onednn-src 0.1.13

/*******************************************************************************
* Copyright 2019 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 "common/c_types_map.hpp"
#include "common/dnnl_thread.hpp"
#include "common/type_helpers.hpp"
#include "common/utils.hpp"

#include "cpu/cpu_primitive.hpp"

#include "cpu/x64/jit_uni_x8s8s32x_convolution.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace x64 {

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

using namespace nstl;

#define wht_blk_off(d, g, ...) \
    (pd()->with_groups() ? (d).blk_off((g), __VA_ARGS__) \
                         : (d).blk_off(__VA_ARGS__))

template <cpu_isa_t isa>
status_t jit_uni_x8s8s32x_convolution_fwd_t<isa>::execute_forward_2d(
        const exec_ctx_t &ctx) const {
    const auto &jcp = pd()->jcp_;
    auto src = CTX_IN_MEM(const char *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const char *, DNNL_ARG_WEIGHTS);
    auto bias = CTX_IN_MEM(const char *, DNNL_ARG_BIAS);
    auto dst = CTX_OUT_MEM(char *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(pd()->jcp_.post_ops, ctx);

    const int32_t *src_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC);
    const int32_t *dst_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | 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 size_t bia_dt_size
            = pd()->with_bias() ? types::data_type_size(bias_d.data_type()) : 0;
    const size_t dst_dt_size = types::data_type_size(dst_d.data_type());

    assert(jcp.ch_block == 1);
    assert(jcp.nb_ch_blocking == 1);
    assert(jcp.nb_oc % jcp.nb_oc_blocking == 0);
    assert(jcp.nb_ch % jcp.nb_ch_blocking == 0);

    const void *src_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC);
    const void *wei_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS);
    const void *dst_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST);

    size_t offset = weights_d.size() - weights_d.additional_buffer_size();
    auto w = const_cast<char *>(weights);
    const int32_t *compensation = (jcp.signed_input)
            ? reinterpret_cast<int32_t *>(&w[offset])
            : nullptr;
    const int32_t *zp_compensation = jcp.src_zero_point
            ? reinterpret_cast<int32_t *>(&w[offset])
                    + (jcp.signed_input ? jcp.ngroups * jcp.oc : 0)
            : nullptr;
    int oc_chunks = jcp.nb_oc / jcp.nb_oc_blocking_thr_chunk;
    int nb_groups = jcp.nb_ch;
    int work_amount = jcp.mb * nb_groups * oc_chunks * jcp.oh * jcp.nb_ow;

    parallel(jcp.nthr, [= COMPAT_THIS_CAPTURE](const int ithr, const int nthr) {
        int start {0}, end {0};
        balance211(work_amount, nthr, ithr, start, end);

        auto p = jit_conv_args_t();

        size_t src_h_stride = src_d.blk_off(0, 0, 1);
        size_t dst_h_stride = dst_d.blk_off(0, 0, 1);
        size_t wht_h_stride = wht_blk_off(weights_d, 0, 0, 0, 1);

        float *dst_scales_inv_ptr = nullptr;
        if (jcp.with_dst_scales) {
            const float *dst_scales_ptr
                    = static_cast<const float *>(dst_scales);
            dst_scales_inv_ptr
                    = ctx.get_scratchpad_grantor().template get<float>(
                              key_conv_dst_scales)
                    + ithr;
            dst_scales_inv_ptr[0] = 1.f / dst_scales_ptr[0];
        }

        int n {0}, g {0}, occ {0}, oh_s {0}, owb {0};
        switch (jcp.loop_order) {
            case loop_cwgn:
                nd_iterator_init(start, occ, oc_chunks, owb, jcp.nb_ow, g,
                        nb_groups, n, jcp.mb, oh_s, jcp.oh);
                break;
            case loop_ngcw:
                nd_iterator_init(start, n, jcp.mb, g, nb_groups, occ, oc_chunks,
                        owb, jcp.nb_ow, oh_s, jcp.oh);
                break;
            case loop_nhwcg:
                nd_iterator_init(start, n, jcp.mb, oh_s, jcp.oh, owb, jcp.nb_ow,
                        occ, oc_chunks, g, nb_groups);
                break;
            default: assert(!"unsupported loop order");
        }
        while (start < end) {
            for (int occ1 = 0; occ1 < jcp.nb_oc_blocking_thr_chunk;
                    occ1 += jcp.nb_oc_blocking) {
                int ocb = occ * jcp.nb_oc_blocking_thr_chunk + occ1;
                int g_oc = (g * jcp.nb_oc + ocb) * jcp.oc_block;

