[Mlir-commits] [mlir] [mlir][Vector] Add a rewrite pattern for better low-precision ext(bit… (PR #65774)

[Oleksandr Alex Zinenko]

================
@@ -0,0 +1,205 @@
+// RUN: mlir-opt %s --test-transform-dialect-interpreter --split-input-file | FileCheck %s
+
+/// Inspect generated assembly and llvm-mca stats
+/// =============================================
+/// mlir-opt --test-transform-dialect-interpreter mlir/test/Dialect/Vector/vector-rewrite-narrow-types.mlir -test-transform-dialect-erase-schedule -test-lower-to-llvm | mlir-translate -mlir-to-llvmir | llc -o - -mcpu=skylake-avx512 | llvm-mca -mcpu=skylake-avx512 | less
+
+/// Rewriting the patterns is about 10x on skylake-ax512 faster according to llvm-mca.
+/// No benefit on -march=arm64 -mcpu=apple-a14 -mattr=+fullfp16 according to llvm-mca.
+
+!source_element_t = i8
+!quantized_element_t = i3
+!expanded_element_t = i16
+!final_element_t = f32
+!mst = memref<1234x!source_element_t>
+!vst = vector<40x!source_element_t>
+!vst_i1 = vector<320xi1>
+!vqt_i1 = vector<288xi1>
+!vqt = vector<96x!quantized_element_t>
+!bvqt = vector<2x3x96x!quantized_element_t>
+!bvtt = vector<2x3x96x!expanded_element_t>
+!vct = vector<96x!final_element_t>
+!bvct = vector<2x3x96x!final_element_t>
+!mtt = memref<1234x!final_element_t>
+
+// CHECK-LABEL: @f1(
+//  CHECK-SAME: %[[A:[a-z0-9]+]]: memref<1234xi8>
+//  CHECK-SAME: %[[IDX:[a-z0-9]+]]: index
+//  CHECK-SAME: %[[B:[a-z0-9]+]]: memref<1234xf32>
+func.func @f1(%m: !mst, %idx : index, %mf: !mtt) {
+
+//              CHECK: %[[MASK:.*]] = arith.constant dense<[
+// CHECK-SAME-COUNT-6: 7, 56, 448, 3584, 28672, 229376, 1835008, 14680064, 117440512, 939524096, 7516192768, 60129542144, 481036337152, 3848290697216, 30786325577728, -35184372088832
----------------
ftynse wrote:

A comment with the hex version of this would be helpful.

https://github.com/llvm/llvm-project/pull/65774