * dil_h),
                            0);
            const int kh = div_up(i_t_overflow, dil_h);
            const int kh_padding = jcp.kh - div_up(i_t_overflow, dil_h)
                    - div_up(i_b_overflow, dil_h);

            const auto ic_off_idx = is_src_layout_nxc ? ch * jcp.ch_block : ch;
            const auto oc_off_idx = is_dst_layout_nxc ? ch * jcp.ch_block : ch;

            auto par_conv = jit_conv_args_t();
            par_conv.src = jcp.is_fused_conv
                    ? src
                    : &src[src_d.blk_off(n, ic_off_idx, ih, iw)];
            par_conv.dst = &dst[dst_d.blk_off(n, oc_off_idx, oh, ow)];

            par_conv.filt = &weights[weights_d.blk_off(ch, 0, 0, kh, kw)];
            if (bias) par_conv.bias = &bias[bias_d.blk_off(ch * jcp.ch_block)];

            par_conv.kh_padding = (size_t)nstl::max(0, kh_padding);

            if (is_src_layout_nxc) {
                // maximize jit work along contiguous dimension
                int work_rem = end - iwork;
                par_conv.ch_blocks = ch + work_rem * ch_step >= jcp.nb_ch
                        ? jcp.nb_ch - ch
                        : work_rem * ch_step;
                assert(jcp.loop_order == loop_nhwcg);
            } else {
                par_conv.ch_blocks
                        = utils::this_block_size(ch, jcp.nb_ch, ch_step);
                assert(jcp.loop_order != loop_nhwcg);
            }

            (*kernel_)(&par_conv);

            if (jcp.loop_order == loop_ngcw) {
                ++iwork;
                utils::nd_iterator_step(n, jcp.mb, chb, chb_work, oh, jcp.oh);
            } else if (jcp.loop_order == loop_nhwcg) {
                utils::nd_iterator_jump(
                        iwork, end, n, jcp.mb, oh, jcp.oh, chb, chb_work);
            } else
                assert(!"unsupported loop order");
        }
    });

    if (pd()->wants_zero_pad_dst()) ctx.zero_pad_output(DNNL_ARG_DST);
}

template struct jit_uni_dw_convolution_fwd_t<sve_512, data_type::f32>;
template struct jit_uni_dw_convolution_fwd_t<sve_256, data_type::f32>;
template struct jit_uni_dw_convolution_fwd_t<sve_128, data_type::f32>;
template struct jit_uni_dw_convolution_fwd_t<sve_256, data_type::bf16>;
template struct jit_uni_dw_convolution_fwd_t<sve_128, data_type::bf16>;

template <cpu_isa_t isa, data_type_t diff_dst_type, data_type_t diff_src_type>
void jit_uni_dw_convolution_bwd_data_t<isa, diff_dst_type,
        diff_src_type>::execute_backward_data(const exec_ctx_t &ctx) const {
    auto diff_dst = CTX_IN_MEM(const diff_dst_data_t *, DNNL_ARG_DIFF_DST);
    auto weights = CTX_IN_MEM(const wei_data_t *, DNNL_ARG_WEIGHTS);
    auto diff_src = CTX_OUT_MEM(diff_src_data_t *, DNNL_ARG_DIFF_SRC);

    const memory_desc_wrapper diff_dst_d(pd()->diff_dst_md());
    const memory_desc_wrapper diff_src_d(pd()->diff_src_md());
    const memory_desc_wrapper weights_d(pd()->weights_md(0));

    const auto &jcp = pd()->jcp_;

    auto kernel_params = [&](int ur_str_w, int iw, int oh, int ih,
                                 int i_t_overflow, int i_b_overflow,
                                 int stride_off_h, int ch, int ch_num, int n) {
        auto par_conv = jit_conv_args_t();

        const int i_l_overflow = div_up(
                nstl::max(0,
                        ((jcp.kw - 1) * (jcp.dilate_w + 1) - iw - jcp.l_pad)),
                (jcp.dilate_w + 1));
        const int i_r_overflow
                = div_up(nstl::max(0,
                                 ((jcp.kw - 1) * (jcp.dilate_w + 1)
                                         - (jcp.iw - 1 - iw) - jcp.r_pad)),
                        (jcp.dilate_w + 1));

        int ow = iw + jcp.l_pad - i_r_overflow * (jcp.dilate_w + 1);
        int stride_off_w = ow % jcp.stride_w;
        ow /= jcp.stride_w;

        par_conv.src = &diff_src[diff_src_d.blk_off(n, ch, ih, iw)];
        par_conv.dst = &diff_dst[diff_dst_d.blk_off(n, ch, oh, ow)];
        par_conv.filt = &weights[weights_d.blk_off(ch, 0, 0,
                i_b_overflow + stride_off_h, i_r_overflow + stride_off_w)];

        par_conv.kh_padding = nstl::max(
                0, jcp.kh - i_t_overflow - i_b_overflow - stride_off_h);
        par_conv.kw_padding = nstl::max(
                0, jcp.kw - i_l_overflow - i_r_overflow - stride_off_w);

        par_conv.ur_str_w = ur_str_w;

        par_conv.ch_blocks = nstl::min(ch + ch_num, jcp.nb_ch) - ch;
        return par_conv;
    };

    const int ext_kw = calculate_extended_filter_size(jcp.kw, jcp.dilate_w);
    const int aux_w
            = nstl::min(jcp.iw, jcp.iw - ext_kw + jcp.r_pad + jcp.stride_w);
    const int chb_work = utils::div_up(jcp.nb_ch, jcp.nb_ch_blocking);
    parallel_nd(jcp.mb, chb_work, jcp.ih, [&](int n, int chb, int ih) {
        int ch = chb * jcp.nb_ch_blocking;
        int ch_num = jcp.nb_ch_blocking;

        const int i_t_overflow
                = div_up(nstl::max(0,
                                 (int)((jcp.kh - 1) * (jcp.dilate_h + 1) - ih
                                         - jcp.t_pad)),
                        (jcp.dilate_h + 1));
        const int i_b_overflow
                = div_up(nstl::max(0,
                                 (int)((jcp.kh - 1) * (jcp.dilate_h + 1)
                                         - (jcp.ih - 1 - ih) - jcp.b_pad)),
                        (jcp.dilate_h + 1));

        int oh = ih + jcp.t_pad - i_b_overflow * (jcp.dilate_h + 1);
        int stride_off_h = oh % jcp.stride_h;
        oh /= jcp.stride_h;

        for (int i_str_w = 0; i_str_w < jcp.stride_w; i_str_w++) {
            // left border
            int iw = i_str_w;
            int l_border = nstl::min(
                    (jcp.kw - 1) * (jcp.dilate_w + 1) - jcp.l_pad, jcp.iw);
            int ur_str_w = 1;
            for (; iw < l_border; iw += jcp.stride_w) {
                jit_conv_args_t par_conv
                        = kernel_params(ur_str_w, iw, oh, ih, i_t_overflow,
                                i_b_overflow, stride_off_h, ch, ch_num, n);
                (*kernel_)(&par_conv);
            }

            // main loop
            ur_str_w = (aux_w - iw) / jcp.stride_w;
            if (ur_str_w > 0) {
                jit_conv_args_t par_conv
                        = kernel_params(ur_str_w, iw, oh, ih, i_t_overflow,
                                i_b_overflow, stride_off_h, ch, ch_num, n);
                (*kernel_)(&par_conv);
                iw += ur_str_w * jcp.stride_w;
            }

            // right border
            ur_str_w = 1;
            for (; iw < jcp.iw; iw += jcp.stride_w) {
                jit_conv_args_t par_conv
                        = kernel_params(ur_str_w, iw, oh, ih, i_t_overflow,
                                i_b_overflow, stride_off_h, ch, ch_num, n);

