Merge branch 'develop' into lwpck-1815

example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp

@@ -83,14 +83,14 @@ using DeviceOpInstanceKKNN =
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   S<4, 32, 1>,
                                                                                   S<1, 0, 2>,
                                                                                   S<1, 0, 2>,
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   1,
                                                                                   1,
                                                                                   S<1, 64, 1, 2>,

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common_wmma.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 // kernel data types
 using InKernelDataType       = FP16;
@@ -23,4 +24,14 @@ using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
 
 #include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv);
+}

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common_wmma.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 // kernel data types
 using InKernelDataType       = I8;
@@ -23,4 +24,14 @@ using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
 
 #include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -163,4 +164,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -285,4 +286,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -71,7 +72,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +278,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
@@ -351,4 +352,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_cross_attention_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp

@@ -28,6 +28,7 @@ Example is GQA-4
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -299,4 +300,14 @@ using ReferenceGemm1Instance =
 
 #include "run_grouped_query_attention_forward_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp

@@ -26,6 +26,7 @@ Shazeer, Noam. “Fast Transformer Decoding: One Write-Head Is All You Need.”
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -284,4 +285,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm_
 
 #include "run_multi_query_attention_forward_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -71,7 +72,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +278,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
@@ -329,4 +330,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_self_attention_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp

@@ -3,6 +3,7 @@
 
 #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp"
 #include "common.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 using OutDataType      = FP16;
 using WeiDataType      = FP16;
@@ -31,4 +32,14 @@ using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDat
 
 #include "run_grouped_conv_bwd_data_example.inc"
 
-int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run_grouped_conv_bwd_data_example(argc, argv);
+}

example/CMakeLists.txt

@@ -67,7 +67,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
     endforeach()
     #Do not build any WMMA examples if gfx11 targets are not on the list
     foreach(source IN LISTS FILE_NAME)
-        if(NOT EX_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
+	if(NOT EX_TARGETS MATCHES "gfx11" AND NOT EX_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
             message("removing wmma example ${source} ")
             list(REMOVE_ITEM FILE_NAME "${source}")
         endif()
@@ -154,7 +154,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
     endforeach()
     #Do not build any WMMA examples if gfx11 targets are not on the list
     foreach(source IN LISTS FILE_NAME)
-        if(NOT EX_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
+	if(NOT EX_TARGETS MATCHES "gfx11" AND NOT EX_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
             message("removing wmma example ${source} ")
             list(REMOVE_ITEM FILE_NAME "${source}")
         endif()
@@ -181,7 +181,7 @@ endfunction(add_example_executable_no_testing EXAMPLE_NAME)
 # add all example subdir
 file(GLOB dir_list LIST_DIRECTORIES true *)
 FOREACH(subdir ${dir_list})
-    IF(IS_DIRECTORY "${subdir}")
+    if(IS_DIRECTORY "${subdir}" AND EXISTS "${subdir}/CMakeLists.txt")
         add_subdirectory(${subdir})
     ENDIF()
 ENDFOREACH()

example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py

@@ -271,7 +271,9 @@ class FmhaBwdApiPool:
                 per_hdim_case = per_hdim_case + FMHA_BWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_BWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
-
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_BWD_KERNEL_HEADER + FMHA_BWD_API.format(F_dispatch = per_dtypes)
 
 # GEMM0: Q@K=S^T

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -278,6 +278,9 @@ class FmhaFwdApiPool:
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
 
 @dataclass

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

@@ -331,6 +331,9 @@ class FmhaFwdSplitKVApiPool:
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_SPLITKV_API.format(F_dispatch = per_dtypes)
 
 @dataclass

include/ck/ck.hpp

@@ -69,6 +69,9 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
 #define __gfx11__
 #endif
+#if defined(__gfx1200__) || defined(__gfx1201__)
+#define __gfx12__
+#endif
 
 // buffer resource
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
@@ -77,7 +80,7 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx103__)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
-#elif defined(__gfx11__)
+#elif defined(__gfx11__) || defined(__gfx12__)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
@@ -89,7 +92,7 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16
 #define CK_USE_AMD_V_DOT4_I32_I8
-#elif defined(__gfx11__)
+#elif defined(__gfx11__) || defined(__gfx12__)
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16
 #define CK_USE_AMD_V_DOT4_I32_I8_GFX11
@@ -110,13 +113,6 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 #define CK_USE_AMD_MFMA_GFX940
 #endif
 
-// WMMA instruction
-#ifndef __HIP_DEVICE_COMPILE__ // for host code
-#define CK_USE_AMD_WMMA
-#elif defined(__gfx11__) // for GPU code
-#define CK_USE_AMD_WMMA
-#endif
-
 // buffer load
 #define CK_USE_AMD_BUFFER_LOAD 1
 

include/ck/host_utility/device_prop.hpp

@@ -84,4 +84,9 @@ inline bool is_gfx11_supported()
            ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx1103";
 }
 
+inline bool is_gfx12_supported()
+{
+    return ck::get_device_name() == "gfx1200" || ck::get_device_name() == "gfx1201";
+}
+
 } // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -487,7 +487,14 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
             // sync point.
             if constexpr(k.value != 0 || KPerInnerLoop == KPerThread)
             {
+#ifdef __gfx12__
+                asm volatile("\
+	        s_barrier_signal -1 \n \
+		s_barrier_wait -1 \
+		" ::);
+#else
                 asm volatile("s_barrier" ::);
+#endif
                 __builtin_amdgcn_sched_barrier(0);
             }
             static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {

include/ck/tensor_operation/gpu/device/device_gemm_streamk_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemm_Streamk_V2 : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t Streamk_sel,
+                        ck::index_t Grid_size,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck