rml_op_type

Operations supported by Radeon ML.

API

Graph Manipulation

Items

Name	Value	Brief Explanation
RML_OP_UNSPECIFIED	0	Unspecified operation type.
RML_OP_ABS	1010	Calculates absolute of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::abs, rml_op_unary_params
RML_OP_ACOS	1020	Calculates arccosine of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::acos, rml_op_unary_params
RML_OP_ADD	1030	Sums input tensors up, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::add, rml_op_binary_params
RML_OP_ASIN	1050	Calculates arcsine of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::asin, rml_op_unary_params
RML_OP_ATAN	1070	Calculates arctangent of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::atan, rml_op_unary_params
RML_OP_AVG	1080	Calculates average between input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::avg, rml_op_binary_params
RML_OP_BATCH_NORM	1090	Performs batch normalization of an input tensor. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::batch_norm, rml_op_batch_norm_params
RML_OP_BIAS_ADD	1100	Adds bias to an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::bias_add, rml_op_bias_add_params
RML_OP_CAST	1110	Casts elements of an input tensor to a specified type. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::cast, rml_op_cast_params
RML_OP_CEIL	1120	Ceils elements of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::ceil, rml_op_unary_params
RML_OP_CELU	1130	Applies CeLU activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::celu, rml_op_celu_params
RML_OP_CLIP	1140	Limits input tensor values within a specified range. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::clip, rml_op_clip_params
RML_OP_CONCAT	1150	Concatenates input tensors into a single tensor along a specified axis. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::concat, rml_op_concat_params
RML_OP_CONST	1160	Stores specified tensor data. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::constant, rml_op_const_params
RML_OP_CONV_2D	1190	Performs 2D convolution of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::conv_2d, rml_op_conv_2d_params
RML_OP_CONV_2D_DEPTHWISE	1200	Performs depthwise 2D convolution of an input tensor. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::conv2d_depthwise, rml_op_conv_2d_params
RML_OP_CONV_2D_TRANSPOSE	1210	Performs transposed 2D convolution of an input tensor. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::conv2d_transpose, rml_op_conv_2d_params
RML_OP_COS	1240	Calculates cosine of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::cos, rml_op_unary_params
RML_OP_DEPTH_TO_SPACE	1250	Rearranges input tensor data from channels into blocks of spatial data. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::depth_to_space, rml_op_depth_to_space_params
RML_OP_DIV	1270	Divides input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::div, rml_op_binary_params
RML_OP_ELU	1300	Applies ELU activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::elu, rml_op_elu_params
RML_OP_EXP	1310	Calculates exponent of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::exp, rml_op_unary_params
RML_OP_FLATTEN	1330	Flattens an input tensor into a 2D matrix. Backend support: DirectML: Patial. Cannot handle some cases. MIOpen: Partial. Cannot handle some cases. Apple MPS: No. See also: rml_op_desc::flatten, rml_op_unary_params
RML_OP_FLOOR	1340	Floors elements of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::floor, rml_op_unary_params
RML_OP_GEMM	1360	Calculates gemm of input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::gemm, rml_op_gemm_params
RML_OP_IDENTITY	1380	Identity operator. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::identity, rml_op_unary_params
RML_OP_LEAKY_RELU	1390	Applies Leaky ReLU activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::leaky_relu, rml_op_leaky_relu_params
RML_OP_LOCAL_RESPONSE_NORM	1410	Performs local response normalization of an input tensor. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::local_response_norm, rml_op_local_response_norm_params
RML_OP_LOG_SOFTMAX	1420	Calculates LogSoftMax of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::log_softmax, rml_op_unary_params
RML_OP_LOGN	1430	Calculates natural logarithm of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::logn, rml_op_unary_params
RML_OP_MAX	1440	Calculates maximum of input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::max, rml_op_binary_params
RML_OP_MIN	1450	Calculates minimum of input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::min, rml_op_binary_params
RML_OP_MUL	1460	Multiplies input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::mul, rml_op_binary_params
RML_OP_NEG	1470	Calculates neg of the input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::neg, rml_op_unary_params
RML_OP_PAD	1490	Describes a data reorganization operator that inflates the input tensor with zeroes (or some other value) on the edges Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::pad, rml_op_pad_params
RML_OP_PARAMETRIC_RELU	1510	Performs Parametric ReLU activation of an input tensor, elsement-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::parametric_relu, rml_op_binary_params
RML_OP_PLACEHOLDER	1520	Special operation that holds information about input data. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::placeholder, rml_op_placeholder_params
RML_OP_POOL_2D_AVG	1560	Performs 2D average pooling of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: No. See also: rml_op_desc::pool_2d_avg, rml_op_pool_2d_params
RML_OP_POOL_2D_GLOBAL_AVG	1570	Performs 2D global average pooling of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: No. See also: rml_op_desc::pool_2d_global_avg, rml_op_pool_2d_global_params
RML_OP_POOL_2D_MAX	1580	Performs 2D max pooling of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::pool_2d_max, rml_op_pool_2d_params
RML_OP_PORT	1625	Returns a specified output tensor of an input operation. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::port, rml_op_port_params
RML_OP_POW	1630	Calculates power of an input tensor, element-wise. Backend support: DirectML: Partial. Wrong results for negative base values. MIOpen: Yes. Apple MPS: Partial. Wrong results for negative base values. See also: rml_op_desc::pow, rml_op_pow_params
RML_OP_QUANTIZE_LINEAR	1640	Transforms a hig-precision input tensor into a low-precision output tensor. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::quantize_linear, rml_op_quantize_linear_params
RML_OP_RECIP	1650	Calculates reciprocal of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::recip, rml_op_unary_params
RML_OP_REDUCE_ADD	1660	Calculates sum of the input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_add, rml_op_reduce_params
RML_OP_REDUCE_ADD_SQUARE	1670	Calculates sum square of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_add_square, rml_op_reduce_params
RML_OP_REDUCE_ARGMAX	1680	Calculates maximum elements indices of input tensor elements along provided axis. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_argmax, rml_op_reduce_params
RML_OP_REDUCE_ARGMIN	1690	Calculates minimum elements indices of input tensor elements along provided axis. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_argmin, rml_op_reduce_params
RML_OP_REDUCE_AVG	1700	Calculates average of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_avg, rml_op_reduce_params
RML_OP_REDUCE_L1	1710	Calculates L1 norm of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_l1, rml_op_reduce_params
RML_OP_REDUCE_L2	1720	Calculates L2 norm of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_l2, rml_op_reduce_params
RML_OP_REDUCE_LOGN_ADD	1730	Calculates log sum of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_logn_add, rml_op_reduce_params
RML_OP_REDUCE_LOGN_ADD_EXP	1740	Calculates log sum exponent of input tensor element along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_logn_add_exp, rml_op_reduce_params
RML_OP_REDUCE_MAX	1750	Calculates maximum of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_max, rml_op_reduce_params
RML_OP_REDUCE_MIN	1760	Calculates minimum of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_min, rml_op_reduce_params
RML_OP_REDUCE_MUL	1770	Calculates product of input tensor elements along provided axes. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::reduce_mul, rml_op_reduce_params
RML_OP_RELU	1780	Applies ReLU activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::relu, rml_op_unary_params
RML_OP_RELU6	1790	Applies ReLU6 activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::relu6, rml_op_unary_params
RML_OP_RESHAPE	1800	Changes the shape of an input tensor without changing tensor data. Backend support: DirectML: Partial. Cannot handle some cases. MIOpen: Partial. Cannot handle some cases. Apple MPS: No. See also: rml_op_desc::reshape, rml_op_reshape_params
RML_OP_RESIZE_2D_NEAREST	1810	Resizes an input tensor along spatial dims using the nearest neighbor algorithm. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::resize_2d, rml_op_resize_2d_params
RML_OP_RESIZE_2D_BICUBIC	1820	Resizes an input tensor along spatial dims using the bicubic algorithm. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::resize_2d_bicubic, rml_op_resize_2d_params
RML_OP_RSQRT	1870	Calculates reciprocal of square root of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::rsqrt, rml_op_unary_params
RML_OP_SELU	1880	Applies Scaled ELU activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::selu, rml_op_selu_params
RML_OP_SHAPE	1890	Outputs the shape of an input tensor. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::shape, rml_op_unary_params
RML_OP_SIGMOID	1900	Applies Sigmoid activation of the input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Partial. Invalid result in case of fp16. See also: rml_op_desc::sigmoid, rml_op_unary_params
RML_OP_SIN	1910	Calculates sine of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::sin, rml_op_unary_params
RML_OP_SLICE	1920	Produces a strided slice of an input tensor along multiple axes. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: No. See also: rml_op_desc::slice, rml_op_slice_params
RML_OP_SOFTMAX	1930	Applies Softmax activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::softmax, rml_op_unary_params
RML_OP_SOFTPLUS	1950	Applies Softplus activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::softplus, rml_op_unary_params
RML_OP_SOFTSIGN	1960	Applies Softsign activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::softsign, rml_op_unary_params
RML_OP_SPACE_TO_DEPTH	1970	Rearranges blocks of spatial tensor data into channels. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::space_to_depth, rml_op_space_to_depth_params
RML_OP_SQRT	1990	Calculates square root of an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::sqrt, rml_op_unary_params
RML_OP_SQUEEZE	2000	Removes single-dimensional entries from the shape of a tensor. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::squeeze, rml_op_squeeze_params
RML_OP_STACK	2010	Stacks a list of tensors with rank R into single tensor with rank R+1. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: No. See also: rml_op_desc::stack, rml_op_stack_params
RML_OP_SUB	2020	Subtracts input tensors, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::sub, rml_op_binary_params
RML_OP_TAN	2030	Calculate tangent of the input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::tan, rml_op_unary_params
RML_OP_TANH	2040	Applies hyperbolic tangent activation to an input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Partial. Invalid result in case of fp16. See also: rml_op_desc::tanh, rml_op_unary_params
RML_OP_THRESHOLDED_RELU	2050	Applies Thresholded ReLU activation of the input tensor, element-wise. Backend support: DirectML: Yes. MIOpen: Yes. Apple MPS: Yes. See also: rml_op_desc::thresholded_relu, rml_op_thresholded_relu_params
RML_OP_TOP_K	2060	Retrieve the top-K largest or smallest elements along a specified axis. Backend support: DirectML: Yes. MIOpen: No. Apple MPS: No. See also: rml_op_desc::top_k, rml_op_top_k_params
RML_OP_TRANSPOSE	2070	Transposes an input tensor. Backend support: DirectML: Partial. MIOpen: Partial. Apple MPS: Partial. See also: rml_op_desc::transpose, rml_op_transpose_params
RML_OP_UNSQUEEZE	2080	Inserts dimensions of size 1 at specified axis indices. Backend support: DirectML: Yes. Some layout transitions are unsupported. MIOpen: No. Apple MPS: No. See also: rml_op_desc::unsqueeze, rml_op_unsqueeze_params