Systems, methods, and apparatuses for tile transpose
Embodiments detailed herein relate to matrix operations. In particular, support for a matrix transpose instruction is detailed. In some embodiments, decode circuitry to decode an instruction having fields for an opcode, a source matrix operand identifier, and a destination matrix operand identifier; and execution circuitry to execute the decoded instruction to transpose each row of elements of the identified source matrix operand into a corresponding column of the identified destination matrix operand are detailed.
1 . An apparatus comprising:
a memory controller to couple to a first memory;
a core to be coupled to the first memory via the memory controller, the core including execution circuitry to execute first instructions of a first instruction set architecture (ISA), the core to cause one or more matrix processing instructions of a second ISA to be communicated to a matrix operations accelerator via the first memory; and
the matrix operations accelerator coupled to the first memory via the memory controller, the matrix operations accelerator to execute the one or more matrix processing instructions of the second ISA to perform matrix operations with a source matrix tile comprising a first plurality of source data elements, the matrix operations accelerator integral to the core or coupled to the core via an interconnect, the matrix operations accelerator to:
receive the one or more matrix processing instructions of the second ISA from the first memory, at least one instruction of the one or more matrix processing instructions having fields for an opcode to indicate a matrix transpose operation and a source matrix tile operand identifier to indicate the source matrix tile to be transposed, the source matrix tile comprising a plurality of 32-bit data elements, and
execute the one or more matrix processing instructions to perform the matrix transpose operation on the source matrix tile to generate a result matrix tile comprising a transpose of the source matrix tile, the matrix transpose operation to transform each row of the 32-bit data elements of the source matrix tile into a corresponding column of 32-bit data elements of the result matrix tile.
2 . The apparatus of claim 1 , wherein the matrix operations accelerator further comprises one or more vector registers to store the source matrix tile and the result matrix tile.
3 . The apparatus of claim 2 , wherein the matrix operations accelerator further comprises a plurality of multiply-accumulate (MAC) circuits to perform parallel multiply and accumulate operations for a matrix multiplication.
4 . The apparatus of claim 3 , wherein the plurality of MAC circuits comprise a 2-dimensional (2D) grid of MAC circuits.
5 . The apparatus of claim 1 , wherein the matrix operations accelerator and one or more cores are integral to a single die.
6 . The apparatus of claim 1 , wherein the execution circuitry is to generate a fault upon a determination that the source matrix tile has a different number of rows than a number of columns of the result matrix tile or the source matrix tile has a different number of columns than a number of rows in the result matrix tile.
7 . The apparatus of claim 1 , further comprising a coherent memory interface to couple the matrix operations accelerator to the memory controller.
8 . A method comprising:
executing, by execution circuitry of a core that is coupled to a first memory via a memory controller, first instructions of a first instruction set architecture (ISA);
causing, by the core, one or more matrix processing instructions of a second ISA to be communicated to a matrix operations accelerator coupled to the first memory via the memory controller; and
executing, by the matrix operations accelerator, the one or more matrix processing instructions of the second ISA to perform matrix operations with a source matrix tile comprising a first plurality of source data elements, the matrix operations accelerator integral to the core or coupled to the core via an interconnect, the matrix operations accelerator to:
receive the one or more matrix processing instructions of the second ISA from the first memory, at least one instruction of the one or more matrix processing instructions having fields for an opcode to indicate a matrix transpose operation and a source matrix tile operand identifier to indicate the source matrix tile to be transposed, the source matrix tile comprising a plurality of 32-bit data elements, and
execute the one or more matrix processing instructions to perform the matrix transpose operation on the source matrix tile to generate a result matrix tile comprising a transpose of the source matrix tile, the matrix transpose operation to transform each row of the 32-bit data elements of the source matrix tile into a corresponding column of 32-bit data elements of the result matrix tile.
9 . The method of claim 8 , further comprising storing the source matrix tile and the result matrix tile in one or more vector registers of the matrix operations accelerator.
10 . The method of claim 9 , wherein the matrix operations accelerator further comprises a plurality of multiply-accumulate (MAC) circuits to perform parallel multiply and accumulate operations for a matrix multiplication.
11 . The method of claim 10 , wherein the plurality of MAC circuits comprise a 2-dimensional (2D) grid of MAC circuits.
12 . The method of claim 8 , wherein the matrix operations accelerator and one or more cores are integral to a single die.
13 . The method of claim 8 , further comprising generating, by the execution circuitry, a fault upon a determination that the source matrix tile has a different number of rows than a number of columns of the result matrix tile or the source matrix tile has a different number of columns than a number of rows in the result matrix tile.
14 . The method of claim 8 , further comprising coupling the matrix operations accelerator to the memory controller via a coherent memory interface.
15 . A non-transitory machine-readable medium having program code stored thereon which, when executed by a machine, causes the machine to perform operations comprising:
executing, by execution circuitry of a core that is coupled to a first memory via a memory controller, first instructions of a first instruction set architecture (ISA);
causing, by the core, one or more matrix processing instructions of a second ISA to be communicated to a matrix operations accelerator coupled to the first memory via the memory controller; and
executing, by the matrix operations accelerator, the one or more matrix processing instructions of the second ISA to perform matrix operations with a source matrix tile comprising a first plurality of source data elements, the matrix operations accelerator integral to the core or coupled to the core via an interconnect, the matrix operations accelerator to:
receive the one or more matrix processing instructions of the second ISA from the first memory, at least one instruction of the one or more matrix processing instructions having fields for an opcode to indicate a matrix transpose operation and a source matrix tile operand identifier to indicate the source matrix tile to be transposed, the source matrix tile comprising a plurality of 32-bit data elements, and
execute the one or more matrix processing instructions to perform the matrix transpose operation on the source matrix tile to generate a result matrix tile comprising a transpose of the source matrix tile, the matrix transpose operation to transform each row of the 32-bit data elements of the source matrix tile into a corresponding column of 32-bit data elements of the result matrix tile.
16 . The non-transitory machine-readable medium of claim 15 , wherein the operations further comprise storing the source matrix tile and the result matrix tile in one or more vector registers of the matrix operations accelerator.
17 . The non-transitory machine-readable medium of claim 16 , wherein the matrix operations accelerator further comprises a plurality of multiply-accumulate (MAC) circuits to perform parallel multiply and accumulate operations for a matrix multiplication.
18 . The non-transitory machine-readable medium of claim 17 , wherein the plurality of MAC circuits comprise a 2-dimensional (2D) grid of MAC circuits.
19 . The non-transitory machine-readable medium of claim 15 , wherein the matrix operations accelerator and one or more cores are integral to a single die.
20 . The non-transitory machine-readable medium of claim 15 , wherein the operations further comprise generating, by the execution circuitry, a fault upon a determination that the source matrix tile has a different number of rows than a number of columns of the result matrix tile or the source matrix tile has a different number of columns than a number of rows in the result matrix tile.
21 . The non-transitory machine-readable medium of claim 15 , wherein the operations further comprise coupling the matrix operations accelerator to the memory controller via a coherent memory interface.