                (*kernel_)(&par_conv);
            }
        }
    });
}

template struct jit_uni_dw_convolution_bwd_data_t<sve_512, data_type::f32>;
template struct jit_uni_dw_convolution_bwd_data_t<sve_256, data_type::f32>;

template <cpu_isa_t isa, data_type_t src_type, data_type_t diff_weights_type>
jit_uni_dw_convolution_bwd_weights_t<isa, src_type, diff_weights_type>::
        jit_uni_dw_convolution_bwd_weights_t(const pd_t *apd)
    : primitive_t(apd), acc_ker_(nullptr), kernel_(nullptr) {}

template <cpu_isa_t isa, data_type_t src_type, data_type_t diff_weights_type>
void jit_uni_dw_convolution_bwd_weights_t<isa, src_type,
        diff_weights_type>::execute_backward_weights(const exec_ctx_t &ctx)
        const {
    auto diff_dst = CTX_IN_MEM(const diff_dst_data_t *, DNNL_ARG_DIFF_DST);
    auto src = CTX_IN_MEM(const src_data_t *, DNNL_ARG_SRC);
    auto diff_weights
            = CTX_OUT_MEM(diff_weights_data_t *, DNNL_ARG_DIFF_WEIGHTS);

    auto diff_wei_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_wei_reduction);
    auto diff_bia_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_bia_reduction);

    const auto &jcp = pd()->jcp_;

    float *diff_bias = nullptr;
    if (jcp.bia_dt == data_type::bf16) {
        diff_bias = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                key_conv_bias_bf16_convert_wsp);
    } else {
        diff_bias = CTX_OUT_MEM(f32_data_t *, DNNL_ARG_DIFF_BIAS);
    }

    const size_t wei_size = jcp.ngroups * jcp.kh * jcp.kw;
    const size_t bias_size = jcp.with_bias ? jcp.ngroups : 0;

    const int ch_block = jcp.ch_block;

    auto set_kernel_params
            = [&](jit_dw_conv_args_t *conv_params, const int batch,
                      const int group, const int oh_start, const int work_size,
                      const unsigned char exec_flag, const size_t kh_padding,
                      const size_t filter_off) {
        const int tpad_underflow_off = jcp.t_pad - filter_off;

        conv_params->exec_flags = exec_flag;
        conv_params->kh_count = jcp.kh - kh_padding;

        const int oh_s = oh_start;
        const int oh_e = oh_start + work_size;
        const int ih_s = oh_s * jcp.stride_h;

        conv_params->filter_pad_off
                = filter_off * jcp.kw * ch_block * jcp.typesize_out;
        conv_params->oh_index = oh_s;
        conv_params->oh_count = oh_e;

        size_t diff_dst_off
                = ((batch * (jcp.ngroups / ch_block) + group) * jcp.oh
                          + oh_start)
                * jcp.ow;

        size_t src_off = ((batch * (jcp.ngroups / ch_block) + group) * jcp.ih
                                 + ih_s - tpad_underflow_off)
                * jcp.iw;

        conv_params->output = &diff_dst[diff_dst_off * ch_block];
        conv_params->input = &src[src_off * ch_block];
    };

    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        assert(nthr == jcp.nthr);

        auto conv_params = jit_dw_conv_args_t();

        /* assign iteration space to thread */
        const int ithr_g = ithr % jcp.nthr_g;
        const int ithr_mb = (ithr / jcp.nthr_g) % jcp.nthr_mb;

