added instances for large N

library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp

@@ -23,6 +23,10 @@ template <typename XElementwise, typename YElementwise, index_t Rank, index_t Re
 using device_elementwise_normalization_f16_instances =
     std::tuple <
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel

test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp

@@ -23,7 +23,7 @@ class TestElementwiseLayernorm : public ::testing::Test
     {
         // M, N
         std::vector<std::vector<ck::index_t>> lengths = {
-            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}};
+            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}, {4096, 8192}};
 
         for(auto length : lengths)
         {