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 <atomic>

#include "oneapi/dnnl/dnnl_types.h"

#include "common/bfloat16.hpp"
#include "common/c_types_map.hpp"
#include "common/dnnl_thread.hpp"
#include "common/type_helpers.hpp"
#include "common/utils.hpp"
#include "cpu/x64/gemm_bf16_convolution.hpp"
#include "cpu/x64/injectors/jit_uni_binary_injector.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 dnnl::impl::cpu::x64::bf16_support;

// Below two stand-alone functions are moved out from execute_backward_data
// and execute_backward_weights to avoid warnings with gcc 6.x and 7.x compilers
// "declared with greater visibility than the type of its field"
// when one lambda function is delcared whithin the other one
void store_bfloat16_in_parallel(bfloat16_t *output_data, const float *acc_data,
        size_t parallel_work, size_t parallel_work_size, bool do_in_parallel) {
    parallel(do_in_parallel ? 0 : 1, [&](const int ithr, const int nthr) {
        size_t start = 0, end = 0;
        balance211(parallel_work, nthr, ithr, start, end);
        if (start < end)
            cvt_float_to_bfloat16(&output_data[start * parallel_work_size],
                    &acc_data[start * parallel_work_size],
                    (end - start) * parallel_work_size);
    });
}

void cvt_acc_to_dst(const conv_gemm_conf_t &jcp, size_t g_start, size_t g_end,
        const float *acc_base, bfloat16_t *diff_weights) {
    const size_t parallel_work_size = jcp.ic * jcp.ks;
    parallel(jcp.nthr == 1 ? 0 : 1, [&](const int ithr, const int nthr) {
        size_t w_start = 0, w_end = 0;
        balance211(parallel_work_size, nthr, ithr, w_start, w_end);
        for_(auto w = w_start; w < w_end; ++w)
        for (auto g = g_start; g < g_end; ++g) {
            const float *__restrict acc_ptr
                    = acc_base + (w * jcp.ngroups + g) * jcp.oc;
            bfloat16_t *__restrict dw_ptr
                    = diff_weights + (w * jcp.ngroups + g) * jcp.oc;
            cvt_float_to_bfloat16(dw_ptr, acc_ptr, jcp.oc);
        }
    });
}

template <data_type_t dst_data_type>
gemm_bf16_convolution_fwd_t<dst_data_type>::pp_ker_t::pp_ker_t(const pd_t *pd)
    : jit_generator_t(jit_name())
    , jcp_(pd->jcp_)
    , do_sum_(dst_data_type != data_type::f32 && jcp_.with_sum)
    , max_data_reg_idx_(31)
    , max_unroll_(12)
    , compute_reg_step_(1)
    , data_reg_base_idx_(0) {
    using namespace types;
    using namespace Xbyak;

    if (!mayiuse(avx512_core))
        // bf16 is not supported
        return;

    const auto &post_ops = jcp_.post_ops;
    if (jcp_.with_eltwise || jcp_.with_binary) {
#define PARAM_OFF(field) offsetof(ker_args_t, field)
        static constexpr bool preserve_gpr = true;
        static constexpr bool preserve_vmm = true;
        static constexpr size_t helper_vmm_idx = 31;
        static constexpr size_t tail_size = 1;
        static constexpr bool use_exact_tail_scalar_bcast = false;
        const binary_injector::rhs_arg_static_params_t rhs_arg_static_params {
                helper_vmm_idx, reserved_eltwise_gpr, r14, r15, preserve_gpr,
                preserve_vmm, PARAM_OFF(post_ops_binary_rhs_arg_vec),
                PARAM_OFF(dst_orig), memory_desc_wrapper(pd->dst_md()),
                tail_size, kreg_rem_mask, use_exact_tail_scalar_bcast};
        const binary_injector::static_params_t binary_static_params {
                this->reg_param, rhs_arg_static_params};
        static constexpr bool save_state = true;
        const eltwise_injector::static_params_t eltwise_static_params {
                save_state, reserved_eltwise_gpr, reserved_eltwise_maskr};

        postops_injector_ = utils::make_unique<
                injector::jit_uni_postops_injector_t<avx512_core>>(
                this, post_ops, binary_static_params, eltwise_static_params);
#undef PARAM_OFF
    }

    if (do_sum_) {
        compute_reg_step_ = 2;
        vreg_sum_scale = Zmm(data_reg_base_idx_++);
    }

    if (jcp_.with_bias) vreg_bias = Zmm(data_reg_base_idx_++);

    vlen_ = cpu_isa_traits_t<avx512_core>::vlen / sizeof(float);

    isa_ = mayiuse(avx512_core_bf16) ? avx512_core_bf16
                                     : bf16_emulation_t::get_isa();

    if (isa_ != avx512_core_bf16) {
        max_data_reg_idx_ = 26;
        bf16_emu_ = utils::make_unique<bf16_emulation_t>(this,
                bf16_emu_reserv_1, bf16_emu_reserv_2, bf16_emu_reserv_3,
                bf16_emu_reserv_4, bf16_emu_reserv_5, bf16_emu_reserv_6);
    }

    max_unroll_
            = (max_data_reg_idx_ - data_reg_base_idx_ + 1) / compute_reg_step_;
}

template <data_type_t dst_data_type>
void gemm_bf16_convolution_fwd_t<dst_data_type>::pp_ker_t::apply_postops(
        const bool apply_mask, const size_t out_offset, const int vmm_idx) {
#define PARAM_OFF(x) offsetof(ker_args_t, x)
    if (jcp_.with_eltwise || jcp_.with_binary) {
        if (jcp_.with_binary) {
            binary_injector::rhs_arg_dynamic_params_t rhs_arg_params;
            rhs_arg_params.vmm_idx_to_out_reg.emplace(vmm_idx, reg_dst);
            rhs_arg_params.vmm_idx_to_out_elem_off_val.emplace(
                    vmm_idx, out_offset * sizeof(dst_data_t));
            if (apply_mask) rhs_arg_params.vmm_tail_idx_.emplace(vmm_idx);

            postops_injector_->compute_vector(vmm_idx, rhs_arg_params);
        } else
            postops_injector_->compute_vector(vmm_idx);
    }
#undef PARAM_OFF
}

template <data_type_t dst_data_type>
void gemm_bf16_convolution_fwd_t<dst_data_type>::pp_ker_t::generate() {
    using namespace Xbyak;
    using namespace utils;

    preamble();

#ifdef _WIN32
    mov(reg_param, rcx);
#endif

#define PARAM_OFF(x) offsetof(ker_args_t, x)
    mov(reg_dst_base, ptr[reg_param + PARAM_OFF(dst)]);
    mov(reg_acc_base, ptr[reg_param + PARAM_OFF(acc)]);
    if (jcp_.with_bias) mov(reg_bias, ptr[reg_param + PARAM_OFF(bias)]);
    mov(reg_dst_str, ptr[reg_param + PARAM_OFF(dst_stride_in_bytes)]);
    mov(reg_acc_str, ptr[reg_param + PARAM_OFF(acc_stride_in_bytes)]);
    mov(reg_len, ptr[reg_param + PARAM_OFF(spatial_length)]);
    mov(reg_oc_iter, ptr[reg_param + PARAM_OFF(oc_work)]);

    if (do_sum_)
        vbroadcastss(vreg_sum_scale, ptr[reg_param + PARAM_OFF(sum_scale)]);
#undef PARAM_OFF

    // Load accumulated value, apply sum (if any), bias (if any)
    // and relu (if any); then convert to destination type and store
    auto compute = [&](size_t offset, int idx, bool apply_mask) {
        auto acc_addr = ptr[reg_acc + offset * sizeof(acc_data_t)];
        auto vreg_dst_ = vreg_dst(idx);

        if (dst_data_type == data_type::bf16 && isa_ != avx512_core_bf16)
            bf16_emu_->init_vcvtneps2bf16();

        if (apply_mask) vreg_dst_ = vreg_dst_ | kreg_rem_mask;
        vmovups(vreg_dst_, acc_addr);

        if (jcp_.with_bias) vaddps(vreg_dst(idx), vreg_dst(idx), vreg_bias);

        auto dst_addr = ptr[reg_dst + offset * sizeof(dst_data_t)];
        if (do_sum_) {
            auto vreg_prev_dst_ = vreg_prev_dst(idx);
            if (dst_data_type == data_type::f32) {
                if (apply_mask) vreg_prev_dst_ = vreg_prev_dst_ | kreg_rem_mask;

                vmovups(vreg_prev_dst_, dst_addr);
            } else if (dst_data_type == data_type::bf16) {
                auto vreg_prev_dst_ymm_ = vreg_prev_dst_ymm(idx);
                if (apply_mask)
                    vreg_prev_dst_ymm_ = vreg_prev_dst_ymm_ | kreg_rem_mask;

                vmovdqu16(vreg_prev_dst_ymm_, dst_addr);
                vpmovzxwd(vreg_prev_dst(idx), vreg_prev_dst_ymm_);
                vpslld(vreg_prev_dst(idx), vreg_prev_dst(idx), 0x10);
            } else
                assert(!"unsupported data type");

            vfmadd231ps(vreg_dst(idx), vreg_prev_dst(idx), vreg_sum_scale);
        }

        apply_postops(apply_mask, offset, vreg_dst_idx(idx));

        if (dst_data_type == data_type::bf16) {
            // TODO: implement store by zmm registers for bf16
            auto vreg_dst_ymm_ = vreg_dst_ymm(idx);
            if (isa_ == avx512_core_bf16)
                vcvtneps2bf16(vreg_dst_ymm_, vreg_dst(idx));
            else
                bf16_emu_->vcvtneps2bf16(vreg_dst_ymm_, vreg_dst(idx));

            if (apply_mask) vreg_dst_ymm_ = vreg_dst_ymm_ | kreg_rem_mask;

            vmovdqu16(dst_addr, vreg_dst_ymm_);
        } else if (dst_data_type == data_type::f32)
            vmovups(dst_addr, vreg_dst_);
        else
            assert(!"unimplemented");
    };

    // Advance all pointers by an immediate
    auto advance_ptrs_imm = [&](size_t offset) {
        add(reg_dst, offset * sizeof(dst_data_t));
        add(reg_acc, offset * sizeof(acc_data_t));
    };

    Xbyak::Label oc_loop, oc_loop_end;

    cmp(reg_oc_iter, 0);
    jle(oc_loop_end, T_NEAR);

    L(oc_loop);

    mov(reg_len_iter, reg_len);
    mov(reg_dst, reg_dst_base);
    mov(reg_acc, reg_acc_base);

    if (jcp_.with_bias) vbroadcastss(vreg_bias, ptr[reg_bias]);

    constexpr int n_unroll = default_unroll_2_pow_; // unroll by powers of 2
            // from 2^n to 2^0
    assert((1 << n_unroll) <= max_unroll_);

    Xbyak::Label l_simd_loop[n_unroll + 2], l_simd_notail;
    for (int i = n_unroll; i >= 0; i--) {
        const int unroll = 1 << i; // 4, 2, 1
        L(l_simd_loop[i + 1]);
        {
            const int loop_len = unroll * vlen_;
            cmp(reg_len_iter, loop_len);
            jl(l_simd_loop[i], T_NEAR);
            for (int j = 0; j < unroll; j++)
                compute(j * vlen_, j, false);

            advance_ptrs_imm(loop_len);
            sub(reg_len_iter, loop_len);
            jmp(l_simd_loop[i + 1], T_NEAR);
        }
    }
    L(l_simd_loop[0]);

    mov(reg_tmp, reg_len_iter); // reg_tmp is rcx, and we need cl for the shift
    mov(reg_rem_mask, 1);
    shl(reg_rem_mask, cl); // reg_tmp == rcx and reg_tail < vlen_ == 16
    sub(reg_rem_mask, 1);
    jz(l_simd_notail, T_NEAR);
    kmovq(kreg_rem_mask, reg_rem_mask);
    compute(0, 0, true);

    L(l_simd_notail);

    add(reg_dst_base, reg_dst_str);
    add(reg_acc_base, reg_acc_str);
    if (jcp_.with_bias) add(reg_bias, sizeof(acc_data_t));

    dec(reg_oc_iter);
    jnz(oc_loop, T_NEAR); // oc_loop end

    L(oc_loop_end);

    postamble();

    if (jcp_.with_eltwise)
        postops_injector_->prepare_table(/* generate = */ true);
}

// operator () specialized for nspc format
template <data_type_t dst_data_type>
void gemm_bf16_convolution_fwd_t<dst_data_type>::pp_ker_t::operator()(
        dst_data_t *dst, const acc_data_t *acc, const acc_data_t *bias,
        float sum_scale, size_t oc_work,
        const void *post_ops_binary_rhs_arg_vec, const void *dst_orig,
        const size_t g_oc_offset) {

    ker_args_t args;
    args.acc = acc;
    args.dst = dst;
    args.bias = bias;
    args.sum_scale = sum_scale;
    args.dst_stride_in_bytes = sizeof(dst_data_t);
    args.acc_stride_in_bytes = sizeof(acc_data_t);
    args.spatial_length = 1;
    args.oc_work = oc_work;

    args.post_ops_binary_rhs_arg_vec = post_ops_binary_rhs_arg_vec;
    args.dst_orig = dst_orig;
    args.g_oc_offset = g_oc_offset;
    jit_generator_t::operator()(&args);
}

// operator () specialized for ncsp format
template <data_type_t dst_data_type>
void gemm_bf16_convolution_fwd_t<dst_data_type>::pp_ker_t::operator()(
        dst_data_t *dst, const acc_data_t *acc, const acc_data_t *bias,
        float sum_scale, size_t dst_stride_in_elements,
        size_t acc_stride_in_elements, size_t sp_len, size_t oc_len,
        const void *post_ops_binary_rhs_arg_vec, const void *dst_orig,
        const size_t g_oc_offset) {
    if (sp_len == 0) return;

    ker_args_t args;
    args.acc = acc;
    args.dst = dst;
    args.bias = bias;
    args.sum_scale = sum_scale;
    args.dst_stride_in_bytes = dst_stride_in_elements * sizeof(dst_data_t);
    args.acc_stride_in_bytes = acc_stride_in_elements * sizeof(acc_data_t);
    args.spatial_length = sp_len;
    args.oc_work = oc_len;

    args.post_ops_binary_rhs_arg_vec = post_ops_binary_rhs_arg_vec;
    args.dst_orig = dst_orig;
    args.g_oc_offset = g_oc_offset;
    jit_generator_t::operator()(&args);
}

template <data_type_t dst_data_type>
status_t gemm_bf16_convolution_fwd_t<dst_data_type>::execute_forward_nspc(
        const exec_ctx_t &ctx) const {
    auto src_base = CTX_IN_MEM(const src_data_t *, DNNL_ARG_SRC);
    auto wei_base = CTX_IN_MEM(const wei_data_t *, DNNL_ARG_WEIGHTS);
    auto dst_base = CTX_OUT_MEM(dst_data_t *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(
                    this->pd()->attr()->post_ops_, ctx);

    const auto &scratchpad = ctx.get_scratchpad_grantor();
    const conv_gemm_conf_t &jcp = pd()->jcp_;

    float *bia_base = nullptr;
    if (jcp.with_bias) {
        if (pd()->desc()->bias_desc.data_type == data_type::bf16) {
            auto bias_in = CTX_IN_MEM(const bfloat16_t *, DNNL_ARG_BIAS);
            bia_base = ctx.get_scratchpad_grantor().template get<float>(
                    key_conv_bias_bf16_convert_wsp);
            cvt_bfloat16_to_float(bia_base, bias_in, jcp.ngroups * jcp.oc);
        } else {
            auto bias_in = CTX_IN_MEM(const float *, DNNL_ARG_BIAS);
            bia_base = const_cast<float *>(bias_in);
        }
    }
    assert(IMPLICATION(jcp.ow_block != jcp.ow, jcp.oh_block == 1));

    std::atomic<status_t> st(status::success);
    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        status_t st_thr = execute_forward_thr_nspc(ithr, nthr, src_base,
                wei_base, bia_base, dst_base, scratchpad,
                post_ops_binary_rhs_arg_vec.data());
        if (st_thr != status::success) st = st_thr;
    });

    return st;
}

template <data_type_t dst_data_type>
status_t gemm_bf16_convolution_fwd_t<dst_data_type>::execute_forward_thr_nspc(
        const int ithr, const int nthr, const src_data_t *src_base,
        const wei_data_t *wei_base, const float *bia_base, dst_data_t *dst_base,
        const memory_tracking::grantor_t &scratchpad,
        const void *post_ops_binary_rhs_arg_vec) const {
    const conv_gemm_conf_t &jcp = pd()->jcp_;

    // Src Format: mb-spatial-groups-input_channels
    const size_t src_mb_stride = static_cast<size_t>(jcp.id) * jcp.ih * jcp.iw
            * jcp.ngroups * jcp.ic;
    const size_t src_g_stride = jcp.ic;
    // Wei Format: spatial-input_channels-groups-output_channels
    const size_t wei_g_stride = pd()->with_groups() ? jcp.oc : 0;

    // Dst Format: mb-spatial-groups-output_channels
    const size_t dst_mb_stride = static_cast<size_t>(jcp.od) * jcp.oh * jcp.ow
            * jcp.ngroups * jcp.oc;
    const size_t dst_g_stride = jcp.oc;
    const size_t dst_os_stride = jcp.ngroups * jcp.oc;

    src_data_t *__restrict col = scratchpad.get<src_data_t>(key_conv_gemm_col)
            + (ptrdiff_t)ithr * jcp.im2col_sz;
    src_data_t *__restrict imtr = scratchpad.get<src_data_t>(key_conv_gemm_imtr)
            + (ptrdiff_t)ithr * jcp.is * jcp.ic;
    acc_data_t *__restrict acc = scratchpad.get<acc_data_t>(key_conv_gemm_acc)
            + (ptrdiff_t)ithr * jcp.oh_block * jcp.ow_block * jcp.oc;

    const auto &post_ops = pd()->attr()->post_ops_;
    const bool do_sum = post_ops.contain(primitive_kind::sum, 0);
    const float sum_scale = do_sum ? post_ops.entry_[0].sum.scale : 0;

    dim_t g {0}, n {0}, ohb {0}, owb {0};
    dim_t start = 0, end = 0;

    const bool is_problem_3d = pd()->ndims() == 5;
    assert(IMPLICATION(is_problem_3d,
            jcp.oh_block == jcp.oh && jcp.ow_block == jcp.ow
                    && jcp.ic_block == jcp.ic));

    const dim_t nb_oh = div_up(jcp.oh, jcp.oh_block);
    const dim_t nb_ow = div_up(jcp.ow, jcp.ow_block);
    const dim_t work_amount = jcp.ngroups * jcp.mb * nb_oh * nb_ow;
    balance211(work_amount, nthr, ithr, start, end);
    nd_iterator_init(start, n, jcp.mb, g, jcp.ngroups, ohb, nb_oh, owb, nb_ow);

    if (jcp.im2col_sz && is_problem_3d) {
        // jit_gemm_convolution_utils::im2col_dt_3d() requires external
        // data initialization by zeroes
        // For performance reasons use uint16_t as a proxy for bfloat16_t
        uint16_t *__restrict col_r
                = reinterpret_cast<uint16_t *__restrict>(col);
        constexpr uint16_t zero_val = 0;

        PRAGMA_OMP_SIMD()
        for (ptrdiff_t i = 0; i < jcp.im2col_sz; i++)
            col_r[i] = zero_val;
    }
    for (dim_t iwork = start; iwork < end; ++iwork) {
        int oh = ohb * jcp.oh_block;
        int ow = owb * jcp.ow_block;
        const src_data_t *__restrict src
                = src_base + n * src_mb_stride + g * src_g_stride;
        const wei_data_t *__restrict wei = wei_base + g * wei_g_stride;

        const int h_step = nstl::min(jcp.oh_block, jcp.oh - oh);
        const int w_step = nstl::min(jcp.ow_block, jcp.ow - ow);
        if (jcp.im2col_sz && is_problem_3d)
            jit_gemm_convolution_utils::transpose_dt(jcp, src, imtr);

        for (int od = 0; od < jcp.od; od++) {
            dst_data_t *__restrict dst = dst_base + n * dst_mb_stride
                    + g * dst_g_stride
                    + ((od * jcp.oh + oh) * jcp.ow + ow) * dst_os_stride;
            if (jcp.im2col_sz) {
                if (is_problem_3d)
                    jit_gemm_convolution_utils::im2col_dt_3d<src_data_t,
                            src_data_t>(jcp, imtr, col, od);
                else
                    jit_gemm_convolution_utils::im2col_dt<src_data_t,
                            src_data_t>(
                            jcp, src, imtr, col, oh, h_step, ow, w_step);
            }

            const dim_t M = jcp.oc;
            const dim_t K = jcp.ks * jcp.ic;
            const dim_t N = h_step * w_step;
            const dim_t LDA = M * jcp.ngroups;
            const dim_t LDB = jcp.im2col_sz ? N : K * jcp.ngroups;
            const char *BT = jcp.im2col_sz ? "T" : "N";
            const float onef = 1.f;
            const float beta = this->beta_;
            const src_data_t *__restrict src_od
                    = src + od * jcp.oh * jcp.ow * jcp.ngroups * jcp.ic;
            const bool acc_needed = dst_data_type == data_type::bf16;
            status_t st = gemm_bf16bf16f32("N", BT, &M, &N, &K, &onef, wei,
                    &LDA, jcp.im2col_sz ? col : (src_data_t *)src_od, &LDB,
                    &beta, acc_needed ? acc : (float *)dst,
                    acc_needed ? &M : &LDA);
            if (st != status::success) return st;

            const bool do_postprocess = pd()->is_postprocess_required();
            if (do_postprocess) {
                parallel_nd_ext(jcp.nthr == 1 ? 0 : 1, N,
                        [&](size_t ithr, size_t nthr, size_t os) {
                    const float *__restrict acc_arr = acc + os * jcp.oc;
                    const float *__restrict bia_arr = (bia_base == nullptr)
                            ? nullptr
                            : bia_base + g * jcp.oc;
                    dst_data_t *__restrict dst_arr = dst + os * dst_os_stride;

                    (*pp_ker_)(dst_arr, acc_needed ? acc_arr : (float *)dst_arr,
                            bia_arr, sum_scale, jcp.oc,
                            post_ops_binary_rhs_arg_vec, dst_base, g *jcp.oc);
                });
            }
        }
        nd_iterator_step(n, jcp.mb, g, jcp.ngroups, ohb, nb_oh, owb, nb_ow);
    }
    return status::success;
}

template <data_type_t dst_data_type>
status_t gemm_bf16_convolution_fwd_t<dst_data_type>::execute_forward_ncsp(
        const exec_ctx_t &ctx) const {
    auto src = CTX_IN_MEM(const src_data_t *, DNNL_ARG_SRC);
    auto weights = CTX_IN_MEM(const wei_data_t *, DNNL_ARG_WEIGHTS);
    auto dst = CTX_OUT_MEM(dst_data_t *, DNNL_ARG_DST);
    const auto post_ops_binary_rhs_arg_vec
            = binary_injector::prepare_binary_args(
                    this->pd()->attr()->post_ops_, ctx);

    bool is_bf16_dst = dst_data_type == data_type::bf16;

