[VectorDistribute] Flush denormals for attention reduction config (#22041)

We are already flushing denormals for intrinsic based attention
configuration, this patch does the same for subgroup reduction based
attention configuration

Author: Groverkss

compiler/src/iree/compiler/Codegen/LLVMGPU/KernelConfig.cpp

@@ -1933,15 +1933,19 @@ static LogicalResult setAttentionReductionConfig(
   decompositionConfig.emplace_back(IREE::LinalgExt::AttentionOp::getPVAttrStr(),
                                    pvAttrDict);
 
+  SmallVector<NamedAttribute, 1> pipelineAttrs;
+  setAttentionPipelineAttributes(target, pipelineAttrs);
+
   // Set attention decomposition control config.
   op.setDecompositionConfigAttr(b.getDictionaryAttr(decompositionConfig));
 
   auto configDict = b.getDictionaryAttr(attrs);
   auto loweringConfig = IREE::GPU::LoweringConfigAttr::get(context, configDict);
+  auto pipelineConfig = DictionaryAttr::get(context, pipelineAttrs);
 
   return setOpConfigAndEntryPointFnTranslation(
       entryPoint, op, loweringConfig, CodeGenPipeline::LLVMGPUVectorDistribute,
-      workgroupSize, targetSubgroupSize);
+      workgroupSize, targetSubgroupSize, pipelineConfig);
 
   return success();
 }

compiler/src/iree/compiler/Codegen/LLVMGPU/test/ROCDL/config_vector_distribute_reduction_gfx942.mlir

@@ -3,6 +3,7 @@
 // RUN:   --pass-pipeline="builtin.module(iree-llvmgpu-select-lowering-strategy)" %s | FileCheck %s
 
 // CHECK:      #iree_codegen.translation_info<pipeline = LLVMGPUVectorDistribute
+// CHECK-SAME: iree_codegen.denormal_fp_math_f32 = #iree_codegen.denormal_fp_math<"preserve-sign">
 
 #pipeline_layout = #hal.pipeline.layout<bindings = [
   #hal.pipeline.binding<storage_buffer>,