                int g_ic = g * jcp.nb_ic * jcp.ic_block;

                int work_rem = end - start;
                int ih_s = -jcp.t_pad + oh_s * jcp.stride_h;
                int oh_e = oh_s + work_rem > jcp.oh ? jcp.oh : oh_s + work_rem;
                if (jcp.loop_order == loop_nhwcg)
                    oh_e = oh_s + 1; // step instead
                int ow_s = owb * jcp.ow_block;
                int iw_s = ow_s * jcp.stride_w;

                auto bias_w = bias ? bias + (bias_d.blk_off(g_oc) * bia_dt_size)
                                   : nullptr;
                const int32_t *compensation_w
                        = (jcp.signed_input) ? compensation + g_oc : nullptr;

                auto dst_w = dst
                        + dst_dt_size * dst_d.blk_off(n, g_oc, oh_s, ow_s);
                auto src_w = src + src_d.blk_off(n, g_ic, ih_s, iw_s);
                auto wht_w = weights + wht_blk_off(weights_d, g, ocb, 0);

                for (int oj = oh_s, ij = ih_s; oj < oh_e;
                        ++oj, ij += jcp.stride_h) {
                    int dilate_h = jcp.dilate_h + 1;
                    int i_t_overflow = nstl::min(
                            jcp.kh, div_up(nstl::max(0, -ij), dilate_h));
                    int i_b_overflow = nstl::min(jcp.kh,
                            div_up(nstl::max(0,
                                           ij - jcp.ih + (jcp.kh - 1) * dilate_h
                                                   + 1),
                                    dilate_h));
                    int kh_padding = nstl::max(
                            0, jcp.kh - i_t_overflow - i_b_overflow);

                    const size_t wei_stride
                            = (jcp.signed_input || jcp.src_zero_point)
                            ? 0
                            : i_t_overflow * wht_h_stride;
                    p.src = src_w + i_t_overflow * dilate_h * src_h_stride;
                    p.dst = dst_w;
                    p.filt = wht_w + wei_stride;
                    p.bias = bias_w;
                    p.compensation = compensation_w;
                    p.zp_compensation = jcp.src_zero_point
                            ? zp_compensation + g_oc
                            : nullptr;
                    p.src_zero_point = src_zero_points;
                    p.dst_zero_point = dst_zero_points;
                    p.oc_blocks = ocb;
                    p.kh_padding = kh_padding;
                    p.src_scales = src_scales;
                    p.wei_scales = jcp.with_wei_scales
                            ? static_cast<const float *>(wei_scales)
                                    + jcp.is_oc_scale * g_oc
                            : nullptr;
                    p.dst_scales = dst_scales_inv_ptr;
                    p.t_overflow = i_t_overflow;
                    p.b_overflow = i_b_overflow;
                    p.owb = owb;

                    p.post_ops_binary_rhs_arg_vec
                            = post_ops_binary_rhs_arg_vec.data();
                    p.dst_orig = dst;

                    (*kernel_)(&p);
                    src_w += src_h_stride * jcp.stride_h;
                    dst_w += dst_dt_size * dst_h_stride;
                }
            }
            switch (jcp.loop_order) {
                case loop_cwgn:
                    nd_iterator_jump(start, end, occ, oc_chunks, owb, jcp.nb_ow,
                            g, nb_groups, n, jcp.mb, oh_s, jcp.oh);
                    break;
                case loop_ngcw:
                    nd_iterator_jump(start, end, n, jcp.mb, g, nb_groups, occ,
                            oc_chunks, owb, jcp.nb_ow, oh_s, jcp.oh);
                    break;
                case loop_nhwcg:
                    ++start;
                    nd_iterator_step(n, jcp.mb, oh_s, jcp.oh, owb, jcp.nb_ow,
                            occ, oc_chunks, g, nb_groups);
                    break;
                default: assert(!"unsupported loop order");
            }
        }
    });
    return status::success;
}

