[WebGPU] Disable automatic subgroup MatMul selection on gen-12hp

onnxruntime/core/providers/webgpu/vendor/intel/math/gemm_subgroup.cc

@@ -154,9 +154,14 @@ std::string CalculateAccStr(const ShaderIndicesHelper* batch_dims, int64_t eleme
 
 }  // namespace
 
+bool CanAutoSelectSubgroupGemm(std::string_view architecture) {
+  return architecture != gpu_arch::kGen12Hp;
+}
+
 bool CanApplySubgroup(const ComputeContext& context, int64_t M, int64_t N, int64_t K, bool transA, bool transB) {
   if (context.AdapterInfo().vendor == std::string_view{"intel"}) {
-    bool use_subgroup = context.HasFeature(wgpu::FeatureName::Subgroups) &&
+    bool use_subgroup = CanAutoSelectSubgroupGemm(context.AdapterInfo().architecture) &&
+                        context.HasFeature(wgpu::FeatureName::Subgroups) &&
                         M >= 64 && N >= 512 && K >= 32 && !transA && !transB;
     return use_subgroup;
   }

onnxruntime/core/providers/webgpu/vendor/intel/math/gemm_subgroup_utils.h

@@ -12,10 +12,14 @@ namespace webgpu {
 namespace intel {
 
 namespace gpu_arch {
+inline constexpr std::string_view kGen12Hp = "gen-12hp";
 inline constexpr std::string_view kXeLpg = "xe-lpg";
 inline constexpr std::string_view kXe3Lpg = "xe-3lpg";
 }  // namespace gpu_arch
 
+// Whether automatic algorithm selection may choose the subgroup GEMM/MatMul kernel on `architecture`.
+bool CanAutoSelectSubgroupGemm(std::string_view architecture);
+
 bool ShouldFoldMatMulSubgroupBatch(std::string_view architecture,
                                    int64_t m,
                                    bool can_fold_batch_into_m);

onnxruntime/core/providers/webgpu/vendor/intel/math/matmul_execution_planner.cc

@@ -5,6 +5,8 @@
 
 #include <utility>
 
+#include "core/providers/webgpu/vendor/intel/math/gemm_subgroup_utils.h"
+
 namespace onnxruntime {
 namespace webgpu {
 namespace intel {
@@ -18,6 +20,7 @@ std::optional<MatMulAlgorithm> IntelMatMulExecutionPlanner::SelectVendorAlgorith
     return MatMulAlgorithm::SubgroupMatrix;
   }
   if (params.has_subgroup_capability &&
+      CanAutoSelectSubgroupGemm(params.adapter_architecture) &&
       params.m >= 64 && params.n >= 512 && params.k >= 32) {
     return MatMulAlgorithm::Subgroup;
   }