[ hgemm ] Move hgemm_padding related files to explicit directory

- Adding padding to matrices is not an optimal solution to approach, but yet can be one sub-optimal option.
- Final goal of this directory would be deleting this directory itself.

**Self evaluation:**
1. Build test:     [X]Passed [ ]Failed [ ]Skipped
2. Run test:     [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: skykongkong8 <ss.kong@samsung.com>

diff --git a/nntrainer/tensor/hgemm/hgemm_pack.cpp b/nntrainer/tensor/hgemm/hgemm_pack.cpp
index 0f4b14708067cbe812f36e10e429b0ab7d62a019..63d1b702fabe1f8c17039563a0f07fd930975f60 100644 (file)
--- a/nntrainer/tensor/hgemm/hgemm_pack.cpp
+++ b/nntrainer/tensor/hgemm/hgemm_pack.cpp
@@ -15,6 +15,7 @@
 #include <assert.h>
 #include <hgemm_common.h>
 #include <hgemm_pack.h>
+#include <hgemm_util.h>
 #include <matrix_transpose_neon.h>
 
 void packing_A1(unsigned int m, unsigned int k, const __fp16 *from,
@@ -404,8 +405,7 @@ void packing_transB16(unsigned int K, unsigned int N, const __fp16 *src,
   __fp16 *dst_off = (__fp16 *)dst;
 
   const unsigned int ld_tile_T = 16;
-  __fp16 *tile_T = new __fp16[8 * ld_tile_T];
-  // __fp16 *tile_T = alignedMalloc(8 * ld_tile_T);
+  __fp16 *tile_T = alignedMalloc(8 * ld_tile_T);
 
   // 1. Do something like 8x16 transpose kernel
   // 2. Save linearized transposed output tile to dst

diff --git a/nntrainer/tensor/hgemm/hgemm_pack.h b/nntrainer/tensor/hgemm/hgemm_pack.h
index 7a671a513fc9382b68e457caa0281cddcb56c66e..b134ee1281b728ca2776fe638c92591291775691 100644 (file)
--- a/nntrainer/tensor/hgemm/hgemm_pack.h
+++ b/nntrainer/tensor/hgemm/hgemm_pack.h
@@ -89,13 +89,13 @@ void packing_B8(unsigned int K, unsigned int N, const __fp16 *src,
 void packing_B16(unsigned int K, unsigned int N, const __fp16 *src,
                  unsigned int ldb, const __fp16 *dst);
 /**
- * @brief
+ * @brief packing function of input matrix B_T
  *
- * @param K
- * @param N
- * @param src
- * @param ldb
- * @param dst
+ * @param M length of the row of the matrix
+ * @param K length of the col of the matrix
+ * @param src input of original source of the matrix
+ * @param ldb leading dimension of the matrix
+ * @param dst output of packed data of the matrix
  */
 void packing_transB16(unsigned int K, unsigned int N, const __fp16 *src,
                       unsigned int ldb, const __fp16 *dst);

diff --git a/nntrainer/tensor/hgemm/hgemm_padding.h b/nntrainer/tensor/hgemm/hgemm_padding.h
deleted file mode 100644 (file)
index f62143a..0000000
--- a/nntrainer/tensor/hgemm/hgemm_padding.h
+++ /dev/null
@@ -1,15 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
- *
- * @file   hgemm_padding.h
- * @date   05 July 2024
- * @see    https://github.com/nnstreamer/nntrainer
- * @author Sungsik Kong <ss.kong@samsung.com>
- * @bug    No known bugs except for NYI items
- * @brief  This is a header file for including both padding matrix A and B
- *
- */
-
-#include <hgemm_padding_a.h>
-#include <hgemm_padding_b.h>

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding.h b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding.h
new file mode 100644 (file)
index 0000000..f62143a
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding.h
@@ -0,0 +1,15 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
+ *
+ * @file   hgemm_padding.h
+ * @date   05 July 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Sungsik Kong <ss.kong@samsung.com>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is a header file for including both padding matrix A and B
+ *
+ */
+
+#include <hgemm_padding_a.h>
+#include <hgemm_padding_b.h>

