[midend/lib/Conversion/LowerLinalgToGemmini] Add pass support for gemmini to run E2E LeNet and some tests.

[shirohasuki]

[midend/lib/Conversion/LowerLinalgToGemmini] Add conv2d_nhwc_fhwc pass for gemmini and add relevant examples and tests.
[midend/lib/Conversion/LowerGemmini] Add scalar print Op for convenient testing and add relevant examples and tests.
[midend/lib/Dialect/Gemmini] Split gemmini conv's RISC and CISC instruction lowering into two separate passes. (Now it's default lowering to RISC type different from native-gemmini, maybe we need a option to config this. Both can pass the LENet test.)

Below are the RISC version and CISC version running LeNet E2E on firesim.

[shirohasuki]

The commits seem to be missing the endline? Although I can see them locally.....

[linuxlonelyeagle]

Maybe you don't need check-same, and feel that the content of check-smae is not important. The same is true for the following example.

[linuxlonelyeagle]

gemmini.tile_conv ({{.*}} ({{.*}} ...， Maybe better since you're not capturing the variable and then using it.

[linuxlonelyeagle]

Do not use auto,The data types here can be confusing for people who are not familiar with this code.

[linuxlonelyeagle]

For unsupported data types, you should return failure.

[linuxlonelyeagle]

Thanks for the work, I just did a quick review.

[linuxlonelyeagle]

use English.

[linuxlonelyeagle]

... context = &getContext();
patterns.add<...>(context);

[linuxlonelyeagle]

The comment should be more explicit and should be placed above the if statement.

[linuxlonelyeagle]

Ping @shirohasuki I don't seem to have approval authority. If no one approves this PR (CI cannot run, please @zhanghb97 in time.

[shirohasuki]

Thank you so much for the REVIEW! I'll modify it and @zhanghb97 for CI.

[linuxlonelyeagle]

Print Ops in Gemmini has always been coupled.Its logic should actually be memref -> affine -> scf -> llvm,In fact, Gemmini is just gemmini -> llvm,So maybe you can separate the printOp patterns into a new pass.You can see that the populate pattern functions from the upstream are introduced in the lower gemmini process.

[shirohasuki]

I think it's right, Separate the printOp pattern into a decoupled pass and it will be like this?

populateGemminiPrintLegalizeForLLVMExportPatterns()
populateGemminiLegalizeForLLVMExportPatterns()
then, affine -> scf -> llvm................

[linuxlonelyeagle]

I think it's right, Separate the printOp pattern into a decoupled pass and it will be like this?

populateGemminiPrintLegalizeForLLVMExportPatterns()
populateGemminiLegalizeForLLVMExportPatterns()
then, affine -> scf -> llvm................

populateGemminiLegalizeForLLVMExportPatterns() should only include patterns that are lower to the gemmini instructions. And populateGemminiPrintLegalizeForLLVMExportPatterns() should only include patterns for print ops, which lower print ops to MLIR upstream ops.and then use the upstream pass, which should be lower to llvm dialect.

gemmini.print => affine/scf
// then use spsteam pass
affine/scf => llvm dialect

[shirohasuki]

I think it's right, Separate the printOp pattern into a decoupled pass and it will be like this?

populateGemminiPrintLegalizeForLLVMExportPatterns()
populateGemminiLegalizeForLLVMExportPatterns()
then, affine -> scf -> llvm................

populateGemminiLegalizeForLLVMExportPatterns() should only include patterns that are lower to the gemmini instructions. And populateGemminiPrintLegalizeForLLVMExportPatterns() should only include patterns for print ops, which lower print ops to MLIR upstream ops.and then use the upstream pass, which should be lower to llvm dialect.

gemmini.print => affine/scf
// then use spsteam pass
affine/scf => llvm dialect

I got it!