|
onnx2versal
|
Multiinstance graph for MxK times KxN that stores weights and biases Requires KxN_RND weight, NCHUNK%8=0, N%4=0 Chunks KxN weights by N dimension into NCHUNK chunks. Each instance has max size = 16384 and 4096 bytes respectively. Places maximum of 3x3 tiles, 8 conv tiles surrounding concat tile (max AIE DMA input=8)
#include <graph_qgemm.h>
Inherits adf::graph.
Public Member Functions | |
| QgemmChunkNGraph (std::vector< TTPARAM > weights, std::vector< int32_t > bias, float x_scale, float w_scale, float y_scale, TT x_zero, TTPARAM w_zero, TT y_zero) | |
Public Attributes | |
| adf::kernel | k [CHUNK_COUNT] |
| ConcatStreamGraph< CONCAT, TT, CHUNK_COUNT, M, NCHUNK, N > | concat_g |
| adf::port< input > | pin [1] |
| adf::port< output > | pout [1] |
Static Public Attributes | |
| static const int | CHUNK_COUNT = (N + NCHUNK - 1) / NCHUNK |