        /* split dimensions */
        int g_start {0}, g_end {0};
        balance211(jcp.nb_ch, jcp.nthr_g, ithr_g, g_start, g_end);

        int mb_start {0}, mb_end {0};
        balance211(jcp.mb, jcp.nthr_mb, ithr_mb, mb_start, mb_end);

        auto i_mb
                = diff_weights_type == data_type::bf16 ? ithr_mb : ithr_mb - 1;
        f32_data_t *diff_wei
                = (ithr_mb == 0 && diff_weights_type == data_type::f32)
                ? (f32_data_t *)diff_weights
                : diff_wei_reduction_buf + i_mb * wei_size;

        auto diff_bia = ithr_mb == 0
                ? diff_bias
                : diff_bia_reduction_buf + (ithr_mb - 1) * bias_size;

        for (int g = g_start; g < g_end; ++g) {
            unsigned char zero_filter_flag = FLAG_ZERO_FILTER;
            unsigned char zero_bias_flag = jcp.with_bias ? FLAG_ZERO_BIAS : 0;

            size_t diff_wei_off = g * jcp.kh * jcp.kw;
            conv_params.filter = &diff_wei[diff_wei_off * ch_block];

            if (jcp.with_bias) conv_params.bias = &diff_bia[g * ch_block];

            for (int mb = mb_start; mb < mb_end; ++mb) {
                for (int oh = 0; oh < jcp.oh; ++oh) {
                    // oh_inp: top input row (no padding) for this output row
                    /**
                     * Calculate the input height coordinate corresponding to the current output height.
                     * This maps from output to input by multiplying by the stride, necessary when stride > 1.
                     */
                    const int oh_inp = oh * jcp.stride_h;
                    // kh_t_padding: kernel rows above valid input
                    const int kh_t_padding = nstl::max(0, jcp.t_pad - oh_inp);
                    // bottom_excess: kernel rows past bottom of padded input
                    const int bottom_excess
                            = (oh_inp + jcp.kh) - (jcp.ih + jcp.t_pad);
                    // kh_b_padding: kernel rows below valid input
                    const int kh_b_padding
                            = bottom_excess > 0 ? bottom_excess : 0;
                    // kh_padding = rows to skip (top + bottom); filter_off = top skip
                    set_kernel_params(&conv_params, mb, g, oh, 1,
                            zero_filter_flag | zero_bias_flag,
                            kh_t_padding + kh_b_padding, kh_t_padding);
                    (*kernel_)(&conv_params);
                    zero_bias_flag &= ~FLAG_ZERO_BIAS;
                    zero_filter_flag &= ~FLAG_ZERO_FILTER;
                }
            }
        }
    });
}

/* TODO: Performing a Parallel Reduction could potentially improve performance;
 * this should be explored in the future if further optimizations are required.
 */
template <>
void jit_uni_dw_convolution_bwd_weights_t<sve_256,
        data_type::bf16>::execute_reduction(const exec_ctx_t &ctx) const {

    auto diff_wei_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_wei_reduction);
    auto diff_bia_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_bia_reduction);
    auto diff_weights
            = CTX_OUT_MEM(diff_weights_data_t *, DNNL_ARG_DIFF_WEIGHTS);

    const auto &jcp = pd()->jcp_;
    assert(jcp.dwei_dt == data_type::bf16);

    const size_t wei_size = jcp.ngroups * jcp.kh * jcp.kw;
    const size_t bias_size = jcp.with_bias ? jcp.ngroups : 0;

    const int ch_block = jcp.ch_block;

    float *diff_bias = nullptr;
    if (jcp.bia_dt == data_type::bf16) {
        diff_bias = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                key_conv_bias_bf16_convert_wsp);
    } else {
        diff_bias = CTX_OUT_MEM(f32_data_t *, DNNL_ARG_DIFF_BIAS);
    }

    /* Apply single-threaded 'mb' reduction */
    if (jcp.with_bias && jcp.nthr_mb > 1) {
        for (int thr_mb = 1; thr_mb < jcp.nthr_mb; ++thr_mb) {
            size_t b_accum_offset = (thr_mb - 1) * bias_size;

