merge from develop and revisison for pr#881

include/ck/utility/type_convert.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/f8_utils.hpp"
+#include "ck/utility/random_gen.hpp"
+
+namespace ck {
+
+// Convert X to Y
+template <typename Y, typename X>
+__host__ __device__ constexpr Y type_convert(X x)
+{
+    static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
+
+    return static_cast<Y>(x);
+}
+
+// convert bfp16 to fp32
+template <>
+inline __host__ __device__ constexpr float type_convert<float, bhalf_t>(bhalf_t x)
+{
+    union
+    {
+        uint32_t int32;
+        float fp32;
+    } u = {uint32_t(x) << 16};
+
+    return u.fp32;
+}
+
+// convert fp32 to bfp16
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    return uint16_t(u.int32 >> 16);
+}
+
+// convert bfp16 to fp16 via fp32
+template <>
+inline __host__ __device__ constexpr half_t type_convert<half_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<half_t>(x_fp32);
+}
+
+// convert fp16 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert bfp16 to int8 via fp32
+template <>
+inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<int8_t>(x_fp32);
+}
+
+// convert int8 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert fp32 to fp8
+template <>
+inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
+    constexpr uint32_t rng           = 0;
+    return utils::cast_to_f8<float, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp8 to fp32
+template <>
+inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    return utils::cast_from_f8<float, negative_zero_nan>(x);
+}
+
+// convert fp16 to fp8
+template <>
+inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
+    constexpr uint32_t rng           = 0;
+    return utils::cast_to_f8<half_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp8 to fp16
+template <>
+inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    return utils::cast_from_f8<half_t, negative_zero_nan>(x);
+}
+
+// Declare a template function for bf16 conversion using RTN
+template <typename Y, typename X>
+__host__ __device__ constexpr Y bf16_convert_rtn(X x);
+
+// Convert fp32 to bf16 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    // When the exponent bits are not all 1s, then the value is zero, normal,
+    // or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
+    // 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
+    // This causes the bfloat16's mantissa to be incremented by 1 if the 16
+    // least significant bits of the float mantissa are greater than 0x8000,
+    // or if they are equal to 0x8000 and the least significant bit of the
+    // bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
+    // the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
+    // has the value 0x7f, then incrementing it causes it to become 0x00 and
+    // the exponent is incremented by one, which is the next higher FP value
+    // to the unrounded bfloat16 value. When the bfloat16 value is subnormal
+    // with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
+    // to a normal value with an exponent of 0x01 and a mantissa of 0x00.
+    // When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
+    // incrementing it causes it to become an exponent of 0xFF and a mantissa
+    // of 0x00, which is Inf, the next higher value to the unrounded value.
+    bool flag0 = ~u.int32 & 0x7f800000;
+
+    // When all of the exponent bits are 1, the value is Inf or NaN.
+    // Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
+    // mantissa bit. Quiet NaN is indicated by the most significant mantissa
+    // bit being 1. Signaling NaN is indicated by the most significant
+    // mantissa bit being 0 but some other bit(s) being 1. If any of the
+    // lower 16 bits of the mantissa are 1, we set the least significant bit
+    // of the bfloat16 mantissa, in order to preserve signaling NaN in case
+    // the bfloat16's mantissa bits are all 0.
+    bool flag1 = !flag0 && (u.int32 & 0xffff);
+
+    u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
+    u.int32 |= flag1 ? 0x10000 : 0x0;                      // Preserve signaling NaN
+
+    return uint16_t(u.int32 >> 16);
+}
+
+// convert fp16 to bfp16 via fp32 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return bf16_convert_rtn<bhalf_t>(x_fp32);
+}
+
+// Declare a template function for fp8 conversion using SR
+template <typename Y, typename X>
+__host__ __device__ constexpr Y f8_convert_sr(X x);
+
+// convert fp32 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
+    constexpr int seed               = 42;
+    // as thread id is not available on host, use 0 for prn generation
+    uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::cast_to_f8<float, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp16 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
+    constexpr int seed               = 42;
+    // as thread id is not available on host, use 0 for prn generation
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::cast_to_f8<half_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+} // namespace ck

