Merge branch 'amd-develop' into amd-master

include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp

@@ -588,6 +588,7 @@ struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -1012,6 +1013,7 @@ struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp

@@ -108,7 +108,8 @@ template <typename ALayout,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched,
           PipelineVersion PipelineVer = PipelineVersion::v1,
-          typename ComputeType        = FloatC>
+          typename ComputeTypeA       = FloatC,
+          typename ComputeTypeB       = ComputeTypeA>
 struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 {
     static constexpr auto I0 = Number<0>{};
@@ -547,8 +548,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         constexpr auto c_block_size =
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
-        return math::max((a_block_space_size_aligned * sizeof(ComputeType) +
-                          b_block_space_size_aligned * sizeof(ComputeType)),
+        return math::max((a_block_space_size_aligned * sizeof(ComputeTypeA) +
+                          b_block_space_size_aligned * sizeof(ComputeTypeB)),
                          c_block_size * sizeof(FloatCShuffle));
     }
 
@@ -750,7 +751,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 FloatA,
-                                                ComputeType,
+                                                ComputeTypeA,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -781,7 +782,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 FloatB,
-                                                ComputeType,
+                                                ComputeTypeB,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -809,13 +810,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-        constexpr index_t KPack =
-            math::max(math::lcm(AK1Number, BK1Number),
-                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1Number, BK1Number),
+            MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ComputeType,
+            ComputeTypeA,
+            ComputeTypeB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -833,10 +835,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ComputeTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ComputeTypeB*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp

@@ -495,6 +495,7 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             FloatAB,
+            FloatAB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp

@@ -494,6 +494,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle
             auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
                 TileMathThreadGroupSize,
                 ABDataType,
+                ABDataType,
                 FloatGemmAcc,
                 decltype(a_block_desc_ak0_m_ak1),
                 decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp

@@ -737,6 +737,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatABAdjusted,
                                                                 FloatABAdjusted,
                                                                 FloatAcc,
                                                                 decltype(a_k0_m_k1_block_desc),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp

@@ -490,6 +490,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
                                                                 FloatAB,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -424,6 +424,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
 
         using BlockwiseGemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatABAdjusted,
                                                                 FloatABAdjusted,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),
@@ -569,6 +570,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             FloatABAdjusted,
+            FloatABAdjusted,
             FloatAcc,
             decltype(a_block_desc_k0_m_k1),
             decltype(b_block_desc_k0_n_k1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -762,6 +762,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ComputeType,
+            ComputeType,
             FloatAcc,
             decltype(a_k0_m_k1_block_desc),
             decltype(b_k0_n_k1_block_desc),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp

@@ -451,6 +451,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
                                                                 FloatAB,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_ak0_m_ak1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp

@@ -471,6 +471,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
                                                                 FloatAB,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp

@@ -489,6 +489,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
                                                                 FloatAB,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),

include/ck/utility/amd_xdlops.hpp

@@ -419,5 +419,70 @@ struct intrin_mfma_f32_16x16x32f8f8<16, 16>
     }
 };
 #endif
+
+#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x16f8bf8;
+
+template <>
+struct intrin_mfma_f32_32x32x16f8bf8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const bf8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+        reg_c.template AsType<float16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(
+                bit_cast<long>(reg_a),
+                bit_cast<long>(reg_b),
+                reg_c.template AsType<float16_t>()[Number<0>{}],
+                0,
+                0,
+                0);
+#else
+        vector_type<f8_t, 8> reg_a_v(reg_a);
+        vector_type<bf8_t, 8> reg_b_v(reg_b);
+
+        static_for<0, 8, 1>{}([&](auto k) {
+            float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
+            float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<bf8_t>()[Number<k>{}]);
+
+            intrin_mfma_f32_32x32x2f32<32, 32>::Run(reg_a_f32, reg_b_f32, reg_c);
+        });
+#endif
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x32f8bf8;
+
+template <>
+struct intrin_mfma_f32_16x16x32f8bf8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const bf8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8(
+            bit_cast<long>(reg_a),
+            bit_cast<long>(reg_b),
+            reg_c.template AsType<float4_t>()[Number<0>{}],
+            0,
+            0,
+            0);
+#else
+        vector_type<f8_t, 8> reg_a_v(reg_a);
+        vector_type<bf8_t, 8> reg_b_v(reg_b);
+
+        static_for<0, 8, 1>{}([&](auto k) {
+            float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
+            float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<bf8_t>()[Number<k>{}]);
+
+            intrin_mfma_f32_16x16x4f32<16, 16>::Run(reg_a_f32, reg_b_f32, reg_c);
+        });
+#endif
+    }
+};
+#endif
 } // namespace ck
 #endif

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -29,6 +29,8 @@ using BF8 = ck::bf8_t;
 
 using Empty_Tuple = ck::Tuple<>;
 
+using BF16_Tuple = ck::Tuple<BF16>;
+
 using F16_Tuple     = ck::Tuple<F16>;
 using F16_F16_Tuple = ck::Tuple<F16, F16>;