Optimized corner turns for local storage and bandwidth reduction

A method, in a data processing system, for optimized corner turns for local storage and bandwidth reduction, the method comprising:

• allocating a first set of buffers in a local store for a first block matrix and the second block matrix, wherein the first set of buffers comprises at most three local store buffers and wherein each buffer within the first set of buffers stores at most one block of data;
• allocating a second set of buffers in the local store for a result block matrix, wherein the second set of buffers comprises at most two local store buffers and wherein each buffer within the second set of buffers stores at most one block of data;
• for a first column of the result block matrix in a block matrix multiplication operation:
• performing a direct memory access operation to load one block in a row of the first block matrix from a system memory into the first set of buffers in the local store;
• for each block in the first column of the result block matrix and each block in a column of the second block matrix:
• performing a direct memory access operation, employing multi-buffering buffering to hide data transfer latency, to load the block of the result block matrix from the system memory to the second set of buffers in the local store according to a visitation order;
• performing a direct memory access operation, employing multi-buffering to hide data transfer latency, to load the block of the column of the second block matrix from the system memory into the first set of buffers in the local store according to the visitation order;
• performing a multiplication operation on the block of the first block matrix and the block of the second block matrix to form a product;
• subtracting the product from the block of the result block matrix to form a result block of the result block matrix in the second set of buffers;
• and
• performing a direct memory access operation to store the result block from the second set of buffers in the local store to the system memory;
• reverse the visitation order;
• and
• repeating the block matrix multiplication operation for a next column of the result block matrix with the reversed visitation order reusing a last loaded block of the column of the second block matrix responsive to reversing the visitation order.