include/ck/utility/workgroup_barrier.hpp

+#pragma once
+#include <hip/hip_runtime.h>
+#include <stdint.h>
+
+namespace ck {
+struct workgroup_barrier
+{
+    __device__ workgroup_barrier(uint32_t* ptr) : base_ptr(ptr) {}
+
+    __device__ uint32_t ld(uint32_t offset)
+    {
+#if 0
+        float d = llvm_amdgcn_raw_buffer_load_fp32(
+                        amdgcn_make_buffer_resource(base_ptr),
+                        0,
+                        offset,
+                        AMDGCN_BUFFER_GLC);
+        union cvt {
+            float f32;
+            uint32_t u32;
+        };
+        cvt x;
+        x.f32 = d;
+        return x.u32;
+#endif
+        return __atomic_load_n(base_ptr + offset, __ATOMIC_RELAXED);
+    }
+
+    __device__ void wait_eq(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) != value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_lt(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) < value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_set(uint32_t offset, uint32_t compare, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(atomicCAS(base_ptr + offset, compare, value) != compare) {}
+        }
+        __syncthreads();
+    }
+
+    // enter critical zoon, assume buffer is zero when launch kernel
+    __device__ void aquire(uint32_t offset) { wait_set(offset, 0, 1); }
+
+    // exit critical zoon, assume buffer is zero when launch kernel
+    __device__ void release(uint32_t offset) { wait_set(offset, 1, 0); }
+
+    __device__ void inc(uint32_t offset)
+    {
+        __syncthreads();
+        if(threadIdx.x == 0)
+        {
+            atomicAdd(base_ptr + offset, 1);
+        }
+    }
+
+    uint32_t* base_ptr;
+};
+} // namespace ck

include/ck/utility/workgroup_synchronization.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "ck/host_utility/hip_check_error.hpp"
+
+namespace ck {
+
+// Initialization flag of Barrier object, can be any value except for zero
+static constexpr int BarrierInitFlag = 0x7856;
+
+// 1) only the first thread-block in the synchronizaton group is supposed to call this function. It
+// is the responsibility of the user to ensure the two integer values in p_control_bits are zeros
+// before calling gms_init().
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_init(int NumWarps, int* p_control_bits)
+{
+    union
+    {
+        int two32[2];
+        unsigned long one64;
+    } regs;
+
+    regs.two32[0] = BarrierInitFlag;
+    regs.two32[1] = NumWarps;
+
+    if(threadIdx.x == 0)
+        atomicCAS(reinterpret_cast<unsigned long*>(p_control_bits), 0, regs.one64);
+};
+
+// all the workgroups in the synchronization group is supposed to call this function
+static __device__ void gms_barrier(int* p_control_bits)
+{
+    constexpr int mask = warpSize - 1;
+
+    if((threadIdx.x & mask) == 0)
+    {
+        // ensure the barrier object is initialized
+        do
+        {
+            const int r0 = __atomic_load_n(&p_control_bits[0], __ATOMIC_RELAXED);
+
+            if(r0 == BarrierInitFlag)
+                break;
+
+        } while(true);
+
+        // go ahead toward the barrier line
+        atomicSub(&p_control_bits[1], 1);
+
+        // wait until all warps have arrived
+        do
+        {
+            const int r1 = __atomic_load_n(&p_control_bits[1], __ATOMIC_RELAXED);
+
+            if(r1 == 0)
+                break;
+
+        } while(true);
+    };
+};
+
+// 1) Only the first thread-block in the synchronizaton group is supposed to call this function.
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_reset(int* p_control_bits)
+{
+    // reset the barrier object
+    if(threadIdx.x == 0)
+        (void)atomicCAS(&p_control_bits[0], BarrierInitFlag, 0);
+};
+
+} // namespace ck

