ROCm: transpose_bin() vs transpose_bin_gpu() to fix transpose_32x32_b…

…its_reversed_diagonale() link problems (issue #99)

src-lib/convolutional_layer.cpp

@@ -151,7 +151,11 @@ namespace
 
 		memset(*t_bit_input, 0, t_bit_input_size * sizeof(char));
 
+#ifdef DARKNET_GPU
+		transpose_bin_gpu((uint8_t*)b, (uint8_t*)*t_bit_input, k, n, bit_align, new_ldb, 8);
+#else
 		transpose_bin((uint32_t*)b, (uint32_t*)*t_bit_input, k, n, bit_align, new_ldb, 8);
+#endif
 
 		return t_intput_size;
 	}

src-lib/gemm.cpp

@@ -342,13 +342,13 @@ void transpose8rS32_reversed_diagonale(unsigned char* A, unsigned char* B, int m
 	B[3 * n] = reverse_byte(y >> 24);  B[2 * n] = reverse_byte(y >> 16);  B[1 * n] = reverse_byte(y >> 8);  B[0 * n] = reverse_byte(y);
 }
 
-#endif // DARKNET_GPU
-
 // transpose by 32-bit
 void transpose_bin(uint32_t *A, uint32_t *B, const int n, const int m, const int lda, const int ldb, const int block_size)
 {
 	TAT(TATPARMS);
 
+	/// @note This function is for CPU-only versions of Darknet.  See im2col_kernels.cu for GPU version.
+
 	//printf("\n n = %d (n mod 32 = %d), m = %d (m mod 32 = %d) \n", n, n % 32, m, m % 32);
 	//printf("\n lda = %d (lda mod 32 = %d), ldb = %d (ldb mod 32 = %d) \n", lda, lda % 32, ldb, ldb % 32);
 
@@ -375,6 +375,8 @@ void transpose_bin(uint32_t *A, uint32_t *B, const int n, const int m, const int
 	}
 }
 
+#endif // DARKNET_GPU
+
 #if (defined(__AVX__) && defined(__x86_64__)) || (defined(_WIN64) && !defined(__MINGW32__) && !defined(_M_ARM64))
 
 #if (defined(_WIN64) && !defined(__MINGW64__))