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.cpp b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.cpp
new file mode 100644 (file)
index 0000000..99eab5c
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.cpp
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
+ *
+ * @file   hgemm_padding_a.cpp
+ * @date   05 July 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Sungsik Kong <ss.kong@samsung.com>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is a source file for padding function used in hgemm
+ *
+ */
+
+#include <arm_neon.h>
+#include <hgemm_padding_a.h>
+#include <hgemm_util.h>
+#include <iostream>
+
+void hgemm_padding_A(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                     unsigned int K, unsigned int M8, unsigned int K8,
+                     bool transA) {
+  if (transA)
+    hgemm_padding_A_Trans(A, Ap, M, K, M8, K8);
+  else
+    hgemm_padding_A_noTrans(A, Ap, M, K, M8, K8);
+}
+
+void hgemm_padding_A_noTrans(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                             unsigned int K, unsigned int M8, unsigned int K8) {
+  if (M != M8 && K != K8) {
+    hgemm_padding_A_noTrans_wrt_MK(A, Ap, M, K, M8, K8);
+  } else if (M != M8) {
+    hgemm_padding_A_noTrans_wrt_M(A, Ap, M, K, M8, K8);
+  } else if (K != K8) {
+    hgemm_padding_A_noTrans_wrt_K(A, Ap, M, K, M8, K8);
+  } else {
+    std::cerr << "Error : No room for matrix A padding!\n";
+  }
+}
+
+void hgemm_padding_A_Trans(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                           unsigned int K, unsigned int M8, unsigned int K8) {
+  if (M != M8 && K != K8) {
+    hgemm_padding_A_Trans_wrt_MK(A, Ap, M, K, M8, K8);
+  } else if (M != M8) {
+    hgemm_padding_A_Trans_wrt_M(A, Ap, M, K, M8, K8);
+  } else if (K != K8) {
+    hgemm_padding_A_Trans_wrt_K(A, Ap, M, K, M8, K8);
+  } else {
+    std::cerr << "Error : No room for matrix A padding!\n";
+  }
+}
+
+void hgemm_padding_A_noTrans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                   unsigned int K, unsigned int M8,
+                                   unsigned int K8) {
+  float16x8_t ZEROS = vmovq_n_f16(0.F);
+
+  for (unsigned int m = 0; m < M; ++m) {
+    for (unsigned int k = 0; k < K; k += 8) {
+      vst1q_f16(&Ap[m * K + k], vld1q_f16(&A[m * K + k]));
+    }
+  }
+  for (unsigned int m = M; m < M8; ++m) {
+    for (unsigned int k = 0; k < K; k += 8) {
+      vst1q_f16(&Ap[m * K + k], ZEROS);
+    }
+  }
+}
+
+void hgemm_padding_A_noTrans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                   unsigned int K, unsigned int M8,
+                                   unsigned int K8) {
+  const unsigned int K8_low = (K >> 3) << 3;
+  float16x8_t ZEROS = vmovq_n_f16(0.F);
+
+  for (unsigned int m = 0; m < M; ++m) {
+    for (unsigned int k = 0; k < K8_low; k += 8) {
+      vst1q_f16(&Ap[m * K8 + k], vld1q_f16(&A[m * K + k]));
+    }
+    for (unsigned int k = K8_low; k < K; ++k) {
+      Ap[m * K8 + k] = A[m * K + k];
+    }
+    for (unsigned int k = K; k < K8; ++k) {
+      Ap[m * K8 + k] = 0.F;
+    }
+  }
+}
+
+void hgemm_padding_A_noTrans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                    unsigned int K, unsigned int M8,
+                                    unsigned int K8) {
+  std::cerr << "Error : hgemm_padding_A_noTrans_wrt_MK NYI!\n";
+}
+
+void hgemm_padding_A_Trans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                 unsigned int K, unsigned int M8,
+                                 unsigned int K8) {
+  const unsigned int M8_low = (M >> 3) << 3;
+
+  for (unsigned int k = 0; k < K; ++k) {
+    for (unsigned int m = 0; m < M8_low; m += 8) {
+      vst1q_f16(&Ap[k * M + m], vld1q_f16(&A[k * M + m]));
+    }
+    for (unsigned int m = M8_low; m < M; ++m) {
+      Ap[k * M + m] = A[k * M + m];
+    }
+    for (unsigned int m = M; m < M8; ++m) {
+      Ap[k * M + m] = 0.F;
+    }
+  }
+}
+
+void hgemm_padding_A_Trans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                 unsigned int K, unsigned int M8,
+                                 unsigned int K8) {
+  std::cerr << "Error : hgemm_padding_A_Trans_wrt_K NYI!\n";
+}
+
+void hgemm_padding_A_Trans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                  unsigned int K, unsigned int M8,
+                                  unsigned int K8) {
+  std::cerr << "Error : hgemm_padding_A_Trans_wrt_MK NYI!\n";
+}

