[all-commits] [llvm/llvm-project] b2729f: [mlir][Vector] Add a vblendps-based impl for trans...

[Nicolas Vasilache via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: b2729fda60dbda595e7b5974279d8f860bce75ab
      https://github.com/llvm/llvm-project/commit/b2729fda60dbda595e7b5974279d8f860bce75ab
  Author: Nicolas Vasilache <nicolas.vasilache at gmail.com>
  Date:   2021-11-23 (Tue, 23 Nov 2021)

  Changed paths:
    M mlir/include/mlir/Dialect/X86Vector/Transforms.h
    M mlir/lib/Dialect/X86Vector/Transforms/AVXTranspose.cpp
    M mlir/test/Dialect/Vector/vector-transpose-lowering.mlir
    A mlir/test/Integration/Dialect/LLVMIR/CPU/X86/test-inline-asm-vector.mlir
    M mlir/test/lib/Dialect/Vector/CMakeLists.txt
    M mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
    M utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel

  Log Message:
  -----------
  [mlir][Vector] Add a vblendps-based impl for transpose8x8 (both intrin and inline_asm)

This revision follows up on the conversation titled:

```[llvm-dev] Understanding and controlling some of the AVX shuffle emission paths```

The revision adds a vblendps-based implementation for transpose8x8 and further distinguishes between and intrinsics and an inline_asm implementation.

This results in roughly 20% fewer cycles as reported by llvm-mca:

After this revision (intrinsic version, resolves to virtually identical assembly as per the llvm-dev discussion, no vblendps instruction is emitted):
```
Iterations:        100
Instructions:      5900
Total Cycles:      2415
Total uOps:        7300

Dispatch Width:    6
uOps Per Cycle:    3.02
IPC:               2.44
Block RThroughput: 24.0

Cycles with backend pressure increase [ 89.90% ]
Throughput Bottlenecks:
  Resource Pressure       [ 89.65% ]
  - SKXPort1  [ 0.04% ]
  - SKXPort2  [ 12.42% ]
  - SKXPort3  [ 12.42% ]
  - SKXPort5  [ 89.52% ]
  Data Dependencies:      [ 37.06% ]
  - Register Dependencies [ 37.06% ]
  - Memory Dependencies   [ 0.00% ]
```

After this revision (inline_asm version, vblendps instructions are indeed emitted):
```
Iterations:        100
Instructions:      6300
Total Cycles:      2015
Total uOps:        7700

Dispatch Width:    6
uOps Per Cycle:    3.82
IPC:               3.13
Block RThroughput: 20.0

Cycles with backend pressure increase [ 83.47% ]
Throughput Bottlenecks:
  Resource Pressure       [ 83.18% ]
  - SKXPort0  [ 14.49% ]
  - SKXPort1  [ 14.54% ]
  - SKXPort2  [ 19.70% ]
  - SKXPort3  [ 19.70% ]
  - SKXPort5  [ 83.03% ]
  - SKXPort6  [ 14.49% ]
  Data Dependencies:      [ 39.75% ]
  - Register Dependencies [ 39.75% ]
  - Memory Dependencies   [ 0.00% ]
```

An accessible copy of the conversation is available [here](https://gist.github.com/nicolasvasilache/68c7f34012584b0e00f335bcb374ede0).

Differential Revision: https://reviews.llvm.org/D114393