template <cpu_isa_t isa>
status_t jit_uni_x8s8s32x_convolution_fwd_t<isa>::execute_forward_1d(
        const exec_ctx_t &ctx) const {
    const auto &jcp = pd()->jcp_;
    auto src = CTX_IN_MEM(const char *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const char *, DNNL_ARG_WEIGHTS);
    auto bias = CTX_IN_MEM(const char *, DNNL_ARG_BIAS);
    auto dst = CTX_OUT_MEM(char *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(pd()->jcp_.post_ops, ctx);

    const int32_t *src_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC);
    const int32_t *dst_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | 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 size_t bia_dt_size
            = pd()->with_bias() ? types::data_type_size(bias_d.data_type()) : 0;
    const size_t dst_dt_size = types::data_type_size(dst_d.data_type());

    assert(jcp.nb_oc % jcp.nb_oc_blocking == 0);
    assert(jcp.nb_ch % jcp.nb_ch_blocking == 0);

    const void *src_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC);
    const void *wei_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS);
    const void *dst_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST);

    size_t extra_data_offset
            = weights_d.size() - weights_d.additional_buffer_size();
    size_t ch_offset = jcp.is_depthwise ? jcp.nb_ch * jcp.ch_block
                                        : jcp.ngroups * jcp.oc;
    auto w = const_cast<char *>(weights);
    const int32_t *compensation = (jcp.signed_input)
            ? reinterpret_cast<int32_t *>(&w[extra_data_offset])
            : nullptr;
    const int32_t *zp_compensation = jcp.src_zero_point
            ? reinterpret_cast<int32_t *>(&w[extra_data_offset])
                    + (jcp.signed_input ? ch_offset : 0)
            : nullptr;

    int oc_chunks = jcp.nb_oc / jcp.nb_oc_blocking;
    int nb_groups = jcp.nb_ch / jcp.nb_ch_blocking;
    int group_block = jcp.ch_block;
    int work_amount = jcp.mb * nb_groups * oc_chunks * jcp.nb_ow;
    parallel(jcp.nthr, [= COMPAT_THIS_CAPTURE](const int ithr, const int nthr) {
        int start {0}, end {0};
        balance211(work_amount, nthr, ithr, start, end);

        auto p = jit_conv_args_t();

        float *dst_scales_inv_ptr = nullptr;
        if (jcp.with_dst_scales) {
            const float *dst_scales_ptr
                    = static_cast<const float *>(dst_scales);
            dst_scales_inv_ptr
                    = ctx.get_scratchpad_grantor().template get<float>(
                              key_conv_dst_scales)
                    + ithr;
            dst_scales_inv_ptr[0] = 1.f / dst_scales_ptr[0];
        }

        int n {0}, gg {0}, occ {0}, owb {0};
        switch (jcp.loop_order) {
            case loop_cwgn:
                nd_iterator_init(start, occ, oc_chunks, owb, jcp.nb_ow, gg,
                        nb_groups, n, jcp.mb);
                break;
            case loop_gncw:
                nd_iterator_init(start, gg, nb_groups, n, jcp.mb, occ,
                        oc_chunks, owb, jcp.nb_ow);
                break;
            case loop_ngcw:
                nd_iterator_init(start, n, jcp.mb, gg, nb_groups, occ,
                        oc_chunks, owb, jcp.nb_ow);
                break;
            case loop_nwcg:
                nd_iterator_init(start, n, jcp.mb, owb, jcp.nb_ow, occ,
                        oc_chunks, gg, nb_groups);
                break;
            default: assert(!"unsupported loop order");
        }
        while (start < end) {
            int ocb = occ * jcp.nb_oc_blocking;
            int gb = gg * jcp.nb_ch_blocking;
            int g = gb * group_block;
            int g_oc = (g * jcp.nb_oc + ocb) * jcp.oc_block;
            int g_ic = g * jcp.nb_ic * jcp.ic_block;
            int ow_s = owb * jcp.ow_block;
            int iw_s = ow_s * jcp.stride_w;