include/ck/version.h.in

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+
+/* the configured version and settings for miopen- Composable Kernel */
+
+#ifndef CK_VERSION_H_
+#define CK_VERSION_H_
+
+// clang-format off
+#define CK_VERSION @CMAKE_PROJECT_VERSION@
+#define CK_VERSION_MAJOR @CMAKE_PROJECT_VERSION_MAJOR@
+#define CK_VERSION_MINOR @CMAKE_PROJECT_VERSION_MINOR@
+#define CK_VERSION_PATCH @CMAKE_PROJECT_VERSION_PATCH@
+#define CK_COMMIT_ID @COMMIT_ID@
+// clang-format on
+
+#endif

library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// dinput descriptor in [N, C, Do, Ho, Wo] order
+// doutput descriptor in [N, C, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename DInDataType,
+          typename DOutDataType,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceAvgPoolBwd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<DInDataType>& dinput,
+                 const Tensor<DOutDataType>& doutput,
+                 std::vector<ck::index_t> window_spatial_lengths,
+                 std::vector<ck::index_t> window_strides,
+                 std::vector<ck::index_t> window_dilations,
+                 std::vector<ck::index_t> dinput_left_pads,
+                 std::vector<ck::index_t> dinput_right_pads)
+            : dinput_{dinput},
+              doutput_{doutput},
+              window_spatial_lengths_{window_spatial_lengths},
+              window_strides_{window_strides},
+              window_dilations_{window_dilations},
+              in_left_pads_{dinput_left_pads},
+              in_right_pads_{dinput_right_pads}
+        {
+        }
+
+        Tensor<DInDataType>& dinput_;
+        const Tensor<DOutDataType>& doutput_;
+
+        std::vector<ck::index_t> window_spatial_lengths_;
+        std::vector<index_t> window_strides_;
+        std::vector<index_t> window_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceAvgPoolBwd::Argument;
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 1, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            // Let input = x, outpu = y
+            // shape of x = [10], y = [6]
+            // window_size = 5, pad = 0, stride = 1, dilation = 1
+            // Forward:
+            // y0 = 1/5 * (x0 + x1 + x2 + x3 + x4)
+            // y1 = 1/5 * (x1 + x2 + x3 + x4 + x5)
+            // ...
+            // y5 = 1/5 * (x5 + x6 + x7 + x8 + x9)
+            // y6 = 1/5 * (x6 + x7 + x8 + x9)
+            // ...
+            // y9 = 1/5 * (x9)
+
+            // Backward:
+            // shape of dy = [6], dx = [10]
+            // dx0 = 1/5 * dy0
+            // dx1 = 1/5 * (dy0 + dy1)
+            // dx2 = 1/5 * (dy0 + dy1 + dy2)
+            // ...
+            // dx4 = 1/5 * (dy0 + dy1 + dy2 + dy3 + dy4)
+            // dx5 = 1/5 * (dy1 + dy2 + dy3 + dy4 + dy5)
+            // ...
+            // dx9 = 1/5 * (dy5 + dy6 + dy7 + dy8 + dy9)
+
+            auto f_ncw = [&](auto n, auto c, auto wi) {
+                std::size_t X  = arg.window_spatial_lengths_[0];
+                std::size_t Wo = arg.doutput_.GetLengths()[2];
+
+                float v_acc = 0;
+
+                for(std::size_t x = 0; x < X; ++x)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(x * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(w_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                  static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                        {
+                            v_acc += ck::type_convert<float>(arg.doutput_(n, c, wo));
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(X);
+                arg.dinput_(n, c, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_ncw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 2, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto c, auto hi, auto wi) {
+                std::size_t Y = arg.window_spatial_lengths_[0];
+                std::size_t X = arg.window_spatial_lengths_[1];
+
+                std::size_t Ho = arg.doutput_.GetLengths()[2];
+                std::size_t Wo = arg.doutput_.GetLengths()[3];
+
+                float v_acc = 0;
+
+                for(std::size_t y = 0; y < Y; ++y)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto h_tmp = static_cast<ck::long_index_t>(hi) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(y * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(h_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                  static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                        {
+                            for(std::size_t x = 0; x < X; ++x)
+                            {
+                                auto w_tmp =
+                                    static_cast<ck::long_index_t>(wi) +
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                    static_cast<ck::long_index_t>(x * arg.window_dilations_[1]);
+                                if(w_tmp % arg.window_strides_[1] == 0)
+                                {
+                                    auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                              static_cast<ck::long_index_t>(arg.window_strides_[1]);
+                                    if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                                    {
+                                        v_acc +=
+                                            ck::type_convert<float>(arg.doutput_(n, c, ho, wo));
