[mlir][spirv] Add definition for GL Length #144041
Conversation
|
@llvm/pr-subscribers-mlir-spirv @llvm/pr-subscribers-mlir Author: Igor Wodiany (IgWod-IMG) ChangesFull diff: https://github.com/llvm/llvm-project/pull/144041.diff 3 Files Affected:
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVGLOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVGLOps.td
index 2ec61758ba8ef..8c4da9b2dce18 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVGLOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVGLOps.td
@@ -1160,6 +1160,46 @@ def SPIRV_GLFMixOp :
// -----
+def SPIRV_GLLengthOp : SPIRV_GLOp<"Length", 66, [
+ Pure,
+ TypesMatchWith<"result type must match operand element type",
+ "operand", "result",
+ "::mlir::getElementTypeOrSelf($_self)">
+ ]> {
+ let summary = "Return the length of a vector x";
+
+ let description = [{
+ Result is the length of vector x, i.e., sqrt(x[0]**2 + x[1]**2 + ...).
+
+ The operand x must be a scalar or vector whose component type is floating-point.
+
+ Result Type must be a scalar of the same type as the component type of x.
+
+ #### Example:
+
+ ```mlir
+ %2 = spirv.GL.Length %0 : vector<3xf32> -> f32
+ %3 = spirv.GL.Length %1 : f32 -> f32
+ ```
+ }];
+
+ let arguments = (ins
+ SPIRV_ScalarOrVectorOf<SPIRV_Float>:$operand
+ );
+
+ let results = (outs
+ SPIRV_Float:$result
+ );
+
+ let assemblyFormat = [{
+ $operand attr-dict `:` type($operand) `->` type($result)
+ }];
+
+ let hasVerifier = 0;
+}
+
+// -----
+
def SPIRV_GLDistanceOp : SPIRV_GLOp<"Distance", 67, [
Pure,
AllTypesMatch<["p0", "p1"]>,
diff --git a/mlir/test/Dialect/SPIRV/IR/gl-ops.mlir b/mlir/test/Dialect/SPIRV/IR/gl-ops.mlir
index 2b75767feaf92..f49be21fcf783 100644
--- a/mlir/test/Dialect/SPIRV/IR/gl-ops.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/gl-ops.mlir
@@ -983,3 +983,69 @@ func.func @unpack_half_2x16_scalar_out(%arg0 : i32) -> () {
%0 = spirv.GL.UnpackHalf2x16 %arg0 : i32 -> f32
return
}
+
+// -----
+
+//===----------------------------------------------------------------------===//
+// spirv.GL.Length
+//===----------------------------------------------------------------------===//
+
+func.func @length(%arg0 : f32) -> () {
+ // CHECK: spirv.GL.Length {{%.*}} : f32 -> f32
+ %0 = spirv.GL.Length %arg0 : f32 -> f32
+ return
+}
+
+func.func @lengthvec(%arg0 : vector<3xf32>) -> () {
+ // CHECK: spirv.GL.Length {{%.*}} : vector<3xf32> -> f32
+ %0 = spirv.GL.Length %arg0 : vector<3xf32> -> f32
+ return
+}
+
+// -----
+
+func.func @length_i32_in(%arg0 : i32) -> () {
+ // expected-error @+1 {{op operand #0 must be 16/32/64-bit float or vector of 16/32/64-bit float values of length 2/3/4/8/16, but got 'i32'}}
+ %0 = spirv.GL.Length %arg0 : i32 -> f32
+ return
+}
+
+// -----
+
+func.func @length_f16_in(%arg0 : f16) -> () {
+ // expected-error @+1 {{op failed to verify that result type must match operand element type}}
+ %0 = spirv.GL.Length %arg0 : f16 -> f32
+ return
+}
+
+// -----
+
+func.func @length_i32vec_in(%arg0 : vector<3xi32>) -> () {
+ // expected-error @+1 {{op operand #0 must be 16/32/64-bit float or vector of 16/32/64-bit float values of length 2/3/4/8/16, but got 'vector<3xi32>'}}
+ %0 = spirv.GL.Length %arg0 : vector<3xi32> -> f32
+ return
+}
+
+// -----
+
+func.func @length_f16vec_in(%arg0 : vector<3xf16>) -> () {
+ // expected-error @+1 {{op failed to verify that result type must match operand element type}}
+ %0 = spirv.GL.Length %arg0 : vector<3xf16> -> f32
+ return
+}
+
+// -----
+
+func.func @length_i32_out(%arg0 : vector<3xf32>) -> () {
+ // expected-error @+1 {{op result #0 must be 16/32/64-bit float, but got 'i32'}}
+ %0 = spirv.GL.Length %arg0 : vector<3xf32> -> i32
+ return
+}
+
+// -----
+
+func.func @length_vec_out(%arg0 : vector<3xf32>) -> () {
+ // expected-error @+1 {{op result #0 must be 16/32/64-bit float, but got 'vector<3xf32>'}}
+ %0 = spirv.GL.Length %arg0 : vector<3xf32> -> vector<3xf32>
+ return
+}
diff --git a/mlir/test/Target/SPIRV/gl-ops.mlir b/mlir/test/Target/SPIRV/gl-ops.mlir
index eacf36bfba9ce..832f7ea2fe314 100644
--- a/mlir/test/Target/SPIRV/gl-ops.mlir
+++ b/mlir/test/Target/SPIRV/gl-ops.mlir
@@ -128,6 +128,10 @@ spirv.module Logical GLSL450 requires #spirv.vce<v1.0, [Shader], []> {
%8 = spirv.GL.FindSMsb %arg3 : vector<3xi32>
// CHECK: {{%.*}} = spirv.GL.FindUMsb {{%.*}} : vector<3xi32>
%9 = spirv.GL.FindUMsb %arg3 : vector<3xi32>
+ // CHECK: {{%.*}} = spirv.GL.Length {{%.*}} : f32 -> f32
+ %10 = spirv.GL.Length %arg0 : f32 -> f32
+ // CHECK: {{%.*}} = spirv.GL.Length {{%.*}} : vector<3xf32> -> f32
+ %11 = spirv.GL.Length %arg1 : vector<3xf32> -> f32
spirv.Return
}
|
|
I have added the canonicalizer. I've considered whether I should put it in the GL canonicalizer, but since in here we transform GL op into GL op, the op availability should not be an issues. |
|
Build failures are unrelated to this change. The problem has been already reported in #144179. |
I looked at this a little bit more closely and I think gl canonicalization is separated out mostly to better organize the code, similar how we have group op definitions in .td files (unlike most dialect that put all ops together). I don't see any logic that decides availability, unless I missed something. One caveat is that gl canon patterns don't run with the generic |
|
The comment in the header file: implies to me that GLSL canonicalization is separate because of the availability reasons, or at least that's my interpretation. In this case that wouldn't be an issue.Having said that, I agree it does make sense to put it with other GL ops -- it is only useful if GL ops are available anyway and it will make the code cleaner. I'll update this PR. |
Oh, I understand what you mean by availability now. Right, we wouldn't want to canonicalize |
|
So these are the buckets of spirv canon patterns:
|
|
I'm glad we're on the same page now :) I think it makes more sense to have all GL canonicalization in one place both from code organization perspective and from user interface -- I think it'd be a bit confusing if some GL canonicalization is run with default canonicalization and some with GL specific canonicalization.
I guess from now on we should be consistent and in both cases run those pattern as a part of gl/cl canonicalization. |
No description provided.