    auto col = ctx.get_scratchpad_grantor().template get<src_data_t>(
            key_conv_gemm_col);
    acc_data_t *acc_base = is_bf16_dst
            ? ctx.get_scratchpad_grantor().template get<acc_data_t>(
                      key_conv_int_dat_in_acc_dt)
            : nullptr;

    const conv_gemm_conf_t &jcp = this->pd()->jcp_;
    const memory_desc_wrapper src_d(pd()->src_md());
    const memory_desc_wrapper dst_d(pd()->dst_md());

    src += src_d.off_l(0);
    dst += dst_d.off_l(0);

    float *bias = nullptr;
    if (jcp.with_bias) {
        if (pd()->desc()->bias_desc.data_type == data_type::bf16) {
            auto bias_in = CTX_IN_MEM(const bfloat16_t *, DNNL_ARG_BIAS);
            bias = ctx.get_scratchpad_grantor().template get<float>(
                    key_conv_bias_bf16_convert_wsp);
            cvt_bfloat16_to_float(bias, bias_in, jcp.ngroups * jcp.oc);
        } else {
            auto bias_in = CTX_IN_MEM(const float *, DNNL_ARG_BIAS);
            bias = const_cast<float *>(bias_in);
        }
    }

    const auto &post_ops = pd()->attr()->post_ops_;
    const bool do_sum = post_ops.contain(primitive_kind::sum, 0);
    const float sum_scale = do_sum ? post_ops.entry_[0].sum.scale : 0;

    const dim_t M = jcp.os * jcp.od;
    const size_t src_step = (size_t)jcp.ic * jcp.ih * jcp.iw * jcp.id;
    const size_t dst_step = (size_t)jcp.oc * M;
    const size_t weights_g_size = (size_t)jcp.ic * jcp.oc * jcp.ks;
    const size_t weights_oc_size = jcp.ic * jcp.ks;

    const dim_t LDB = weights_oc_size;
    const dim_t work_amount
            = (size_t)jcp.ngroups * jcp.mb * jcp.od * jcp.os_nb_block;
    const bool is_problem_3d = pd()->ndims() == 5;
    std::atomic<status_t> st(status::success);

    auto inner_ker = [&](const int ic, const int oc, const int groups,
                             const int od, const int spatial,
                             const src_data_t *src, const wei_data_t *weights,
                             src_data_t *col, dst_data_t *dst_im,
                             acc_data_t *acc, int ic_block, int oc_block) {
        const dim_t os_block = nstl::min(
                (dim_t)jcp.os_block, (dim_t)jcp.os - spatial * jcp.os_block);

        if (jcp.im2col_sz) {
            if (!is_problem_3d) {
                jit_gemm_convolution_utils::im2col<src_data_t>(jcp, src, col,
                        spatial * jcp.os_block, os_block, ic, ic_block);
            } else {
                assert(jcp.ic_block == jcp.ic);
                jit_gemm_convolution_utils::im2col_3d<src_data_t>(
                        jcp, src, col, od, spatial * jcp.os_block, os_block);
            }
        }

        const acc_data_t one = 1.0;
        const dim_t N = oc_block;
        const dim_t K = ic_block * jcp.ks;
        const dim_t m = os_block;
        const dim_t LDA = jcp.im2col_sz ? m : M;
        const dim_t LDC = is_bf16_dst ? m : M;
        const float beta = (ic == 0) ? this->beta_ : one;
        auto out_off = spatial * jcp.os_block + od * jcp.os;
        dst_data_t *dst_local = dst_im + out_off;

        status_t st_thr = gemm_bf16bf16f32("N", "N", &m, &N, &K, &one,
                jcp.im2col_sz ? col : src + ic * M + out_off, &LDA, weights,
                &LDB, &beta, acc, &LDC);

        if (st_thr != status::success) {
            st = st_thr;
            return;
        }

        if (this->pd()->is_postprocess_required() && ic + ic_block >= jcp.ic) {
            size_t acc_str = LDC;
            size_t dst_str = M;
            float *bias_ptr = bias ? bias + groups * jcp.oc + oc : nullptr;
            (*pp_ker_)(dst_local, acc, bias_ptr, sum_scale, dst_str, acc_str, m,
                    oc_block, post_ops_binary_rhs_arg_vec.data(), dst,
                    groups *jcp.oc + oc);
        }
    };

    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        src_data_t *_col = col + (ptrdiff_t)ithr * jcp.im2col_sz;
        if (is_problem_3d) {
            // jit_gemm_convolution_utils::im2col_3d() requires external
            // data initialization by zeroes
            for (ptrdiff_t i = 0; i < jcp.im2col_sz; i++)
                _col[i] = (src_data_t)0;
        }
        dim_t g {0}, n {0}, od {0}, nb_os {0};
        dim_t start = 0, end = 0;
        dim_t oc_start = 0, oc_end = 0;

        assert(jcp.loop_order == gemm_loop_lbr);
        balance2D(nthr, ithr, work_amount, start, end, jcp.oc, oc_start, oc_end,
                dim_t(jcp.nthr_oc));

        nd_iterator_init(start, g, jcp.ngroups, n, jcp.mb, od, jcp.od, nb_os,
                jcp.os_nb_block);
        for (dim_t iwork = start; iwork < end; ++iwork) {
            for_(dim_t oc = (dim_t)oc_start; oc < (dim_t)oc_end;
                    oc += jcp.oc_block)
            for (dim_t ic = 0; ic < jcp.ic; ic += jcp.ic_block) {
                const src_data_t *_src = src + (n * jcp.ngroups + g) * src_step;
                const wei_data_t *_weights = weights + g * weights_g_size
                        + oc * weights_oc_size + ic * jcp.ks;
                dst_data_t *_dst_im
                        = dst + (n * jcp.ngroups + g) * dst_step + oc * M;
                auto out_off = nb_os * jcp.os_block + od * jcp.os;
                dst_data_t *dst_local = _dst_im + out_off;
                const dim_t sizeof_cacheline_float = 16;
                acc_data_t *_acc = is_bf16_dst ? acc_base
                                + ithr
                                        * rnd_up(jcp.oc_block * jcp.os_block,
                                                sizeof_cacheline_float)
                                               : (acc_data_t *)dst_local;

                const dim_t ic_block = nstl::min(jcp.ic - ic, jcp.ic_block);
                const dim_t oc_block
                        = nstl::min(dim_t(oc_end) - oc, jcp.oc_block);

                inner_ker(ic, oc, g, od, nb_os, _src, _weights, _col, _dst_im,
                        _acc, ic_block, oc_block);
            }
            nd_iterator_step(g, jcp.ngroups, n, jcp.mb, od, jcp.od, nb_os,
                    jcp.os_nb_block);
        }
    });

    return st;
}

template <data_type_t diff_src_data_type>
status_t gemm_bf16_convolution_bwd_data_t<diff_src_data_type>::
        execute_backward_data_nspc(const exec_ctx_t &ctx) const {

    auto diff_dst_base = CTX_IN_MEM(const diff_dst_data_t *, DNNL_ARG_DIFF_DST);
    auto wei_base = CTX_IN_MEM(const wei_data_t *, DNNL_ARG_WEIGHTS);
    auto diff_src_base = CTX_OUT_MEM(diff_src_data_t *, DNNL_ARG_DIFF_SRC);

    const auto &scratchpad = ctx.get_scratchpad_grantor();
    const conv_gemm_conf_t &jcp = pd()->jcp_;

    std::atomic<status_t> st(status::success);
    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        status_t st_thr = execute_backward_data_thr_nspc(
                ithr, nthr, diff_src_base, wei_base, diff_dst_base, scratchpad);
        if (st_thr != status::success) st = st_thr;
    });

    return st;
}

template <data_type_t diff_src_data_type>
status_t gemm_bf16_convolution_bwd_data_t<
        diff_src_data_type>::execute_backward_data_thr_nspc(const int ithr,
        const int nthr, diff_src_data_t *diff_src_base,
        const wei_data_t *wei_base, const diff_dst_data_t *diff_dst_base,
        const memory_tracking::grantor_t &scratchpad) const {

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

    // Diff_dst Format: mb-spatial-groups-output_channels
    const size_t diff_dst_mb_stride = static_cast<size_t>(jcp.od) * jcp.oh
            * jcp.ow * jcp.ngroups * jcp.oc;
    const size_t diff_dst_g_stride = jcp.oc;

    // Wei Format: spatial-input_channels-groups-output_channels
    const size_t wei_g_stride = pd()->with_groups() ? jcp.oc : 0;

    // Diff_src Format: mb-spatial-groups-input_channels
    const size_t diff_src_mb_stride = static_cast<size_t>(jcp.id) * jcp.ih
            * jcp.iw * jcp.ngroups * jcp.ic;
    const size_t diff_src_g_stride = jcp.ic;
    const size_t diff_src_os_stride = jcp.ngroups * jcp.ic;

    // threads share work across mini-batch and groups
    const dim_t work_amount = jcp.ngroups * jcp.mb;

    acc_data_t *__restrict col = scratchpad.get<acc_data_t>(key_conv_gemm_col)
            + (ptrdiff_t)ithr * jcp.im2col_sz;
    acc_data_t *__restrict acc = scratchpad.get<acc_data_t>(key_conv_gemm_acc)
            + (ptrdiff_t)ithr * jcp.is * jcp.id * jcp.ic;

    dim_t n {0}, g {0};
    dim_t start = 0, end = 0;

    balance211(work_amount, nthr, ithr, start, end);
    nd_iterator_init(start, n, jcp.mb, g, jcp.ngroups);

    for (dim_t iwork = start; iwork < end; ++iwork) {
        const diff_dst_data_t *__restrict diff_dst = diff_dst_base
                + n * diff_dst_mb_stride + g * diff_dst_g_stride;
        const wei_data_t *__restrict wei = wei_base + g * wei_g_stride;
        diff_src_data_t *__restrict diff_src = diff_src_base
                + n * diff_src_mb_stride + g * diff_src_g_stride;

        const dim_t M = jcp.ks * jcp.ic;
        const dim_t N = jcp.os * jcp.od;
        const dim_t K = jcp.oc;

        const float onef = 1.0f, zerof = 0.0f;
        const dim_t LD = K * jcp.ngroups;

        status_t st = gemm_bf16bf16f32("T", "N", &M, &N, &K, &onef, wei, &LD,
                diff_dst, &LD, &zerof, jcp.im2col_sz ? col : acc, &M);
        if (st != status::success) return st;

        if (jcp.im2col_sz)
            jit_gemm_convolution_utils::col2im_dt<acc_data_t>(jcp, col, acc);

        const bool is_diff_src_bf16 = diff_src_data_type == data_type::bf16;

        if (is_diff_src_bf16 && jcp.ngroups == 1 && jcp.nthr != 1) {
            cvt_float_to_bfloat16((bfloat16_t *)diff_src, (const float *)acc,
                    static_cast<size_t>(jcp.is) * jcp.id * jcp.ic);
        } else if (is_diff_src_bf16) {
            parallel_nd_ext(jcp.nthr == 1 ? 0 : 1,
                    static_cast<size_t>(jcp.is) * jcp.id,
                    [&](size_t ithr, size_t nthr, size_t is) {
                diff_src_data_t *__restrict diff_src_loc
                        = diff_src + is * diff_src_os_stride;
                const acc_data_t *__restrict acc_loc = acc + is * jcp.ic;
                cvt_float_to_bfloat16((bfloat16_t *)diff_src_loc,
                        (const float *)acc_loc, jcp.ic);
            });
        } else {
            assert(diff_src_data_type == data_type::f32);
            parallel_nd_ext(jcp.nthr == 1 ? 0 : 1,
                    static_cast<size_t>(jcp.is) * jcp.id,
                    [&](size_t ithr, size_t nthr, size_t is) {
                diff_src_data_t *__restrict diff_src_loc
                        = diff_src + is * diff_src_os_stride;
                const acc_data_t *__restrict acc_loc = acc + is * jcp.ic;
                PRAGMA_OMP_SIMD()
                for (int ic = 0; ic < jcp.ic; ++ic)
                    diff_src_loc[ic] = acc_loc[ic];
            });
        }
        nd_iterator_step(n, jcp.mb, g, jcp.ngroups);
    }
    return status::success;
}

template <data_type_t diff_src_data_type>
status_t gemm_bf16_convolution_bwd_data_t<diff_src_data_type>::
        execute_backward_data_ncsp(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);

    auto col = ctx.get_scratchpad_grantor().template get<acc_data_t>(
            key_conv_gemm_col);
    acc_data_t *acc_base = diff_src_data_type == data_type::bf16
            ? ctx.get_scratchpad_grantor().template get<acc_data_t>(
                      key_conv_int_dat_in_acc_dt)
            : nullptr;

    const conv_gemm_conf_t &jcp = this->pd()->jcp_;

    const dim_t M = jcp.os * jcp.od;
    const size_t src_step = (size_t)jcp.ic * jcp.ih * jcp.iw * jcp.id;
    const size_t dst_step = (size_t)jcp.oc * M;
    const size_t weights_g_size = (size_t)jcp.ic * jcp.oc * jcp.ks;

    const dim_t m = jcp.os_block;
    const dim_t K = jcp.oc;
    const dim_t N = jcp.ic * jcp.ks;

    const dim_t work_amount = (size_t)jcp.ngroups * jcp.mb;
    const bool is_problem_3d = pd()->ndims() == 5;

    std::atomic<status_t> st(status::success);

    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        acc_data_t *_col = col + (ptrdiff_t)ithr * jcp.im2col_sz;

        dim_t g {0}, n {0};
        dim_t start = 0, end = 0;
        balance211(work_amount, nthr, ithr, start, end);
        nd_iterator_init(start, g, jcp.ngroups, n, jcp.mb);
        for (dim_t iwork = start; iwork < end; ++iwork) {

            diff_src_data_t *diff_src_local
                    = diff_src + (n * jcp.ngroups + g) * src_step;
            acc_data_t *acc = diff_src_data_type == data_type::bf16
                    ? acc_base + ithr * rnd_up(src_step, 16)
                    : (acc_data_t *)diff_src_local;

            if (is_problem_3d && jcp.im2col_sz > 0) {
                // jit_gemm_convolution_utils::col2im_3d() assumes that the
                // accumulator is initialized by zeroes
                for (size_t i = 0; i < src_step; i++)
                    acc[i] = (acc_data_t)0;
            }

            const wei_data_t *_weights = weights + g * weights_g_size;
            for_(int od = 0; od < jcp.od; ++od)
            for (int os_nb = 0; os_nb < jcp.os_nb_block; ++os_nb) {
                auto out_off = os_nb * m + od * jcp.os;
                const diff_dst_data_t *_diff_dst
                        = diff_dst + (n * jcp.ngroups + g) * dst_step + out_off;
                const dim_t os_block
                        = nstl::min((dim_t)jcp.os_block, jcp.os - os_nb * m);
                const dim_t LDC = jcp.im2col_sz ? os_block : M;

                const acc_data_t zero = 0.0, one = 1.0;
                status_t st_thr = gemm_bf16bf16f32("N", "T", &os_block, &N, &K,
                        &one, _diff_dst, &M, _weights, &N, &zero,
                        jcp.im2col_sz ? _col : acc + out_off, &LDC);

                if (st_thr != status::success) {
                    st = st_thr;
                    return;
                }