            p.bias = bias ? bias + (bias_d.blk_off(g_oc) * bia_dt_size)
                          : nullptr;
            p.compensation = (jcp.signed_input) ? compensation + g_oc : nullptr;
            p.zp_compensation
                    = jcp.src_zero_point ? zp_compensation + g_oc : nullptr;
            p.src_zero_point = src_zero_points;
            p.dst_zero_point = dst_zero_points;
            p.dst = dst + dst_dt_size * dst_d.blk_off(n, g_oc, ow_s);
            p.src = src + src_d.blk_off(n, g_ic, iw_s);
            p.filt = weights + wht_blk_off(weights_d, gb, ocb, 0);
            p.src_scales = src_scales;
            p.wei_scales = jcp.with_wei_scales
                    ? static_cast<const float *>(wei_scales)
                            + jcp.is_oc_scale * g_oc
                    : nullptr;
            p.dst_scales = dst_scales_inv_ptr;
            p.oc_blocks = jcp.is_depthwise ? gb : ocb;
            p.kh_padding = jcp.kh;
            p.t_overflow = 0;
            p.b_overflow = 0;
            p.owb = owb;

            p.post_ops_binary_rhs_arg_vec = post_ops_binary_rhs_arg_vec.data();
            p.dst_orig = dst;

            (*kernel_)(&p);

            ++start;
            switch (jcp.loop_order) {
                case loop_cwgn:
                    nd_iterator_step(occ, oc_chunks, owb, jcp.nb_ow, gg,
                            nb_groups, n, jcp.mb);
                    break;
                case loop_gncw:
                    nd_iterator_step(gg, nb_groups, n, jcp.mb, occ, oc_chunks,
                            owb, jcp.nb_ow);
                    break;
                case loop_ngcw:
                    nd_iterator_step(n, jcp.mb, gg, nb_groups, occ, oc_chunks,
                            owb, jcp.nb_ow);
                    break;
                case loop_nwcg:
                    nd_iterator_step(n, jcp.mb, owb, jcp.nb_ow, occ, oc_chunks,
                            gg, nb_groups);
                    break;
                default: assert(!"unsupported loop order");
            }
        }
    });
    return status::success;
}

template <cpu_isa_t isa>
status_t jit_uni_x8s8s32x_convolution_fwd_t<isa>::execute_forward_2d_dw(
        const exec_ctx_t &ctx) const {
    const auto &jcp = pd()->jcp_;
    auto src = CTX_IN_MEM(const char *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const char *, DNNL_ARG_WEIGHTS);
    auto bias = CTX_IN_MEM(const char *, DNNL_ARG_BIAS);
    auto dst = CTX_OUT_MEM(char *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(pd()->jcp_.post_ops, ctx);

    const int32_t *src_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC);
    const int32_t *dst_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | 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 size_t bia_dt_size
            = pd()->with_bias() ? types::data_type_size(bias_d.data_type()) : 0;
    const size_t dst_dt_size = types::data_type_size(dst_d.data_type());

    assert(jcp.ic_block == 1);
    assert(jcp.oc_block == 1);
    assert(jcp.nb_ic == 1);
    assert(jcp.nb_oc == 1);
    assert(jcp.nb_oc_blocking == 1);
    assert(jcp.nb_ch % jcp.nb_ch_blocking == 0);

    const void *src_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC);
    const void *wei_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS);
    const void *dst_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST);

    size_t offset = weights_d.size() - weights_d.additional_buffer_size();
    auto w = const_cast<char *>(weights);
    const int32_t *compensation = (jcp.signed_input)
            ? reinterpret_cast<int32_t *>(&w[offset])
            : nullptr;
    const int32_t *zp_compensation = jcp.src_zero_point
            ? reinterpret_cast<int32_t *>(&w[offset])
                    + (jcp.signed_input ? jcp.nb_ch * jcp.ch_block : 0)
            : nullptr;
    int nb_groups = jcp.nb_ch / jcp.nb_ch_blocking;
    int group_block = jcp.ch_block;

    parallel_nd(jcp.mb, jcp.oh, jcp.nb_ow, nb_groups,
            [= COMPAT_THIS_CAPTURE](dim_t n, dim_t oh_s, dim_t owb, dim_t gg) {
        auto p = jit_conv_args_t();

        size_t src_h_stride = src_d.blk_off(0, 0, 1);
        size_t wht_h_stride = wht_blk_off(weights_d, 0, 0, 0, 1);

        int gb = gg * jcp.nb_ch_blocking;
        int g = gb * group_block;

        int ih_s = -jcp.t_pad + oh_s * jcp.stride_h;
        int ow_s = owb * jcp.ow_block;
        int iw_s = ow_s * jcp.stride_w;

        auto bias_w = bias ? bias + (bias_d.blk_off(g) * bia_dt_size) : nullptr;
        const int32_t *compensation_w
                = jcp.signed_input ? compensation + g : nullptr;

        auto dst_w = dst + dst_dt_size * dst_d.blk_off(n, g, oh_s, ow_s);
        auto src_w = src + src_d.blk_off(n, g, ih_s, iw_s);
        auto wht_w = weights + wht_blk_off(weights_d, gb, 0);

        float *dst_scales_inv_ptr = nullptr;
        if (jcp.with_dst_scales) {
            const float *dst_scales_ptr
                    = static_cast<const float *>(dst_scales);
            dst_scales_inv_ptr
                    = ctx.get_scratchpad_grantor().template get<float>(
                              key_conv_dst_scales)
                    + g;
            dst_scales_inv_ptr[0] = 1.f / dst_scales_ptr[0];
        }

        const float *wei_scales_ptr = jcp.with_wei_scales
                ? static_cast<const float *>(wei_scales) + jcp.is_oc_scale * g
                : nullptr;

        int dilate_h = jcp.dilate_h + 1;
        int i_t_overflow
                = nstl::min(jcp.kh, div_up(nstl::max(0, -ih_s), dilate_h));
        int i_b_overflow = nstl::min(jcp.kh,
                div_up(nstl::max(
                               0, ih_s - jcp.ih + (jcp.kh - 1) * dilate_h + 1),
                        dilate_h));
        int kh_padding = nstl::max(0, jcp.kh - i_t_overflow - i_b_overflow);

        size_t wei_stride = (jcp.signed_input || jcp.src_zero_point)
                ? 0
                : i_t_overflow * wht_h_stride;
        p.src = src_w + i_t_overflow * dilate_h * src_h_stride;
        p.dst = dst_w;
        p.filt = wht_w + wei_stride;
        p.bias = bias_w;
        p.compensation = compensation_w;
        p.zp_compensation = jcp.src_zero_point ? zp_compensation + g : nullptr;
        p.src_zero_point = src_zero_points;
        p.dst_zero_point = dst_zero_points;
        p.oc_blocks = gb;
        p.kh_padding = kh_padding;
        p.src_scales = src_scales;
        p.wei_scales = wei_scales_ptr;
        p.dst_scales = dst_scales_inv_ptr;
        p.t_overflow = i_t_overflow;
        p.b_overflow = i_b_overflow;
        p.owb = owb;

        p.post_ops_binary_rhs_arg_vec = post_ops_binary_rhs_arg_vec.data();
        p.dst_orig = dst;

        (*kernel_)(&p);
    });
    return status::success;
}

template <cpu_isa_t isa>
status_t jit_uni_x8s8s32x_convolution_fwd_t<isa>::execute_forward_3d(
        const exec_ctx_t &ctx) const {
    const auto &jcp = pd()->jcp_;
    auto src = CTX_IN_MEM(const char *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const char *, DNNL_ARG_WEIGHTS);
    auto bias = CTX_IN_MEM(const char *, DNNL_ARG_BIAS);
    auto dst = CTX_OUT_MEM(char *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(pd()->jcp_.post_ops, ctx);

    const int32_t *src_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC);
    const int32_t *dst_zero_points = CTX_IN_MEM(
            const int32_t *, DNNL_ARG_ATTR_ZERO_POINTS | 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 size_t bia_dt_size
            = pd()->with_bias() ? types::data_type_size(bias_d.data_type()) : 0;
    const size_t dst_dt_size = types::data_type_size(dst_d.data_type());

    assert(jcp.ch_block == 1);
    assert(jcp.nb_ch_blocking == 1);
    assert(jcp.nb_oc % jcp.nb_oc_blocking == 0);
    assert(jcp.nb_ch % jcp.nb_ch_blocking == 0);

