[Mlir-commits] [mlir] [mlir][spirv] Mark several SPIR-V TOSA Ext Inst ops as NoMemoryEffects (PR #191814)

[Davide Grohmann]

https://github.com/davidegrohmann updated https://github.com/llvm/llvm-project/pull/191814

>From cf2ce4eb228ee214378863c298a6ff66b755f819 Mon Sep 17 00:00:00 2001
From: Davide Grohmann <davide.grohmann at arm.com>
Date: Mon, 13 Apr 2026 15:19:07 +0200
Subject: [PATCH 1/2] [mlir][spirv] Mark several SPIR-V TOSA Ext Inst ops as
 NoMemoryEffects

Initially such ops were marked Pure wrongly since they could overflow
or underflow the accumulator and result in undefined behavior.

Signed-off-by: Davide Grohmann <davide.grohmann at arm.com>
Change-Id: I0148e504f752a3799732a3413f82e2398dbb9630
---
 .../mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td      | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
index 7fdd00ef2e03b..3a2a0e832b745 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
@@ -150,7 +150,7 @@ class SPIRV_TosaBoolElementwiseBinaryOp<string mnemonic, int opcode, list<Trait>
 }
 
 class SPIRV_TosaConvolutionOp<string mnemonic, int opcode, list<Trait> traits = []> :
-  SPIRV_TosaOpWithResult<mnemonic, opcode, !listconcat(traits, [Pure,
+  SPIRV_TosaOpWithResult<mnemonic, opcode, !listconcat(traits, [NoMemoryEffect,
     TypeConstraintImplicationOn<"input", I8, "output", [I32]>,
     TypeConstraintImplicationOn<"input", I16, "output", [I64]>,
     TypeConstraintImplicationOn<"input", BF16, "output", [BF16]>,
@@ -273,7 +273,7 @@ def SPIRV_TosaArgMaxOp : SPIRV_TosaOpWithResult<"ArgMax", 0, [Pure,
 }
 
 
-def SPIRV_TosaAvgPool2DOp : SPIRV_TosaOpWithResult<"AvgPool2D", 1, [Pure,
+def SPIRV_TosaAvgPool2DOp : SPIRV_TosaOpWithResult<"AvgPool2D", 1, [NoMemoryEffect,
   TypeImpliesAccType<"input", I8, ["INT32"]>,
   TypeImpliesAccType<"input", I16, ["INT32"]>,
   TypeImpliesAccType<"input", F16, ["FP16", "FP32"]>,
@@ -287,7 +287,8 @@ def SPIRV_TosaAvgPool2DOp : SPIRV_TosaOpWithResult<"AvgPool2D", 1, [Pure,
     window of size given by <kernel size> is passed over the input tensor, with
     the mean value being placed in the output tensor. When calculating the
     average, only the number of valid input tensor values, but not padding, are
-    used to calculate the divisor.
+    used to calculate the divisor. The behaviour is undefined if the accumulated
+    result overflows or underflows.
 
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_avg_pool2d
@@ -343,6 +344,8 @@ def SPIRV_TosaConv2DOp : SPIRV_TosaConvolutionOp<"Conv2D", 2> {
 
     Input and weight have respective zero point values provided in input_zp and weight_zp.
 
+    The behaviour is undefined if the accumulated result overflows or underflows.
+
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_conv2d
       * https://www.mlplatform.org/tosa/tosa_spec_1_0_1.html#_conv2d
@@ -396,6 +399,8 @@ def SPIRV_TosaConv3DOp : SPIRV_TosaConvolutionOp<"Conv3D", 3> {
 
     Input and weight have respective zero point values provided in input_zp and weight_zp.
 
+    The behaviour is undefined if the accumulated result overflows or underflows.
+
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_conv3d
       * https://www.mlplatform.org/tosa/tosa_spec_1_0_1.html#_conv3d
@@ -450,6 +455,8 @@ def SPIRV_TosaDepthwiseConv2DOp : SPIRV_TosaConvolutionOp<"DepthwiseConv2D", 4>
 
     Input and weight have respective zero point values provided in input_zp and weight_zp.
 
+    The behaviour is undefined if the accumulated result overflows or underflows.
+
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_depthwise_conv2d
       * https://www.mlplatform.org/tosa/tosa_spec_1_0_1.html#_depthwise_conv2d
@@ -546,7 +553,7 @@ def SPIRV_TosaFFT2DOp : SPIRV_TosaOpWithComplexResult<"FFT2D", 5, [Pure]> {
 }
 
 
-def SPIRV_TosaMatMulOp : SPIRV_TosaOpWithResult<"MatMul", 6, [Pure,
+def SPIRV_TosaMatMulOp : SPIRV_TosaOpWithResult<"MatMul", 6, [NoMemoryEffect,
   TypeConstraintImplicationOn<"A", I8, "output", [I32]>,
   TypeConstraintImplicationOn<"A", I16, "output", [I64]>,
   TypeConstraintImplicationOn<"A", BF16, "output", [F32]>,
@@ -558,6 +565,7 @@ def SPIRV_TosaMatMulOp : SPIRV_TosaOpWithResult<"MatMul", 6, [Pure,
   let description = [{
     Performs two dimensional matrix multiplications.
     A, B are the inputs with respective zero point values in A_zp, B_zp.
+    The behaviour is undefined if the accumulated result overflows or underflows.
 
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_matmul
@@ -708,6 +716,8 @@ def SPIRV_TosaTransposeConv2DOp : SPIRV_TosaConvolutionOp<"TransposeConv2D", 9>
 
     Input and weight have respective zero point values provided in input_zp and weight_zp.
 
+    The behaviour is undefined if the accumulated result overflows or underflows.
+
     References:
       * https://github.khronos.org/SPIRV-Registry/extended/TOSA.001000.1.html#_transpose_conv2d
       * https://www.mlplatform.org/tosa/tosa_spec_1_0_1.html#_transpose_conv2d

>From 3c303cf5f09142d508c86185d991c50ca982eb1a Mon Sep 17 00:00:00 2001
From: Davide Grohmann <davide.grohmann at arm.com>
Date: Tue, 14 Apr 2026 10:18:04 +0200
Subject: [PATCH 2/2] Enhance description for Matrix Multiplication operator

Adds detailed description for Matrix Multiplication operator.
---
 mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
index 3a2a0e832b745..2c086c9e48ebb 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTosaOps.td
@@ -564,7 +564,9 @@ def SPIRV_TosaMatMulOp : SPIRV_TosaOpWithResult<"MatMul", 6, [NoMemoryEffect,
 
   let description = [{
     Performs two dimensional matrix multiplications.
+
     A, B are the inputs with respective zero point values in A_zp, B_zp.
+
     The behaviour is undefined if the accumulated result overflows or underflows.
 
     References: