GemmKernels

Collaboration diagram for GemmKernels:

Classes
class	GemmReluScalarMKNKStream< M, K, N, IS_RELU >
	Scalar implementation for MK*NK, streams input, outputs, weights, stores bias GemmReluScalarMKNKStream<7,36,10> total = 12214. More...
class	GemmReluScalarMKNK< M, K, N, IS_RELU >
	Scalar implementation for MK*NK, stores weights and biases, Running GemmReluScalarMKNK<7,36,10> total = 6515. More...
class	GemmReluScalarMKKN< M, K, N, IS_RELU >
	Scalar implementation for MK*KN, stores weights and biases, GemmReluScalarMKKN<7,36,10> total = 7798. More...
class	GemmReluMKKNTwoAccsStream< M, K, N, IS_RELU >
	Vector implementation for MK*KN, streams input, outputs, weights, stores bias, requires (2*K + N)*4 <= 24576 and N % 8 == 0 GemmReluMKKNStream<7,36,10> total = 2774 (slower than GemmReluMKKNStream for larger N) More...
class	GemmReluMKKNStream< M, K, N, IS_RELU >
	Vector implementation for MK*KN, streams input, outputs, weights, stores bias, requires (4*K + 3*N)*4 <= 24576 and N % 8 == 0 GemmReluMKKNStream<7,36,10> total = 3100. More...
class	GemmReluMKKN< M, K, N, IS_RELU >
	Vector implementation for MK*KN, stores weights and biases, requires K%4=0, N%8=0 GemmReluMKKN<7,36,10> total = 1470. More...

Detailed Description

Design notes

• Single acc/fpmac: 393 cycles, no fpmac interleaving
• upd_w v8 better than load v16 from pointer, allows interleaving: 406 -> 698
• interleaving ops: 406 -> 365
• kstep=4 to 8: 365 -> 339
• Typically K > N for downsampling, M=1 if each net does an instance
• If chunking by N, for N%16=0, K<=128, for N%8=0, K<=256