    const void *src_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC);
    const void *wei_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS);
    const void *dst_scales
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST);

    size_t offset = weights_d.size() - weights_d.additional_buffer_size();
    auto w = const_cast<char *>(weights);
    const int32_t *compensation = (jcp.signed_input)
            ? reinterpret_cast<int32_t *>(&w[offset])
            : nullptr;
    const int32_t *zp_compensation = jcp.src_zero_point
            ? reinterpret_cast<int32_t *>(&w[offset])
                    + (jcp.signed_input ? jcp.ngroups * jcp.oc : 0)
            : nullptr;
    int oc_chunks = jcp.nb_oc / jcp.nb_oc_blocking_thr_chunk;
    int nb_groups = jcp.nb_ch;
    int work_amount
            = jcp.mb * nb_groups * oc_chunks * jcp.od * jcp.oh * jcp.nb_ow;

    parallel(jcp.nthr, [= COMPAT_THIS_CAPTURE](const int ithr, const int nthr) {
        int start {0}, end {0};
        balance211(work_amount, nthr, ithr, start, end);

        auto p = jit_conv_args_t();

        float *dst_scales_inv_ptr = nullptr;
        if (jcp.with_dst_scales) {
            const float *dst_scales_ptr
                    = static_cast<const float *>(dst_scales);
            dst_scales_inv_ptr
                    = ctx.get_scratchpad_grantor().template get<float>(
                              key_conv_dst_scales)
                    + ithr;
            dst_scales_inv_ptr[0] = 1.f / dst_scales_ptr[0];
        }

        size_t src_d_stride = src_d.blk_off(0, 0, 1);
        size_t src_h_stride = src_d.blk_off(0, 0, 0, 1);
        size_t dst_h_stride = dst_d.blk_off(0, 0, 0, 1);
        size_t wht_d_stride = wht_blk_off(weights_d, 0, 0, 0, 1);
        size_t wht_h_stride = wht_blk_off(weights_d, 0, 0, 0, 0, 1);

        int n {0}, g {0}, occ {0}, od_s {0}, oh_s {0}, owb {0};
        switch (jcp.loop_order) {
            case loop_cwgn:
                nd_iterator_init(start, occ, oc_chunks, owb, jcp.nb_ow, g,
                        nb_groups, n, jcp.mb, od_s, jcp.od, oh_s, jcp.oh);
                break;
            case loop_ngcw:
                nd_iterator_init(start, n, jcp.mb, g, nb_groups, occ, oc_chunks,
                        owb, jcp.nb_ow, od_s, jcp.od, oh_s, jcp.oh);
                break;
            case loop_nhwcg:
                nd_iterator_init(start, n, jcp.mb, od_s, jcp.od, oh_s, jcp.oh,
                        owb, jcp.nb_ow, occ, oc_chunks, g, nb_groups);
                break;
            default: assert(!"unsupported loop order");
        }
        while (start < end) {
            for (int occ1 = 0; occ1 < jcp.nb_oc_blocking_thr_chunk;
                    occ1 += jcp.nb_oc_blocking) {
                int ocb = occ * jcp.nb_oc_blocking_thr_chunk + occ1;
                int g_oc = (g * jcp.nb_oc + ocb) * jcp.oc_block;

                int g_ic = g * jcp.nb_ic * jcp.ic_block;

                int work_rem = end - start;
                int ih_s = -jcp.t_pad + oh_s * jcp.stride_h;
                int oh_e = oh_s + work_rem > jcp.oh ? jcp.oh : oh_s + work_rem;
                if (jcp.loop_order == loop_nhwcg)
                    oh_e = oh_s + 1; // step instead
                int ow_s = owb * jcp.ow_block;
                int iw_s = ow_s * jcp.stride_w;
                int id_s = -jcp.f_pad + od_s * jcp.stride_d;
                int dilate_d = jcp.dilate_d + 1;
                int d_f_overflow = nstl::min(
                        jcp.kd, div_up(nstl::max(0, -id_s), dilate_d));
                int d_back_overflow = nstl::min(jcp.kd,
                        div_up(nstl::max(0,
                                       id_s - jcp.id + (jcp.kd - 1) * dilate_d
                                               + 1),
                                dilate_d));

                int kd_padding
                        = nstl::max(0, jcp.kd - d_f_overflow - d_back_overflow);

                auto bias_w = bias ? bias + (bias_d.blk_off(g_oc) * bia_dt_size)
                                   : nullptr;
                const int32_t *compensation_w
                        = (jcp.signed_input) ? compensation + g_oc : nullptr;
                p.zp_compensation
                        = jcp.src_zero_point ? zp_compensation + g_oc : nullptr;
                p.src_zero_point = src_zero_points;
                p.dst_zero_point = dst_zero_points;

                auto dst_w = dst
                        + dst_dt_size
                                * dst_d.blk_off(n, g_oc, od_s, oh_s, ow_s);
                auto src_w = src + src_d.blk_off(n, g_ic, id_s, ih_s, iw_s)
                        + d_f_overflow * dilate_d * src_d_stride;
                auto wht_w = weights + wht_blk_off(weights_d, g, ocb, 0)
                        + ((jcp.signed_input || jcp.src_zero_point)
                                          ? 0
                                          : d_f_overflow)
                                * wht_d_stride;

                for (int oj = oh_s, ij = ih_s; oj < oh_e;
                        ++oj, ij += jcp.stride_h) {
                    int dilate_h = jcp.dilate_h + 1;
                    int i_t_overflow = nstl::min(
                            jcp.kh, div_up(nstl::max(0, -ij), dilate_h));
                    int i_b_overflow = nstl::min(jcp.kh,
                            div_up(nstl::max(0,
                                           ij - jcp.ih + (jcp.kh - 1) * dilate_h
                                                   + 1),
                                    dilate_h));
                    int kh_padding = nstl::max(
                            0, jcp.kh - i_t_overflow - i_b_overflow);

                    size_t wei_stride = (jcp.signed_input || jcp.src_zero_point)
                            ? 0
                            : wht_h_stride * i_t_overflow;
                    p.src = src_w + i_t_overflow * dilate_h * src_h_stride;
                    p.dst = dst_w;
                    p.filt = wht_w + wei_stride;
                    p.bias = bias_w;
                    p.compensation = compensation_w;
                    p.oc_blocks = ocb;
                    p.kh_padding = kh_padding;
                    p.kd_padding = kd_padding;
                    p.src_scales = src_scales;
                    p.wei_scales = jcp.with_wei_scales
                            ? static_cast<const float *>(wei_scales)
                                    + jcp.is_oc_scale * g_oc
                            : nullptr;
                    p.dst_scales = dst_scales_inv_ptr;
                    p.t_overflow = i_t_overflow;
                    p.b_overflow = i_b_overflow;
                    p.f_overflow = d_f_overflow;
                    p.back_overflow = d_back_overflow;
                    p.owb = owb;

                    p.post_ops_binary_rhs_arg_vec
                            = post_ops_binary_rhs_arg_vec.data();
                    p.dst_orig = dst;

                    (*kernel_)(&p);
                    src_w += src_h_stride * jcp.stride_h;
                    dst_w += dst_dt_size * dst_h_stride;
                }
            }
            switch (jcp.loop_order) {
                case loop_cwgn:
                    nd_iterator_jump(start, end, occ, oc_chunks, owb, jcp.nb_ow,
                            g, nb_groups, n, jcp.mb, od_s, jcp.od, oh_s,
                            jcp.oh);
                    break;
                case loop_ngcw:
                    nd_iterator_jump(start, end, n, jcp.mb, g, nb_groups, occ,
                            oc_chunks, owb, jcp.nb_ow, od_s, jcp.od, oh_s,
                            jcp.oh);
                    break;
                case loop_nhwcg:
                    ++start;
                    nd_iterator_step(n, jcp.mb, od_s, jcp.od, oh_s, jcp.oh, owb,
                            jcp.nb_ow, occ, oc_chunks, g, nb_groups);
                    break;
                default: assert(!"unsupported loop order");
            }
        }
    });
    return status::success;
}

template struct jit_uni_x8s8s32x_convolution_fwd_t<sse41>;
template struct jit_uni_x8s8s32x_convolution_fwd_t<avx2>;

} // namespace x64
} // namespace cpu
} // namespace impl
} // namespace dnnl