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.h b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.h
new file mode 100644 (file)
index 0000000..f7b6885
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_a.h
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
+ *
+ * @file   hgemm_padding_a.h
+ * @date   05 July 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Sungsik Kong <ss.kong@samsung.com>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is a header file for padding function used in hgemm
+ *
+ */
+
+/**
+ * @brief Padding function for matrix A in HGEMM
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param transA Whether the matrix A is transposed or not
+ */
+void hgemm_padding_A(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                     unsigned int K, unsigned int M8, unsigned int K8,
+                     bool transA);
+
+/**
+ * @brief Padding function for non-transposed matrix A in HGEMM
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_noTrans(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                             unsigned int K, unsigned int M8, unsigned int K8);
+
+/**
+ * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. M
+ * direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_noTrans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                   unsigned int K, unsigned int M8,
+                                   unsigned int K8);
+/**
+ * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. K
+ * direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_noTrans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                   unsigned int K, unsigned int M8,
+                                   unsigned int K8);
+
+/**
+ * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. M and K
+ * direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_noTrans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                    unsigned int K, unsigned int M8,
+                                    unsigned int K8);
+/**
+ * @brief Padding function for transposed matrix A in HGEMM
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_Trans(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                           unsigned int K, unsigned int M8, unsigned int K8);
+/**
+ * @brief Padding function for transposed matrix A in HGEMM w.r.t. M direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_Trans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                 unsigned int K, unsigned int M8,
+                                 unsigned int K8);
+/**
+ * @brief Padding function for transposed matrix A in HGEMM w.r.t. K direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_Trans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                 unsigned int K, unsigned int M8,
+                                 unsigned int K8);
+/**
+ * @brief Padding function for transposed matrix A in HGEMM w.r.t. M and K
+ * direction
+ *
+ * @param A src matrix to pad
+ * @param Ap dst matrix after padding
+ * @param M row length of matrix A
+ * @param K col length of matrix A
+ * @param M8 Least multiple of 8 that is bigger than or equal to M
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ */
+void hgemm_padding_A_Trans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
+                                  unsigned int K, unsigned int M8,
+                                  unsigned int K8);