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(Y * X);
+                arg.dinput_(n, c, hi, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2],
+                                       arg.dinput_.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 3, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            auto f_ncdhw = [&](auto n, auto c, auto di, auto hi, auto wi) {
+                std::size_t Z = arg.window_spatial_lengths_[0];
+                std::size_t Y = arg.window_spatial_lengths_[1];
+                std::size_t X = arg.window_spatial_lengths_[2];
+
+                std::size_t Do = arg.doutput_.GetLengths()[2];
+                std::size_t Ho = arg.doutput_.GetLengths()[3];
+                std::size_t Wo = arg.doutput_.GetLengths()[4];
+
+                float v_acc = 0;
+
+                for(std::size_t z = 0; z < Z; ++z)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto d_tmp = static_cast<ck::long_index_t>(di) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(z * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(d_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto do_ = static_cast<ck::long_index_t>(d_tmp) /
+                                   static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
+                        {
+                            for(std::size_t y = 0; y < Y; ++y)
+                            {
+                                auto h_tmp =
+                                    static_cast<ck::long_index_t>(hi) +
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                    static_cast<ck::long_index_t>(y * arg.window_dilations_[1]);
+                                if(h_tmp % arg.window_strides_[1] == 0)
+                                {
+                                    auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                              static_cast<ck::long_index_t>(arg.window_strides_[1]);
+                                    if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                                    {
+                                        for(std::size_t x = 0; x < X; ++x)
+                                        {
+                                            auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                                         static_cast<ck::long_index_t>(
+                                                             arg.in_left_pads_[2]) -
+                                                         static_cast<ck::long_index_t>(
+                                                             x * arg.window_dilations_[2]);
+
+                                            if(w_tmp % arg.window_strides_[2] == 0)
+                                            {
+                                                auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                                          static_cast<ck::long_index_t>(
+                                                              arg.window_strides_[2]);
+                                                if(wo >= 0 &&
+                                                   ck::type_convert<std::size_t>(wo) < Wo)
+                                                {
+                                                    v_acc += ck::type_convert<float>(
+                                                        arg.doutput_(n, c, do_, ho, wo));
+                                                }
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(Z * Y * X);
+                arg.dinput_(n, c, di, hi, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_ncdhw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2],
+                                       arg.dinput_.GetLengths()[3],
+                                       arg.dinput_.GetLengths()[4])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.dinput_.GetNumOfDimension() == NDimSpatial + 2 &&
+                 arg.doutput_.GetNumOfDimension() == NDimSpatial + 2))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            return RunAvgPoolBwd<NDimSpatial>(arg);
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<DInDataType>& dinput,
+                             const Tensor<DOutDataType>& doutput,
+                             std::vector<ck::index_t> window_spatial_lengths,
+                             std::vector<ck::index_t> window_strides,
+                             std::vector<ck::index_t> window_dilations,
+                             std::vector<ck::index_t> dinput_left_pads,
+                             std::vector<ck::index_t> dinput_right_pads)
+    {
+        if(window_spatial_lengths.size() != NDimSpatial || window_strides.size() != NDimSpatial ||
+           window_dilations.size() != NDimSpatial || dinput_left_pads.size() != NDimSpatial ||
+           dinput_right_pads.size() != NDimSpatial)
+            throw std::runtime_error("dimension is incorrect");
+
+        return Argument{dinput,
+                        doutput,
+                        window_spatial_lengths,
+                        window_strides,
+                        window_dilations,
+                        dinput_left_pads,
+                        dinput_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceAvgPoolBwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck