block PadK specialization in attention kernels

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

@@ -194,6 +194,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
             MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
 
+    // FIXME: pad K
+    static_assert(!matrix_padder.PadK, "KPadding is currently not supported");
+
     static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
     {
         const auto a_grid_desc_mraw_kraw = [&]() {

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp

@@ -203,6 +203,9 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
         GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
             MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
 
+    // FIXME: pad K
+    static_assert(!matrix_padder.PadK, "KPadding is currently not supported");
+
     static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
     {
         const auto a_grid_desc_mraw_kraw = [&]() {

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp

@@ -26,7 +26,8 @@ using S = ck::Sequence<Is...>;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
-static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto GemmPadded =
+    ck::tensor_operation::device::GemmSpecialization::MNOPadding; // Padding K is currently flawed
 
 // c[g, m, n] = a[g, m, k] * b[g, n, k]
 using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =

test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp

@@ -107,6 +107,7 @@ TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16)
 
 using ck::tensor_operation::device::GemmSpecialization;
 
+// TODO: enable KPadding tests when it is implemented
 TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
 {
     int P = 120; // requires padding
@@ -117,19 +118,19 @@ TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
     EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
-    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    // EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
     // clang-format on
 }
 
@@ -138,12 +139,12 @@ TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMismatch)
     // IsSupported(M, N, K, O)
     // clang-format off
     EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
-    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
     // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
-    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
-    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    // EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
     // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
-    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
     // clang-format on
 }