            for (int g = 0; g < jcp.nb_ch; ++g) {
                /* Reduction on Bias */
                PRAGMA_OMP_SIMD()
                for (int g_block = 0; g_block < ch_block; ++g_block) {
                    size_t bias_offset = g * ch_block + g_block;
                    diff_bias[bias_offset]
                            += diff_bia_reduction_buf[b_accum_offset
                                    + bias_offset];
                }
            }
        }
    }
    if (jcp.bia_dt == data_type::bf16) {
        auto diff_bias_in = CTX_OUT_MEM(bf16_data_t *, DNNL_ARG_DIFF_BIAS);
        cvt_float_to_bfloat16(diff_bias_in, diff_bias, jcp.ngroups);
    }
    /* Apply single-threaded 'mb' reduction */
    if (jcp.nthr_mb > 1) {
        for (int thr_mb = 2; thr_mb < jcp.nthr_mb; ++thr_mb) {
            size_t mb_accum_offset = thr_mb * wei_size;
            acc_ker_->accumulate(&diff_wei_reduction_buf[0],
                    &diff_wei_reduction_buf[mb_accum_offset], wei_size);
        }
        add_floats_and_cvt_to_bfloat16((bfloat16_t *)&(diff_weights[0]),
                (float *)&diff_wei_reduction_buf[0],
                (float *)&diff_wei_reduction_buf[wei_size], wei_size);
    } else {
        cvt_float_to_bfloat16((bfloat16_t *)&(diff_weights[0]),
                (const float *)&(diff_wei_reduction_buf[0]), wei_size);
    }
}

template <cpu_isa_t isa, data_type_t src_type, data_type_t diff_weights_type>
void jit_uni_dw_convolution_bwd_weights_t<isa, src_type,
        diff_weights_type>::execute_reduction(const exec_ctx_t &ctx) const {

    auto diff_wei_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_wei_reduction);
    auto diff_bia_reduction_buf
            = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                    key_conv_bia_reduction);
    auto diff_weights = CTX_OUT_MEM(f32_data_t *, DNNL_ARG_DIFF_WEIGHTS);

    const auto &jcp = pd()->jcp_;

    const size_t wei_size = jcp.ngroups * jcp.kh * jcp.kw;
    const size_t bias_size = jcp.with_bias ? jcp.ngroups : 0;

    const int ch_block = jcp.ch_block;

    assert(diff_weights_type == data_type::f32
            && jcp.dwei_dt == data_type::f32);

    float *diff_bias = nullptr;
    if (jcp.bia_dt == data_type::bf16) {
        diff_bias = ctx.get_scratchpad_grantor().template get<f32_data_t>(
                key_conv_bias_bf16_convert_wsp);
    } else {
        diff_bias = CTX_OUT_MEM(f32_data_t *, DNNL_ARG_DIFF_BIAS);
    }

    /* Apply single-threaded 'mb' reduction */
    for (int thr_mb = 1; thr_mb < jcp.nthr_mb; ++thr_mb) {
        size_t mb_accum_offset = (thr_mb - 1) * wei_size;
        size_t b_accum_offset = (thr_mb - 1) * bias_size;

        for (int g = 0; g < jcp.nb_ch; ++g) {
            /* Reduction on Bias */
            if (jcp.with_bias) {
                PRAGMA_OMP_SIMD()
                for (int g_block = 0; g_block < ch_block; ++g_block) {
                    size_t bias_offset = g * ch_block + g_block;
                    diff_bias[bias_offset]
                            += diff_bia_reduction_buf[b_accum_offset
                                    + bias_offset];
                }
            }
        }
        acc_ker_->accumulate(&diff_weights[0],
                &diff_wei_reduction_buf[mb_accum_offset], wei_size);
    }

    if (jcp.bia_dt == data_type::bf16) {
        auto diff_bias_in = CTX_OUT_MEM(bf16_data_t *, DNNL_ARG_DIFF_BIAS);
        cvt_float_to_bfloat16(diff_bias_in, diff_bias, jcp.ngroups);
    }
}

template struct jit_uni_dw_convolution_bwd_weights_t<sve_512, data_type::f32>;
template struct jit_uni_dw_convolution_bwd_weights_t<sve_256, data_type::f32>;
} // namespace aarch64
} // namespace cpu
} // namespace impl
} // namespace dnnl