Add fp8 xdl gemm

include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp

@@ -29,7 +29,9 @@ enum struct MfmaInstr
     mfma_i32_16x16x16i8,
     mfma_i32_32x32x16i8,
     mfma_i32_16x16x32i8,
-    mfma_f64_16x16x4f64
+    mfma_f64_16x16x4f64,
+    mfma_f32_32x32x16f8f8,
+    mfma_f32_16x16x32f8f8
 };
 
 template <MfmaInstr instr>
@@ -454,6 +456,50 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
     }
 };
 
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x16f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
 template <typename base_type, index_t MPerXdlops, index_t NPerXdlops>
 struct MfmaSelector
 {
@@ -594,6 +640,18 @@ struct MfmaSelector
     }
 #endif
 
+    template <>
+        static constexpr auto GetMfma<f8_t, 32, 32>()
+        {
+            return MfmaInstr::mfma_f32_32x32x16f8f8;
+        }
+
+        template <>
+        static constexpr auto GetMfma<f8_t, 16, 16>()
+        {
+            return MfmaInstr::mfma_f32_16x16x32f8f8;
+        }
+
     static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{};
 
     __host__ __device__ constexpr MfmaSelector()
@@ -794,7 +852,7 @@ struct XdlopsGemm
     {
         static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value ||
                           is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
-                          is_same<base_type, int8_t>::value,
+                          is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value,
                       "base base_type must be double, float, half, bfloat16, and int8_t!");
 
         static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {

include/ck/utility/amd_xdlops.hpp

@@ -354,5 +354,68 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
 #endif
     }
 };
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x16f8f8;
+
+template <>
+struct intrin_mfma_f32_32x32x16f8f8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__)
+        reg_c.template AsType<float16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
+                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<f8_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<f8_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_16x16x32f8f8;
+
+template <>
+struct intrin_mfma_f32_16x16x32f8f8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__)
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
+            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<f8_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<f8_t>()[Number<k>{}]);
+
+            intrin_mfma_f32_16x16x4f32<16, 16>::Run(reg_a_f32, reg_b_f32, reg_c);
+        });
+#endif
+    }
+};
 } // namespace ck
 #endif