Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/marian-nmt/FBGEMM.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authorDaya Khudia <dskhudia@fb.com>2019-06-07 20:03:21 +0300
committerFacebook Github Bot <facebook-github-bot@users.noreply.github.com>2019-06-07 20:19:53 +0300
commitbf2f45f35cc0d7b6f420894652824a377f764714 (patch)
treed767a1aa5d993a6e40e82ca3b42b6ee8d5ed3e31 /src
parent8197494f3ae280941639c72bc1342a9faa8e2ad6 (diff)
Remove duplicated header and undo some changes in D15399811
Summary: Delete duplicated header Remove #ifndef and replace with pragma once. Reviewed By: jianyuh Differential Revision: D15669744 fbshipit-source-id: 8895f6c867e626ac5813a8952837435e76b09370
Diffstat (limited to 'src')
-rw-r--r--src/FbgemmI8DepthwiseAvx2.h174
1 files changed, 0 insertions, 174 deletions
diff --git a/src/FbgemmI8DepthwiseAvx2.h b/src/FbgemmI8DepthwiseAvx2.h
deleted file mode 100644
index e2730df..0000000
--- a/src/FbgemmI8DepthwiseAvx2.h
+++ /dev/null
@@ -1,174 +0,0 @@
-/*
- * Copyright (c) Facebook, Inc. and its affiliates.
- * All rights reserved.
- * This source code is licensed under the BSD-style license found in the
- * LICENSE file in the root directory of this source tree.
- */
-#ifndef I8DEPTHWISE_H
-#define I8DEPTHWISE_H
-
-#include <cstdint>
-#include "fbgemm/FbgemmBuild.h"
-
-namespace fbgemm {
-
-// KERNEL_PROD is the product of all kernels.
-// For example, KERNEL_PROD = 9 for 3x3, and 27 for 3x3x3.
-template <int KERNEL_PROD>
-class FBGEMM_API PackedDepthWiseConvMatrix {
- public:
- // smat in RSG layout
- PackedDepthWiseConvMatrix(int K, const std::int8_t* smat);
- virtual ~PackedDepthWiseConvMatrix();
-
- const std::int8_t* PackedMat() const {
- return pmat_;
- }
-
- private:
- int K_;
- std::int8_t* pmat_;
-}; // Packed3x3ConvMatrix
-
-using Packed3x3ConvMatrix = PackedDepthWiseConvMatrix<3 * 3>;
-using Packed3x3x3ConvMatrix = PackedDepthWiseConvMatrix<3 * 3 * 3>;
-
-/**
- * Depth-wise 3x3 convolution with pad=1 and stride=1 and K a multiple of 8
- * @params A The input image in NHWK layout
- * @params Bp The pre-packed filter
- */
-FBGEMM_API void depthwise_3x3_pad_1(
- int N,
- int H,
- int W,
- int K,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- const Packed3x3ConvMatrix& Bp,
- std::int32_t* C,
- int thread_id = 0,
- int num_threads = 1);
-
-/**
- * Depth-wise 3x3 convolution with pad=1 and stride=1 and K a multiple of 8
- * This version is fused with requantization.
- *
- * @col_offsets nullptr if col_offsets are folded into bias
- */
-FBGEMM_API void depthwise_3x3_pad_1(
- int N,
- int H,
- int W,
- int K,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- std::int32_t B_zero_point,
- const Packed3x3ConvMatrix& Bp,
- float C_multiplier,
- std::int32_t C_zero_point,
- std::uint8_t* C,
- const std::int32_t* col_offsets,
- const std::int32_t* bias,
- bool fuse_relu = false,
- int thread_id = 0,
- int num_threads = 1);
-
-/**
- * Depth-wise 3x3 convolution with pad=1 and stride=1 and K a multiple of 8
- * This version is fused with requantization and uses per-channel quantization.
- *
- * @col_offsets nullptr if col_offsets are folded into bias
- */
-FBGEMM_API void depthwise_3x3_per_channel_quantization_pad_1(
- int N,
- int H,
- int W,
- int K,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- const std::int32_t* B_zero_point,
- const Packed3x3ConvMatrix& Bp,
- const float* C_multiplier,
- std::int32_t C_zero_point,
- std::uint8_t* C,
- const std::int32_t* col_offsets,
- const std::int32_t* bias,
- bool fuse_relu = false,
- int thread_id = 0,
- int num_threads = 1);
-
-FBGEMM_API void depthwise_3x3x3_pad_1(
- int N,
- int T,
- int H,
- int W,
- int K,
- int stride_t,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- const Packed3x3x3ConvMatrix& Bp,
- std::int32_t* C,
- int thread_id = 0,
- int num_threads = 1);
-
-/**
- * @col_offsets nullptr if col_offsets are folded into bias
- */
-FBGEMM_API void depthwise_3x3x3_pad_1(
- int N,
- int T,
- int H,
- int W,
- int K,
- int stride_t,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- std::int32_t B_zero_point,
- const Packed3x3x3ConvMatrix& Bp,
- float C_multiplier,
- std::int32_t C_zero_point,
- std::uint8_t* C,
- const std::int32_t* col_offsets,
- const std::int32_t* bias,
- bool fuse_relu = false,
- int thread_id = 0,
- int num_threads = 1);
-
-/**
- * @col_offsets nullptr if col_offsets are folded into bias
- */
-FBGEMM_API void depthwise_3x3x3_per_channel_quantization_pad_1(
- int N,
- int T,
- int H,
- int W,
- int K,
- int stride_t,
- int stride_h,
- int stride_w,
- std::int32_t A_zero_point,
- const std::uint8_t* A,
- const std::int32_t* B_zero_point,
- const Packed3x3x3ConvMatrix& Bp,
- const float* C_multiplier,
- std::int32_t C_zero_point,
- std::uint8_t* C,
- const std::int32_t* col_offsets,
- const std::int32_t* bias,
- bool fuse_relu = false,
- int thread_id = 0,
- int num_threads = 1);
-
-} // namespace fbgemm
-#endif