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.cpp b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.cpp
new file mode 100644 (file)
index 0000000..f0ef105
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.cpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
+ *
+ * @file   hgemm_padding_b.cpp
+ * @date   05 July 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Sungsik Kong <ss.kong@samsung.com>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is a source file for padding function used in hgemm
+ *
+ */
+
+#include <arm_neon.h>
+#include <hgemm_padding_b.h>
+#include <hgemm_util.h>
+#include <iostream>
+
+void hgemm_padding_B(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                     unsigned int N, unsigned int K8, unsigned int N16,
+                     bool transB) {
+  if (transB) {
+    hgemm_padding_B_Trans(B, Bp, K, N, K8, N16);
+  } else {
+    hgemm_padding_B_noTrans(B, Bp, K, N, K8, N16);
+  }
+}
+
+void hgemm_padding_B_noTrans(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                             unsigned int N, unsigned int K8,
+                             unsigned int N16) {
+  if (K != K8 && N != N16) {
+    hgemm_padding_B_noTrans_wrt_KN(B, Bp, K, N, K8, N16);
+  } else if (K != K8) {
+    hgemm_padding_B_noTrans_wrt_K(B, Bp, K, N, K8, N16);
+  } else if (N != N16) {
+    hgemm_padding_B_noTrans_wrt_N(B, Bp, K, N, K8, N16);
+  } else {
+    std::cerr << "Error : No room for matrix B padding\n";
+  }
+}
+
+void hgemm_padding_B_Trans(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                           unsigned int N, unsigned int K8, unsigned int N16) {
+  if (K != K8 && N != N16) {
+    hgemm_padding_B_Trans_wrt_KN(B, Bp, K, N, K8, N16);
+  } else if (K != K8) {
+    hgemm_padding_B_Trans_wrt_K(B, Bp, K, N, K8, N16);
+  } else if (N != N16) {
+    hgemm_padding_B_Trans_wrt_N(B, Bp, K, N, K8, N16);
+  } else {
+    std::cerr << "Error : No room for matrix B padding\n";
+  }
+}
+
+void hgemm_padding_B_noTrans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                   unsigned int N, unsigned int K8,
+                                   unsigned int N16) {
+  std::cerr << "Error : hgemm_padding_B_noTrans_wrt_N NYI!\n";
+}
+
+void hgemm_padding_B_noTrans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                   unsigned int N, unsigned int K8,
+                                   unsigned int N16) {
+  float16x8_t ZEROS = vmovq_n_f16(0.F);
+
+  for (unsigned int k = 0; k < K; ++k) {
+    for (unsigned int n = 0; n < N; n += 8) {
+      vst1q_f16(&Bp[k * N + n], vld1q_f16(&B[k * N + n]));
+    }
+  }
+  for (unsigned int k = K; k < K8; ++k) {
+    for (unsigned int n = 0; n < N; n += 8) {
+      vst1q_f16(&Bp[k * N + n], ZEROS);
+    }
+  }
+}
+
+void hgemm_padding_B_noTrans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                    unsigned int N, unsigned int K8,
+                                    unsigned int N16) {
+  std::cerr << "Error : hgemm_padding_B_noTrans_wrt_KN NYI!\n";
+}
+
+void hgemm_padding_B_Trans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                 unsigned int N, unsigned int K8,
+                                 unsigned int N16) {
+  std::cerr << "Error : hgemm_padding_B_Trans_wrt_N NYI!\n";
+}
+
+void hgemm_padding_B_Trans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                 unsigned int N, unsigned int K8,
+                                 unsigned int N16) {
+  const unsigned int K8_low = (K >> 3) << 3;
+  float16x8_t ZEROS = vmovq_n_f16(0.F);
+
+  for (unsigned int n = 0; n < N; ++n) {
+    for (unsigned int k = 0; k < K8_low; k += 8) {
+      vst1q_f16(&Bp[n * K8 + k], vld1q_f16(&B[n * K + k]));
+    }
+    for (unsigned int k = K8_low; k < K; ++k) {
+      Bp[n * K8 + k] = B[n * K + k];
+    }
+    for (unsigned int k = K; k < K8; ++k) {
+      Bp[n * K8 + k] = 0.F;
+    }
+  }
+}
+
+void hgemm_padding_B_Trans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                  unsigned int N, unsigned int K8,
+                                  unsigned int N16) {
+  std::cerr << "Error : hgemm_padding_B_Trans_wrt_KN NYI!\n";
+}

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.h b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.h
new file mode 100644 (file)
index 0000000..53e1697
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/hgemm_padding_b.h
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
+ *
+ * @file   hgemm_padding_b.h
+ * @date   05 July 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Sungsik Kong <ss.kong@samsung.com>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is a header file for padding function used in hgemm
+ *
+ */
+
+/**
+ * @brief Padding function for matrix B in HGEMM
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ * @param transB Whether the matrix B is transposed or not
+ */
+void hgemm_padding_B(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                     unsigned int N, unsigned int K8, unsigned int N16,
+                     bool transB);
+
+/**
+ * @brief Padding function for non-transposed matrix B in HGEMM
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_noTrans(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                             unsigned int N, unsigned int K8, unsigned int N16);
+/**
+ * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. N
+ * direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_noTrans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                   unsigned int N, unsigned int K8,
+                                   unsigned int N16);
+/**
+ * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. K
+ * direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_noTrans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                   unsigned int N, unsigned int K8,
+                                   unsigned int N16);
+/**
+ * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. N and K
+ * direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_noTrans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                    unsigned int N, unsigned int K8,
+                                    unsigned int N16);
+/**
+ * @brief Padding function for transposed matrix B in HGEMM
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_Trans(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                           unsigned int N, unsigned int K8, unsigned int N16);
+/**
+ * @brief Padding function for transposed matrix B in HGEMM w.r.t. N direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_Trans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                 unsigned int N, unsigned int K8,
+                                 unsigned int N16);
+/**
+ * @brief Padding function for transposed matrix B in HGEMM w.r.t. K direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_Trans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                 unsigned int N, unsigned int K8,
+                                 unsigned int N16);
+
+/**
+ * @brief Padding function for transposed matrix B in HGEMM w.r.t. K and N
+ * direction
+ *
+ * @param B src matrix to pad
+ * @param Bp dst matrix after padding
+ * @param K row length of matrix B
+ * @param N col length of matrix B
+ * @param K8 Least multiple of 8 that is bigger than or equal to K
+ * @param N16 Least multiple of 16 that is bigger than or equal to N
+ */
+void hgemm_padding_B_Trans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
+                                  unsigned int N, unsigned int K8,
+                                  unsigned int N16);

diff --git a/nntrainer/tensor/hgemm/hgemm_padding/meson.build b/nntrainer/tensor/hgemm/hgemm_padding/meson.build
new file mode 100644 (file)
index 0000000..c647774
--- /dev/null
+++ b/nntrainer/tensor/hgemm/hgemm_padding/meson.build
@@ -0,0 +1,20 @@
+hgemm_padding_headers = [
+  'hgemm_padding.h',
+  'hgemm_padding_a.h',
+  'hgemm_padding_b.h',
+]
+
+
+hgemm_padding_sources = [
+  'hgemm_padding_a.cpp',
+  'hgemm_padding_b.cpp',
+]
+
+foreach s : hgemm_padding_sources
+  nntrainer_sources += meson.current_source_dir() / s
+endforeach
+
+foreach h : hgemm_padding_headers
+  nntrainer_headers += meson.current_source_dir() / h
+endforeach
+

diff --git a/nntrainer/tensor/hgemm/hgemm_padding_a.cpp b/nntrainer/tensor/hgemm/hgemm_padding_a.cpp
deleted file mode 100644 (file)
index 99eab5c..0000000
--- a/nntrainer/tensor/hgemm/hgemm_padding_a.cpp
+++ /dev/null
@@ -1,124 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
- *
- * @file   hgemm_padding_a.cpp
- * @date   05 July 2024
- * @see    https://github.com/nnstreamer/nntrainer
- * @author Sungsik Kong <ss.kong@samsung.com>
- * @bug    No known bugs except for NYI items
- * @brief  This is a source file for padding function used in hgemm
- *
- */
-
-#include <arm_neon.h>
-#include <hgemm_padding_a.h>
-#include <hgemm_util.h>
-#include <iostream>
-
-void hgemm_padding_A(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                     unsigned int K, unsigned int M8, unsigned int K8,
-                     bool transA) {
-  if (transA)
-    hgemm_padding_A_Trans(A, Ap, M, K, M8, K8);
-  else
-    hgemm_padding_A_noTrans(A, Ap, M, K, M8, K8);
-}
-
-void hgemm_padding_A_noTrans(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                             unsigned int K, unsigned int M8, unsigned int K8) {
-  if (M != M8 && K != K8) {
-    hgemm_padding_A_noTrans_wrt_MK(A, Ap, M, K, M8, K8);
-  } else if (M != M8) {
-    hgemm_padding_A_noTrans_wrt_M(A, Ap, M, K, M8, K8);
-  } else if (K != K8) {
-    hgemm_padding_A_noTrans_wrt_K(A, Ap, M, K, M8, K8);
-  } else {
-    std::cerr << "Error : No room for matrix A padding!\n";
-  }
-}
-
-void hgemm_padding_A_Trans(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                           unsigned int K, unsigned int M8, unsigned int K8) {
-  if (M != M8 && K != K8) {
-    hgemm_padding_A_Trans_wrt_MK(A, Ap, M, K, M8, K8);
-  } else if (M != M8) {
-    hgemm_padding_A_Trans_wrt_M(A, Ap, M, K, M8, K8);
-  } else if (K != K8) {
-    hgemm_padding_A_Trans_wrt_K(A, Ap, M, K, M8, K8);
-  } else {
-    std::cerr << "Error : No room for matrix A padding!\n";
-  }
-}
-
-void hgemm_padding_A_noTrans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                   unsigned int K, unsigned int M8,
-                                   unsigned int K8) {
-  float16x8_t ZEROS = vmovq_n_f16(0.F);
-
-  for (unsigned int m = 0; m < M; ++m) {
-    for (unsigned int k = 0; k < K; k += 8) {
-      vst1q_f16(&Ap[m * K + k], vld1q_f16(&A[m * K + k]));
-    }
-  }
-  for (unsigned int m = M; m < M8; ++m) {
-    for (unsigned int k = 0; k < K; k += 8) {
-      vst1q_f16(&Ap[m * K + k], ZEROS);
-    }
-  }
-}
-
-void hgemm_padding_A_noTrans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                   unsigned int K, unsigned int M8,
-                                   unsigned int K8) {
-  const unsigned int K8_low = (K >> 3) << 3;
-  float16x8_t ZEROS = vmovq_n_f16(0.F);
-
-  for (unsigned int m = 0; m < M; ++m) {
-    for (unsigned int k = 0; k < K8_low; k += 8) {
-      vst1q_f16(&Ap[m * K8 + k], vld1q_f16(&A[m * K + k]));
-    }
-    for (unsigned int k = K8_low; k < K; ++k) {
-      Ap[m * K8 + k] = A[m * K + k];
-    }
-    for (unsigned int k = K; k < K8; ++k) {
-      Ap[m * K8 + k] = 0.F;
-    }
-  }
-}
-
-void hgemm_padding_A_noTrans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                    unsigned int K, unsigned int M8,
-                                    unsigned int K8) {
-  std::cerr << "Error : hgemm_padding_A_noTrans_wrt_MK NYI!\n";
-}
-
-void hgemm_padding_A_Trans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                 unsigned int K, unsigned int M8,
-                                 unsigned int K8) {
-  const unsigned int M8_low = (M >> 3) << 3;
-
-  for (unsigned int k = 0; k < K; ++k) {
-    for (unsigned int m = 0; m < M8_low; m += 8) {
-      vst1q_f16(&Ap[k * M + m], vld1q_f16(&A[k * M + m]));
-    }
-    for (unsigned int m = M8_low; m < M; ++m) {
-      Ap[k * M + m] = A[k * M + m];
-    }
-    for (unsigned int m = M; m < M8; ++m) {
-      Ap[k * M + m] = 0.F;
-    }
-  }
-}
-
-void hgemm_padding_A_Trans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                 unsigned int K, unsigned int M8,
-                                 unsigned int K8) {
-  std::cerr << "Error : hgemm_padding_A_Trans_wrt_K NYI!\n";
-}
-
-void hgemm_padding_A_Trans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                  unsigned int K, unsigned int M8,
-                                  unsigned int K8) {
-  std::cerr << "Error : hgemm_padding_A_Trans_wrt_MK NYI!\n";
-}

diff --git a/nntrainer/tensor/hgemm/hgemm_padding_a.h b/nntrainer/tensor/hgemm/hgemm_padding_a.h
deleted file mode 100644 (file)
index f7b6885..0000000
--- a/nntrainer/tensor/hgemm/hgemm_padding_a.h
+++ /dev/null
@@ -1,136 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
- *
- * @file   hgemm_padding_a.h
- * @date   05 July 2024
- * @see    https://github.com/nnstreamer/nntrainer
- * @author Sungsik Kong <ss.kong@samsung.com>
- * @bug    No known bugs except for NYI items
- * @brief  This is a header file for padding function used in hgemm
- *
- */
-
-/**
- * @brief Padding function for matrix A in HGEMM
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param transA Whether the matrix A is transposed or not
- */
-void hgemm_padding_A(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                     unsigned int K, unsigned int M8, unsigned int K8,
-                     bool transA);
-
-/**
- * @brief Padding function for non-transposed matrix A in HGEMM
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_noTrans(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                             unsigned int K, unsigned int M8, unsigned int K8);
-
-/**
- * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. M
- * direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_noTrans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                   unsigned int K, unsigned int M8,
-                                   unsigned int K8);
-/**
- * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. K
- * direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_noTrans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                   unsigned int K, unsigned int M8,
-                                   unsigned int K8);
-
-/**
- * @brief Padding function for non-transposed matrix A in HGEMM w.r.t. M and K
- * direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_noTrans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                    unsigned int K, unsigned int M8,
-                                    unsigned int K8);
-/**
- * @brief Padding function for transposed matrix A in HGEMM
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_Trans(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                           unsigned int K, unsigned int M8, unsigned int K8);
-/**
- * @brief Padding function for transposed matrix A in HGEMM w.r.t. M direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_Trans_wrt_M(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                 unsigned int K, unsigned int M8,
-                                 unsigned int K8);
-/**
- * @brief Padding function for transposed matrix A in HGEMM w.r.t. K direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_Trans_wrt_K(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                 unsigned int K, unsigned int M8,
-                                 unsigned int K8);
-/**
- * @brief Padding function for transposed matrix A in HGEMM w.r.t. M and K
- * direction
- *
- * @param A src matrix to pad
- * @param Ap dst matrix after padding
- * @param M row length of matrix A
- * @param K col length of matrix A
- * @param M8 Least multiple of 8 that is bigger than or equal to M
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- */
-void hgemm_padding_A_Trans_wrt_MK(const __fp16 *A, __fp16 *Ap, unsigned int M,
-                                  unsigned int K, unsigned int M8,
-                                  unsigned int K8);

diff --git a/nntrainer/tensor/hgemm/hgemm_padding_b.cpp b/nntrainer/tensor/hgemm/hgemm_padding_b.cpp
deleted file mode 100644 (file)
index f0ef105..0000000
--- a/nntrainer/tensor/hgemm/hgemm_padding_b.cpp
+++ /dev/null
@@ -1,114 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
- *
- * @file   hgemm_padding_b.cpp
- * @date   05 July 2024
- * @see    https://github.com/nnstreamer/nntrainer
- * @author Sungsik Kong <ss.kong@samsung.com>
- * @bug    No known bugs except for NYI items
- * @brief  This is a source file for padding function used in hgemm
- *
- */
-
-#include <arm_neon.h>
-#include <hgemm_padding_b.h>
-#include <hgemm_util.h>
-#include <iostream>
-
-void hgemm_padding_B(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                     unsigned int N, unsigned int K8, unsigned int N16,
-                     bool transB) {
-  if (transB) {
-    hgemm_padding_B_Trans(B, Bp, K, N, K8, N16);
-  } else {
-    hgemm_padding_B_noTrans(B, Bp, K, N, K8, N16);
-  }
-}
-
-void hgemm_padding_B_noTrans(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                             unsigned int N, unsigned int K8,
-                             unsigned int N16) {
-  if (K != K8 && N != N16) {
-    hgemm_padding_B_noTrans_wrt_KN(B, Bp, K, N, K8, N16);
-  } else if (K != K8) {
-    hgemm_padding_B_noTrans_wrt_K(B, Bp, K, N, K8, N16);
-  } else if (N != N16) {
-    hgemm_padding_B_noTrans_wrt_N(B, Bp, K, N, K8, N16);
-  } else {
-    std::cerr << "Error : No room for matrix B padding\n";
-  }
-}
-
-void hgemm_padding_B_Trans(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                           unsigned int N, unsigned int K8, unsigned int N16) {
-  if (K != K8 && N != N16) {
-    hgemm_padding_B_Trans_wrt_KN(B, Bp, K, N, K8, N16);
-  } else if (K != K8) {
-    hgemm_padding_B_Trans_wrt_K(B, Bp, K, N, K8, N16);
-  } else if (N != N16) {
-    hgemm_padding_B_Trans_wrt_N(B, Bp, K, N, K8, N16);
-  } else {
-    std::cerr << "Error : No room for matrix B padding\n";
-  }
-}
-
-void hgemm_padding_B_noTrans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                   unsigned int N, unsigned int K8,
-                                   unsigned int N16) {
-  std::cerr << "Error : hgemm_padding_B_noTrans_wrt_N NYI!\n";
-}
-
-void hgemm_padding_B_noTrans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                   unsigned int N, unsigned int K8,
-                                   unsigned int N16) {
-  float16x8_t ZEROS = vmovq_n_f16(0.F);
-
-  for (unsigned int k = 0; k < K; ++k) {
-    for (unsigned int n = 0; n < N; n += 8) {
-      vst1q_f16(&Bp[k * N + n], vld1q_f16(&B[k * N + n]));
-    }
-  }
-  for (unsigned int k = K; k < K8; ++k) {
-    for (unsigned int n = 0; n < N; n += 8) {
-      vst1q_f16(&Bp[k * N + n], ZEROS);
-    }
-  }
-}
-
-void hgemm_padding_B_noTrans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                    unsigned int N, unsigned int K8,
-                                    unsigned int N16) {
-  std::cerr << "Error : hgemm_padding_B_noTrans_wrt_KN NYI!\n";
-}
-
-void hgemm_padding_B_Trans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                 unsigned int N, unsigned int K8,
-                                 unsigned int N16) {
-  std::cerr << "Error : hgemm_padding_B_Trans_wrt_N NYI!\n";
-}
-
-void hgemm_padding_B_Trans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                 unsigned int N, unsigned int K8,
-                                 unsigned int N16) {
-  const unsigned int K8_low = (K >> 3) << 3;
-  float16x8_t ZEROS = vmovq_n_f16(0.F);
-
-  for (unsigned int n = 0; n < N; ++n) {
-    for (unsigned int k = 0; k < K8_low; k += 8) {
-      vst1q_f16(&Bp[n * K8 + k], vld1q_f16(&B[n * K + k]));
-    }
-    for (unsigned int k = K8_low; k < K; ++k) {
-      Bp[n * K8 + k] = B[n * K + k];
-    }
-    for (unsigned int k = K; k < K8; ++k) {
-      Bp[n * K8 + k] = 0.F;
-    }
-  }
-}
-
-void hgemm_padding_B_Trans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                  unsigned int N, unsigned int K8,
-                                  unsigned int N16) {
-  std::cerr << "Error : hgemm_padding_B_Trans_wrt_KN NYI!\n";
-}

diff --git a/nntrainer/tensor/hgemm/hgemm_padding_b.h b/nntrainer/tensor/hgemm/hgemm_padding_b.h
deleted file mode 100644 (file)
index 53e1697..0000000
--- a/nntrainer/tensor/hgemm/hgemm_padding_b.h
+++ /dev/null
@@ -1,135 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Copyright (C) 2024 Sungsik Kong <ss.kong@samsung.com>
- *
- * @file   hgemm_padding_b.h
- * @date   05 July 2024
- * @see    https://github.com/nnstreamer/nntrainer
- * @author Sungsik Kong <ss.kong@samsung.com>
- * @bug    No known bugs except for NYI items
- * @brief  This is a header file for padding function used in hgemm
- *
- */
-
-/**
- * @brief Padding function for matrix B in HGEMM
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- * @param transB Whether the matrix B is transposed or not
- */
-void hgemm_padding_B(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                     unsigned int N, unsigned int K8, unsigned int N16,
-                     bool transB);
-
-/**
- * @brief Padding function for non-transposed matrix B in HGEMM
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_noTrans(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                             unsigned int N, unsigned int K8, unsigned int N16);
-/**
- * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. N
- * direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_noTrans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                   unsigned int N, unsigned int K8,
-                                   unsigned int N16);
-/**
- * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. K
- * direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_noTrans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                   unsigned int N, unsigned int K8,
-                                   unsigned int N16);
-/**
- * @brief Padding function for non-transposed matrix B in HGEMM w.r.t. N and K
- * direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_noTrans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                    unsigned int N, unsigned int K8,
-                                    unsigned int N16);
-/**
- * @brief Padding function for transposed matrix B in HGEMM
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_Trans(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                           unsigned int N, unsigned int K8, unsigned int N16);
-/**
- * @brief Padding function for transposed matrix B in HGEMM w.r.t. N direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_Trans_wrt_N(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                 unsigned int N, unsigned int K8,
-                                 unsigned int N16);
-/**
- * @brief Padding function for transposed matrix B in HGEMM w.r.t. K direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_Trans_wrt_K(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                 unsigned int N, unsigned int K8,
-                                 unsigned int N16);
-
-/**
- * @brief Padding function for transposed matrix B in HGEMM w.r.t. K and N
- * direction
- *
- * @param B src matrix to pad
- * @param Bp dst matrix after padding
- * @param K row length of matrix B
- * @param N col length of matrix B
- * @param K8 Least multiple of 8 that is bigger than or equal to K
- * @param N16 Least multiple of 16 that is bigger than or equal to N
- */
-void hgemm_padding_B_Trans_wrt_KN(const __fp16 *B, __fp16 *Bp, unsigned int K,
-                                  unsigned int N, unsigned int K8,
-                                  unsigned int N16);

diff --git a/nntrainer/tensor/hgemm/meson.build b/nntrainer/tensor/hgemm/meson.build
index 90ef0a3774a7716c5c7e8865c370f13bc508875a..e100f63ade7ca626ffb0d84c2e201551d16fa315 100644 (file)
--- a/nntrainer/tensor/hgemm/meson.build
+++ b/nntrainer/tensor/hgemm/meson.build
@@ -10,6 +10,10 @@ subdir('hgemm_kernel')
 nntrainer_inc += include_directories('hgemm_kernel')
 nntrainer_inc_abs += meson.current_source_dir() / 'hgemm_kernel'
 
+subdir('hgemm_padding')
+nntrainer_inc += include_directories('hgemm_padding')
+nntrainer_inc_abs += meson.current_source_dir() / 'hgemm_padding'
+
 hgemm_sources = [
     'hgemm.cpp',
     'hgemm_padding_a.cpp',