Visualization¶
The generated CostModelEvaluation object(s) (from e.g. the API call) can be visualized in multiple ways.
Temporal mapping¶
The temporal mapping can be visualized by a function which prints it to the terminal. The code block demonstrates how to use it:
from zigzag.utils import pickle_load
from zigzag.visualization.results.print_mapping import print_mapping
cmes = pickle_load("zigzag/visualization/list_of_cmes.pickle")
cme = cmes[0]
print_mapping(cme)
The function will show the loops of the temporal mapping and for each operand shows at which memory level it resides. For example:
********* Temporal Mapping - CostModelEvaluation(layer=LayerNode_0, core=1) *********
O (O): [[('FX', 11), ('FY', 11)], [('OY', 7), ('OY', 2), ('OX', 14), ('K', 12)], []]
W (I2): [[], [('FX', 11), ('FY', 11), ('OY', 7), ('OY', 2), ('OX', 14)], [('K', 12)]]
I (I1): [[('FX', 11), ('FY', 11), ('OY', 7), ('OY', 2), ('OX', 14), ('K', 12)], []]
-------------------------------------------------------------------------------------
Temporal Loops O W I
-------------------------------------------------------------------------------------
for K in [0:12) sram_2MB dram sram_2MB
-------------------------------------------------------------------------------------
for OX in [0:14) sram_2MB sram_32KB sram_2MB
-------------------------------------------------------------------------------------
for OY in [0:2) sram_2MB sram_32KB sram_2MB
-------------------------------------------------------------------------------------
for OY in [0:7) sram_2MB sram_32KB sram_2MB
-------------------------------------------------------------------------------------
for FY in [0:11) rf_2B sram_32KB sram_2MB
-------------------------------------------------------------------------------------
for FX in [0:11) rf_2B sram_32KB sram_2MB
-------------------------------------------------------------------------------------
The top loop is the outer-most for loop, where as the bottom loop is the inner-most. Going from bottom to top, loops are allocated to the innermost memories of the memory hierarchy for each operand. The names of the memories match the names of the MemoryInstance object used to create the memory level using the add_memory() call in the MemoryHierarchy.
Energy and latency breakdown¶
The energy and latency breakdown of a list of CostModelEvaluation objects can be plotted using the bar_plot_cost_model_evaluations_breakdown function:
from zigzag.utils import pickle_load
from zigzag.visualization.results.plot_cme import bar_plot_cost_model_evaluations_breakdown
cmes = pickle_load("zigzag/visualization/list_of_cmes.pickle")
bar_plot_cost_model_evaluations_breakdown(cmes, "outputs/breakdown.jpg")
This will produce a bar chart, for example:
