Enhanced artificial intelligence for performance validation of core integraeted circuit features

1 . An apparatus for using artificial intelligence to validate performance of integrated circuit features, the apparatus comprising processing circuitry coupled to memory, the processing circuitry configured to:

extract, from instruction files, microinstructions source and destination registers;

generate a dependency graph based on the microinstructions source and destination registers, the dependency graph comprising macroinstructions as nodes and dependencies between the macroinstructions as edges between the nodes;

generate, based on the dependency graph, a frequency distribution of instructions from trace files, performance univariate autoregressive conditionally heteroscedastic (Perf uarch) stat files, and register transfer language (RTL) stat files, predictors for a machine learning model, the predictors comprising instruction frequencies, instruction dependencies, and knob values associated with enabling the integrated circuit features;

generate, based on the Perf uarch stat files and the RTL stat files, ratios of Perf uarch stats to RTL stats as target stat ratios;

generate, using the predictors and the machine learning model, predicted ratios of Perf uarch stats to RTL stats; and

generate, using greedy constrained optimization, based on the target stat ratios and the predicted ratios, recommended traces for debugging,

wherein the recommended traces are a subset of traces in the trace files.

2 . The apparatus of claim 1 , wherein the subset of the traces is a minimum number of traces from among the traces in the trace files to indicate all divergence between the target stat ratios and the predicted ratios.

3 . The apparatus of claim 2 , wherein the divergence is based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

4 . The apparatus of claim 1 , wherein the processing circuitry is further configured to:

sample the traces in the trace files multiple time intervals; and

generate, based on the sampling, matrices using respective traces of the traces in the trace files as rows and stats from the Perf uarch stat files and the RTL stat files as columns, wherein a respective entry of a matrix is a predicted ratio of a Perf uarch stat to a RTL stat for a respective trace,

wherein to generate the recommended traces for debugging is further based on the matrices.

5 . The apparatus of claim 4 , wherein the greedy constrained optimization comprises generating, based on the matrices, second matrices with respective matrix entries of 0 or 1 based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

6 . The apparatus of claim 1 , wherein the processing circuitry is further configured to:

apply a page rank algorithm to the dependency graph,

wherein the predictors are based on the page rank algorithm.

7 . The apparatus of claim 1 , wherein to generate the ratios of Perf uarch stats to RTL stats as target stat ratios is based on a sum of stats updated due to a computation or execution of instructions, stats updated due to known bugs in RTL code, and stats updated due to unknown bugs.

8 . A non-transitory computer-readable storage medium comprising instructions to cause processing circuitry of a device for using artificial intelligence to validate performance of integrated circuit features, upon execution of the instructions by the processing circuitry, to:

extract, from instruction files, microinstructions source and destination registers;

generate a dependency graph based on the microinstructions source and destination registers, the dependency graph comprising macroinstructions as nodes and dependencies between the macroinstructions as edges between the nodes;

generate, based on the dependency graph, a frequency distribution of instructions from trace files, performance univariate autoregressive conditionally heteroscedastic (Perf uarch) stat files, and register transfer language (RTL) stat files, predictors for a machine learning model, the predictors comprising instruction frequencies, instruction dependencies, and knob values associated with enabling the integrated circuit features;

generate, based on the Perf uarch stat files and the RTL stat files, ratios of Perf uarch stats to RTL stats as target stat ratios;

generate, using the predictors and the machine learning model, predicted ratios of Perf uarch stats to RTL stats; and

generate, using greedy constrained optimization, based on the target stat ratios and the predicted ratios, recommended traces for debugging,

wherein the recommended traces are a subset of traces in the trace files.

9 . The non-transitory computer-readable storage medium of claim 8 , wherein the subset of the traces is a minimum number of traces from among the traces in the trace files to indicate all divergence between the target stat ratios and the predicted ratios.

10 . The non-transitory computer-readable storage medium of claim 9 , wherein the divergence is based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

11 . The non-transitory computer-readable storage medium of claim 8 , wherein execution of the instructions further causes the processing circuitry to:

sample the traces in the trace files multiple time intervals; and

generate, based on the sampling, matrices using respective traces of the traces in the trace files as rows and stats from the Perf uarch stat files and the RTL stat files as columns, wherein a respective entry of a matrix is a predicted ratio of a Perf uarch stat to a RTL stat for a respective trace,

wherein to generate the recommended traces for debugging is further based on the matrices.

12 . The non-transitory computer-readable storage medium of claim 11 , wherein the greedy constrained optimization comprises generating, based on the matrices, second matrices with respective matrix entries of 0 or 1 based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

13 . The non-transitory computer-readable storage medium of claim 8 , wherein execution of the instructions further causes the processing circuitry to:

apply a page rank algorithm to the dependency graph,

wherein the predictors are based on the page rank algorithm.

14 . The non-transitory computer-readable storage medium of claim 8 , wherein to generate the ratios of Perf uarch stats to RTL stats as target stat ratios is based on a sum of stats updated due to a computation or execution of instructions, stats updated due to known bugs in RTL code, and stats updated due to unknown bugs.

15 . A method for using artificial intelligence to validate performance of integrated circuit features, the method comprising:

extracting, by processing circuitry of a device, from instruction files, microinstructions source and destination registers;

generating, by the processing circuitry, a dependency graph based on the microinstructions source and destination registers, the dependency graph comprising macroinstructions as nodes and dependencies between the macroinstructions as edges between the nodes;

generating, by the processing circuitry, based on the dependency graph, a frequency distribution of instructions from trace files, performance univariate autoregressive conditionally heteroscedastic (Perf uarch) stat files, and register transfer language (RTL) stat files, predictors for a machine learning model, the predictors comprising instruction frequencies, instruction dependencies, and knob values associated with enabling the integrated circuit features;

generating, by the processing circuitry, based on the Perf uarch stat files and the RTL stat files, ratios of Perf uarch stats to RTL stats as target stat ratios;

generating, by the processing circuitry, using the predictors and the machine learning model, predicted ratios of Perf uarch stats to RTL stats; and

generating, by the processing circuitry, using greedy constrained optimization, based on the target stat ratios and the predicted ratios, recommended traces for debugging,

wherein the recommended traces are a subset of traces in the trace files.

16 . The method of claim 15 , wherein the subset of the traces is a minimum number of traces from among the traces in the trace files to indicate all divergence between the target stat ratios and the predicted ratios.

17 . The method of claim 16 , wherein the divergence is based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

18 . The method of claim 15 , further comprising:

sampling the traces in the trace files multiple time intervals; and

generating, based on the sampling, matrices using respective traces of the traces in the trace files as rows and stats from the Perf uarch stat files and the RTL stat files as columns, wherein a respective entry of a matrix is a predicted ratio of a Perf uarch stat to a RTL stat for a respective trace,

wherein generating the recommended traces for debugging is further based on the matrices.

19 . The method of claim 18 , wherein the greedy constrained optimization comprises generating, based on the matrices, second matrices with respective matrix entries of 0 or 1 based on a difference between a respective target stat ratio and a respective predicted ratio being greater than a threshold divergence.

20 . The method of claim 15 , further comprising:

applying a page rank algorithm to the dependency graph,

wherein the predictors are based on the page rank algorithm.