From 5b0873c31ad3229ec8968dfc2be25b915fc95376 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= Date: Sat, 22 Mar 2025 23:19:49 +0100 Subject: [PATCH] Fix split N for large images in groupd conv fwd (#2004) * Fix split N for large images in groupd conv fwd * Fix comments --- .../transform_conv_bwd_data_to_gemm_v1.hpp | 9 +++++---- .../operator_transform/transform_conv_fwd_to_gemm.hpp | 7 ++++--- ...device_grouped_conv_fwd_xdl_large_tensor_instance.hpp | 4 +++- 3 files changed, 12 insertions(+), 8 deletions(-) diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp index 8df0d885b9..0ddfd0a7c8 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp +++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -106,9 +106,10 @@ struct TransformConvBwdDataToGemm_v1 } else { - // Not possible to support even after split N. - // Too large tensor. - return N; + // Split Convolution's N dimension into N workgroups. However + // this still might not result in sufficiently small tensor, + // but at least later on we could divide the image as well. + return 1; } } else diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp index 3db94deccb..c291f3994c 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp +++ b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp @@ -83,9 +83,10 @@ struct TransformConvFwdToGemm } else { - // Not possible to support even after split N. - // Too large tensor. - return N; + // Split Convolution's N dimension into N workgroups. However + // this still might not result in sufficiently small tensor, + // but at least later on we could divide the image as well. + return 1; } } else diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp index d317d270ce..0a85cde3bc 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" @@ -46,6 +46,7 @@ using device_grouped_conv_fwd_xdl_large_tensor_bf16_instances = std::tuple< // generic instance DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> // clang-format on >; @@ -65,6 +66,7 @@ using device_grouped_conv_fwd_xdl_large_tensor_f16_instances = std::tuple< // generic instance DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8> // clang-format on >;