                if (jcp.im2col_sz) {
                    if (!is_problem_3d)
                        jit_gemm_convolution_utils::col2im(
                                jcp, _col, acc, os_nb * jcp.os_block, os_block);
                    else
                        jit_gemm_convolution_utils::col2im_3d(jcp, _col, acc,
                                od, os_nb * jcp.os_block, os_block);
                }
            }
            if (diff_src_data_type == data_type::bf16) {
                size_t spatial_size = (size_t)jcp.ih * jcp.iw * jcp.id;
                store_bfloat16_in_parallel((bfloat16_t *)diff_src_local,
                        (const float *)acc, jcp.ic, spatial_size,
                        jcp.nthr == 1);
            }
            nd_iterator_step(g, jcp.ngroups, n, jcp.mb);
        }
    });

    return st;
}

template <data_type_t diff_wei_data_type>
void gemm_bf16_convolution_bwd_weights_t<
        diff_wei_data_type>::bf16_bwd_weights_reduction_par_nspc(int ithr_mb,
        int nthr_mb, size_t g_start, size_t g_end, const conv_gemm_conf_t &jcp,
        const acc_data_t *weights_reduce_base,
        diff_wei_data_t *weights_base) const {
    assert(nthr_mb > 1); // no reduction for nthr_mb == 1

    const bool is_bf16_out = diff_wei_data_type == data_type::bf16;
    const dim_t weights_g_size = jcp.oc;
    dim_t weights_start {0}, weights_end {0};
    balance211(jcp.ks * jcp.ic, nthr_mb, ithr_mb, weights_start, weights_end);

    for (auto tidx = 1; tidx < nthr_mb; ++tidx) {
        const acc_data_t *ws_base
                = weights_reduce_base + tidx * weights_g_size * jcp.ks * jcp.ic;
        for_(auto w = weights_start; w < weights_end; ++w)
        for (auto g = g_start; g < g_end; ++g) {
            const acc_data_t *ws_ptr = ws_base + w * jcp.oc;
            float *wei_reduced = is_bf16_out
                    ? (float *)weights_reduce_base + w * jcp.oc
                    : (float *)weights_base + (w * jcp.ngroups + g) * jcp.oc;
            if (is_bf16_out && tidx == nthr_mb - 1) {
                // the last iteration for bfloat16 requires conversion
                // and store to diff_weights array
                diff_wei_data_t *dwei_ptr
                        = weights_base + (w * jcp.ngroups + g) * jcp.oc;
                add_floats_and_cvt_to_bfloat16(
                        (bfloat16_t *)(dwei_ptr), wei_reduced, ws_ptr, jcp.oc);
            } else {
                acc_ker_->accumulate(wei_reduced, ws_ptr, jcp.oc);
            }
        }
    }
}

template <data_type_t diff_wei_data_type>
void gemm_bf16_convolution_bwd_weights_t<
        diff_wei_data_type>::bf16_bwd_weights_reduction_par_ncsp(int ithr_mb,
        int nthr_mb, const conv_gemm_conf_t &jcp,
        const acc_data_t *weights_reduce_base,
        diff_wei_data_t *weights_base) const {
    assert(nthr_mb > 1); // no reduction for nthr_mb == 1

    const bool is_bf16_out = diff_wei_data_type == data_type::bf16;
    const size_t weights_g_size = (size_t)jcp.ic * jcp.oc * jcp.ks;
    size_t weights_start {0}, weights_end {0};
    balance211(weights_g_size, nthr_mb, ithr_mb, weights_start, weights_end);

    if (weights_start >= weights_end) return; // nothing to do

    size_t acc_size = weights_end - weights_start;
    float *wei_reduced = is_bf16_out
            ? (float *)weights_reduce_base + weights_start
            : (float *)weights_base + weights_start;
    if (!is_bf16_out) {
        // f32 diff_weights require initialization by weights_reduce
        // for thr_mb = 0
        for (size_t i = 0; i < acc_size; i++)
            wei_reduced[i] = ((float *)weights_reduce_base + weights_start)[i];
    }

    for (int thr_mb = 1; thr_mb < nthr_mb; ++thr_mb) {
        float *wei_to_reduce = (float *)weights_reduce_base
                + thr_mb * weights_g_size + weights_start;

        if (is_bf16_out && thr_mb == nthr_mb - 1)
            // the last iteration for bfloat16 requires conversion
            // and store to diff_weights array
            add_floats_and_cvt_to_bfloat16(
                    (bfloat16_t *)(weights_base + weights_start), wei_reduced,
                    wei_to_reduce, acc_size);
        else
            acc_ker_->accumulate(wei_reduced, wei_to_reduce, acc_size);
    }
}

template <data_type_t diff_wei_data_type>
status_t gemm_bf16_convolution_bwd_weights_t<diff_wei_data_type>::
        execute_backward_weights_nspc(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_wei_data_t *, DNNL_ARG_DIFF_WEIGHTS);

    auto col = ctx.get_scratchpad_grantor().template get<src_data_t>(
            key_conv_gemm_col);
    auto wei_reduction = ctx.get_scratchpad_grantor().template get<acc_data_t>(
            key_conv_wei_reduction);
    const conv_gemm_conf_t &jcp = this->pd()->jcp_;

    acc_data_t *acc_base = diff_wei_data_type == data_type::bf16
            ? ctx.get_scratchpad_grantor().template get<acc_data_t>(
                      key_conv_int_dat_in_acc_dt)
            : (acc_data_t *)diff_weights;

    float *diff_bias = nullptr;
    if (jcp.with_bias) {
        if (pd()->desc()->diff_bias_desc.data_type == data_type::bf16)
            diff_bias = ctx.get_scratchpad_grantor().template get<float>(
                    key_conv_bias_bf16_convert_wsp);
        else
            diff_bias = CTX_OUT_MEM(float *, DNNL_ARG_DIFF_BIAS);
    }

    const dim_t K = jcp.os * static_cast<size_t>(jcp.od);
    const size_t src_step
            = static_cast<size_t>(jcp.ic) * jcp.ih * jcp.iw * jcp.id;
    const size_t dst_step = jcp.oc * K;
    const size_t weights_g_size = jcp.oc;

    const dim_t k = jcp.os;
    const dim_t M = jcp.oc;
    const dim_t N = static_cast<dim_t>(jcp.ic) * jcp.ks;
    const dim_t LDB = jcp.ngroups * jcp.oc;
    const dim_t LDA = jcp.im2col_sz ? jcp.oh * jcp.ow : jcp.ngroups * jcp.ic;
    const bool is_problem_3d = pd()->ndims() == 5;

    std::atomic<status_t> st(status::success);

    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        int ithr_g, nthr_g, ithr_mb, nthr_mb;
        size_t g_start {0}, g_end {0}, mb_start {0}, mb_end {0};

        const int mb_for_balance = jcp.need_wei_reduction ? jcp.mb : 1;
        jit_gemm_convolution_utils::bwd_weights_balance(ithr, nthr, jcp.ngroups,
                mb_for_balance, ithr_g, nthr_g, ithr_mb, nthr_mb);

        assert(IMPLICATION(!jcp.need_wei_reduction, nthr_mb == 1));

        const int need_reduction = nthr_mb != 1;
        src_data_t *__restrict imtr
                = ctx.get_scratchpad_grantor().template get<src_data_t>(
                          key_conv_gemm_imtr)
                + (ptrdiff_t)ithr * jcp.id * jcp.ic * jcp.is;

        if (ithr_g != -1 && ithr_mb != -1) {
            balance211((size_t)jcp.ngroups, nthr_g, ithr_g, g_start, g_end);
            balance211((size_t)jcp.mb, nthr_mb, ithr_mb, mb_start, mb_end);

            assert(IMPLICATION((g_end - g_start) > 1, need_reduction == 0));

            src_data_t *_col = col + (ptrdiff_t)ithr * jcp.im2col_sz;
            if (is_problem_3d) {
                // jit_gemm_convolution_utils::im2col_3d() requires external
                // data initialization by zeroes
                // For performance reasons use uint16_t as proxy for bfloat16_t
                uint16_t *__restrict _col_r
                        = reinterpret_cast<uint16_t *__restrict>(_col);
                constexpr uint16_t zero_val = 0;

                PRAGMA_OMP_SIMD()
                for (ptrdiff_t i = 0; i < jcp.im2col_sz; i++)
                    _col_r[i] = zero_val;
            }

            acc_data_t *weights_reduce_base = wei_reduction
                    + ithr_g * nthr_mb * weights_g_size * jcp.ks * jcp.ic;
            acc_data_t *weights_reduce = weights_reduce_base
                    + ithr_mb * weights_g_size * jcp.ks * jcp.ic;

            const bool use_diff_wei
                    = ithr_mb == 0 && diff_wei_data_type == data_type::f32;
            for (size_t g = g_start; g < g_end; ++g) {
                acc_data_t *_diff_weights = use_diff_wei
                        ? (acc_data_t *)diff_weights + g * weights_g_size
                        : need_reduction ? weights_reduce
                                         : acc_base + g * weights_g_size;
                const dim_t LDC = use_diff_wei ? jcp.ngroups * jcp.oc
                        : need_reduction       ? jcp.oc
                                               : jcp.ngroups * jcp.oc;
                for (size_t mb = mb_start; mb < mb_end; ++mb) {
                    const src_data_t *_src
                            = src + mb * jcp.ngroups * src_step + g * jcp.ic;
                    if (jcp.im2col_sz && is_problem_3d)
                        jit_gemm_convolution_utils::transpose_dt(
                                jcp, _src, imtr);
                    for (int od = 0; od < jcp.od; ++od) {
                        const diff_dst_data_t *_diff_dst = diff_dst
                                + mb * jcp.ngroups * dst_step
                                + od * k * jcp.ngroups * jcp.oc + g * jcp.oc;

                        if (jcp.im2col_sz) {
                            if (is_problem_3d)
                                jit_gemm_convolution_utils::im2col_dt_3d<
                                        src_data_t, src_data_t>(
                                        jcp, imtr, _col, od);
                            else
                                jit_gemm_convolution_utils::im2col_dt<
                                        src_data_t, src_data_t>(jcp, _src, imtr,
                                        _col, 0, jcp.oh, 0, jcp.ow);
                        }
                        const float zero = 0.0f, one = 1.0f;
                        status_t st_thr = gemm_bf16bf16f32("N",
                                jcp.im2col_sz ? "N" : "T", &M, &N, &k, &one,
                                _diff_dst, &LDB,
                                jcp.im2col_sz
                                        ? _col
                                        : _src + od * k * jcp.ngroups * jcp.ic,
                                &LDA, mb == mb_start && od == 0 ? &zero : &one,
                                _diff_weights, &LDC);
                        if (st_thr != status::success) {
                            st = st_thr;
                            // Finish the loops early if failure occured.
                            g = g_end;
                            mb = mb_end;
                            od = jcp.od;
                        }
                    }
                }
            }
            if (need_reduction && dnnl_thr_syncable()) {
                dnnl_thr_barrier();
                if (st != status::success) return;
                bf16_bwd_weights_reduction_par_nspc(ithr_mb, nthr_mb, g_start,
                        g_end, jcp, weights_reduce_base, diff_weights);
            } else if (diff_wei_data_type == data_type::bf16
                    && g_end > g_start) {
                cvt_acc_to_dst(jcp, g_start, g_end, (const float *)acc_base,
                        (bfloat16_t *)diff_weights);
            }
        } else {
            if (need_reduction && dnnl_thr_syncable()) dnnl_thr_barrier();
        }
    });

    if (jcp.need_wei_reduction && !dnnl_thr_syncable()) {
        parallel(jcp.nthr, [&](const int ithr, const int nthr) {
            int ithr_g, nthr_g, ithr_mb, nthr_mb;
            size_t g_start {0}, g_end {0}, mb_start {0}, mb_end {0};

            const int mb_for_balance = jcp.need_wei_reduction ? jcp.mb : 1;
            jit_gemm_convolution_utils::bwd_weights_balance(ithr, nthr,
                    jcp.ngroups, mb_for_balance, ithr_g, nthr_g, ithr_mb,
                    nthr_mb);

            assert(IMPLICATION(!jcp.need_wei_reduction, nthr_mb == 1));
            const int need_reduction = nthr_mb != 1;

            if (need_reduction && ithr_g != -1 && ithr_mb != -1) {
                balance211((size_t)jcp.ngroups, nthr_g, ithr_g, g_start, g_end);
                balance211((size_t)jcp.mb, nthr_mb, ithr_mb, mb_start, mb_end);

                assert(IMPLICATION((g_end - g_start) > 1, need_reduction == 0));

                acc_data_t *weights_reduce_base = wei_reduction
                        + ithr_g * nthr_mb * weights_g_size * jcp.ic * jcp.ks;

                bf16_bwd_weights_reduction_par_nspc(ithr_mb, nthr_mb, g_start,
                        g_end, jcp, weights_reduce_base, diff_weights);
            }
        });
    }

    if (jcp.with_bias) {
        parallel_nd(jcp.ngroups, jcp.oc, [&](dim_t g, dim_t oc) {
            acc_data_t db = 0;
            const dim_t offset_base = g * jcp.oc + oc;
            for_(dim_t mb = 0; mb < jcp.mb; ++mb)
            for_(dim_t od = 0; od < jcp.od; ++od)
            for (dim_t oh = 0; oh < jcp.oh; ++oh) {
                const dim_t width_stride = jcp.ngroups * jcp.oc;
                const diff_dst_data_t *__restrict diff_dst_arr = diff_dst
                        + offset_base
                        + ((mb * jcp.od + od) * jcp.oh + oh) * jcp.ow
                                * width_stride;

                PRAGMA_OMP_SIMD(reduction(+ : db))
                for (dim_t ow = 0; ow < jcp.ow; ++ow) {
                    db += diff_dst_arr[ow * width_stride];
                }
            }
            diff_bias[g * jcp.oc + oc] = db;
        });

        if (pd()->desc()->diff_bias_desc.data_type == data_type::bf16) {
            auto diff_bias_in = CTX_OUT_MEM(bfloat16_t *, DNNL_ARG_DIFF_BIAS);
            cvt_float_to_bfloat16(
                    diff_bias_in, diff_bias, jcp.ngroups * jcp.oc);
        }
    }
    return st;
}

template <data_type_t diff_wei_data_type>
status_t gemm_bf16_convolution_bwd_weights_t<diff_wei_data_type>::
        execute_backward_weights_ncsp(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_wei_data_t *, DNNL_ARG_DIFF_WEIGHTS);

    auto col = ctx.get_scratchpad_grantor().template get<src_data_t>(
            key_conv_gemm_col);
    auto wei_reduction = ctx.get_scratchpad_grantor().template get<acc_data_t>(
            key_conv_wei_reduction);

    const conv_gemm_conf_t &jcp = this->pd()->jcp_;

    acc_data_t *acc_base = diff_wei_data_type == data_type::bf16
            ? ctx.get_scratchpad_grantor().template get<acc_data_t>(
                      key_conv_int_dat_in_acc_dt)
            : (acc_data_t *)diff_weights;

    float *diff_bias = nullptr;
    if (jcp.with_bias) {
        if (pd()->desc()->diff_bias_desc.data_type == data_type::bf16)
            diff_bias = ctx.get_scratchpad_grantor().template get<float>(
                    key_conv_bias_bf16_convert_wsp);
        else
            diff_bias = CTX_OUT_MEM(float *, DNNL_ARG_DIFF_BIAS);
    }

    const dim_t K = jcp.os * jcp.od;
    const size_t src_step = (size_t)jcp.ic * jcp.ih * jcp.iw * jcp.id;
    const size_t dst_step = (size_t)jcp.oc * K;
    const size_t weights_g_size = (size_t)jcp.ic * jcp.oc * jcp.ks;

    const dim_t k = jcp.os_block;
    const dim_t N = jcp.oc;
    const dim_t M = jcp.ic * jcp.ks;
    const bool is_problem_3d = pd()->ndims() == 5;

    std::atomic<status_t> st(status::success);
    parallel(jcp.nthr, [&](const int ithr, const int nthr) {
        int ithr_g, nthr_g, ithr_mb, nthr_mb;
        size_t g_start {0}, g_end {0}, mb_start {0}, mb_end {0};

        const int mb_for_balance = jcp.need_wei_reduction ? jcp.mb : 1;
        jit_gemm_convolution_utils::bwd_weights_balance(ithr, nthr, jcp.ngroups,
                mb_for_balance, ithr_g, nthr_g, ithr_mb, nthr_mb);

        assert(IMPLICATION(!jcp.need_wei_reduction, nthr_mb == 1));
        const int need_reduction = nthr_mb != 1;

        if (ithr_g != -1 && ithr_mb != -1) {
            balance211((size_t)jcp.ngroups, nthr_g, ithr_g, g_start, g_end);
            balance211((size_t)jcp.mb, nthr_mb, ithr_mb, mb_start, mb_end);

            assert(IMPLICATION((g_end - g_start) > 1, need_reduction == 0));

            src_data_t *_col = col + (ptrdiff_t)ithr * jcp.im2col_sz;
            // non-blocked jit_gemm_convolution_utils::im2col_3d() requires
            // external data initialization by zeroes
            const bool outer_padding = jcp.os_nb_block == 1;
            if (outer_padding && is_problem_3d) {
                for (ptrdiff_t i = 0; i < jcp.im2col_sz; i++)
                    _col[i] = (src_data_t)0;
            }

            acc_data_t *weights_reduce_base
                    = wei_reduction + ithr_g * nthr_mb * weights_g_size;
            acc_data_t *weights_reduce
                    = weights_reduce_base + ithr_mb * weights_g_size;

            for (size_t g = g_start; g < g_end; ++g) {
                acc_data_t *acc = need_reduction
                        ? weights_reduce
                        : (acc_base + g * weights_g_size);
                for (size_t mb = mb_start; mb < mb_end; ++mb) {
                    const src_data_t *_src
                            = src + (mb * jcp.ngroups + g) * src_step;
                    for_(int od = 0; od < jcp.od; ++od)
                    for (int os_nb = 0; os_nb < jcp.os_nb_block; ++os_nb) {
                        auto out_off = os_nb * k + od * jcp.os;
                        const dim_t os_block = nstl::min(
                                (dim_t)jcp.os_block, jcp.os - os_nb * k);
                        const diff_dst_data_t *_diff_dst = diff_dst
                                + (mb * jcp.ngroups + g) * dst_step + out_off;

                        if (jcp.im2col_sz) {
                            if (!is_problem_3d)
                                jit_gemm_convolution_utils::im2col<src_data_t>(
                                        jcp, _src, _col, os_nb * jcp.os_block,
                                        os_block, 0, jcp.ic);
                            else
                                jit_gemm_convolution_utils::im2col_3d<
                                        src_data_t>(jcp, _src, _col, od,
                                        os_nb * jcp.os_block, os_block);
                        }

                        const dim_t LDA = jcp.im2col_sz ? os_block : K;
                        const acc_data_t zero = 0.0, one = 1.0;
                        status_t st_thr = gemm_bf16bf16f32("T", "N", &M, &N,
                                &os_block, &one,
                                jcp.im2col_sz ? _col : _src + out_off, &LDA,
                                _diff_dst, &K,
                                mb == mb_start && os_nb == 0 && od == 0 ? &zero
                                                                        : &one,
                                acc, &M);

                        if (st_thr != status::success) {
                            st = st_thr;
                            // Finish the loops early if failure occured.
                            g = g_end;
                            mb = mb_end;
                            od = jcp.od;
                            os_nb = jcp.os_nb_block;
                        }
                    }
                }
            }
            if (need_reduction && dnnl_thr_syncable()) {
                dnnl_thr_barrier();
                if (st != status::success) return;
                diff_wei_data_t *weights_base
                        = diff_weights + g_start * weights_g_size;
                bf16_bwd_weights_reduction_par_ncsp(ithr_mb, nthr_mb, jcp,
                        weights_reduce_base, weights_base);
            } else if (diff_wei_data_type == data_type::bf16
                    && g_end > g_start) {
                const size_t weights_g_size = (size_t)jcp.ic * jcp.oc * jcp.ks;
                const size_t work_size = (g_end - g_start) * weights_g_size;
                bfloat16_t *diff_weights_local
                        = (bfloat16_t *)diff_weights + g_start * weights_g_size;
                const float *acc_local
                        = (const float *)acc_base + g_start * weights_g_size;
                store_bfloat16_in_parallel(diff_weights_local, acc_local,
                        work_size, 1, jcp.nthr == 1);
            }
        } else {
            if (need_reduction && dnnl_thr_syncable()) dnnl_thr_barrier();
        }
    });

    if (st != status::success) return st;

    if (jcp.need_wei_reduction && !dnnl_thr_syncable()) {
        parallel(jcp.nthr, [&](const int ithr, const int nthr) {
            int ithr_g, nthr_g, ithr_mb, nthr_mb;
            size_t g_start {0}, g_end {0}, mb_start {0}, mb_end {0};

            const int mb_for_balance = jcp.need_wei_reduction ? jcp.mb : 1;
            jit_gemm_convolution_utils::bwd_weights_balance(ithr, nthr,
                    jcp.ngroups, mb_for_balance, ithr_g, nthr_g, ithr_mb,
                    nthr_mb);

            assert(IMPLICATION(!jcp.need_wei_reduction, nthr_mb == 1));
            const int need_reduction = nthr_mb != 1;

            if (need_reduction && ithr_g != -1 && ithr_mb != -1) {
                balance211((size_t)jcp.ngroups, nthr_g, ithr_g, g_start, g_end);
                balance211((size_t)jcp.mb, nthr_mb, ithr_mb, mb_start, mb_end);

                assert(IMPLICATION((g_end - g_start) > 1, need_reduction == 0));

                acc_data_t *weights_reduce_base
                        = wei_reduction + ithr_g * nthr_mb * weights_g_size;

                diff_wei_data_t *weights_base
                        = diff_weights + g_start * weights_g_size;
                bf16_bwd_weights_reduction_par_ncsp(ithr_mb, nthr_mb, jcp,
                        weights_reduce_base, weights_base);
            }
        });
    }

    if (jcp.with_bias) {
        parallel_nd(jcp.ngroups, jcp.oc, [&](size_t g, size_t oc) {
            acc_data_t db = 0;
            dim_t offset_ = g * dst_step + oc * K;
            for (dim_t mb = 0; mb < jcp.mb; ++mb) {
                dim_t offset = offset_ + mb * jcp.ngroups * dst_step;
                for_(dim_t od = 0; od < jcp.od; ++od)
                for (dim_t oh = 0; oh < jcp.oh; ++oh) {
                    PRAGMA_OMP_SIMD(reduction(+ : db) linear(offset))
                    for (dim_t ow = 0; ow < jcp.ow; ++ow) {
                        db += diff_dst[offset];
                        offset++;
                    }
                }
            }
            diff_bias[g * jcp.oc + oc] = db;
        });

        if (pd()->desc()->diff_bias_desc.data_type == data_type::bf16) {
            auto diff_bias_in = CTX_OUT_MEM(bfloat16_t *, DNNL_ARG_DIFF_BIAS);
            cvt_float_to_bfloat16(
                    diff_bias_in, diff_bias, jcp.ngroups * jcp.oc);
        }
    }

    return st;
}

template struct gemm_bf16_convolution_fwd_t<data_type::f32>;
template struct gemm_bf16_convolution_fwd_t<data_type::bf16>;
template struct gemm_bf16_convolution_bwd_data_t<data_type::f32>;
template struct gemm_bf16_convolution_bwd_data_t<data_type::bf16>;
template struct gemm_bf16_convolution_bwd_weights_t<data_type::f32>;
template struct gemm_bf16_convolution_bwd_weights_t<data_type::bf16>;

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