Merge branch 'feature/add-splitkv-instance' into...

Merge branch 'feature/add-splitkv-instance' into feature/support-vllm-kcache-layout-add-splitkv-instance

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp

@@ -421,11 +421,11 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
             }
 
             hip_check_error(
-                hipMemcpyWithStream(arg.p_workspace_,
-                                    arg.gemm_kernel_args_.data(),
-                                    arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
-                                    hipMemcpyHostToDevice,
-                                    stream_config.stream_id_));
+                hipMemcpyAsync(arg.p_workspace_,
+                               arg.gemm_kernel_args_.data(),
+                               arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
+                               hipMemcpyHostToDevice,
+                               stream_config.stream_id_));
 
             float ave_time = 0;
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp

@@ -41,7 +41,7 @@ __global__ void
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
-    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
 
     GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
         karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
@@ -76,7 +76,7 @@ __global__ void
     __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
-    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
 
     GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
         karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
@@ -639,27 +639,27 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
 
     struct SplitKBatchOffset
     {
-        __device__ SplitKBatchOffset(Argument& karg)
+        __device__ SplitKBatchOffset(Argument& karg, index_t k_id)
         {
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
             {
-                a_k_split_offset = blockIdx.z * karg.KRead;
+                a_k_split_offset = k_id * karg.KRead;
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
             {
-                a_k_split_offset = blockIdx.z * karg.KRead * karg.StrideA;
+                a_k_split_offset = k_id * karg.KRead * karg.StrideA;
             }
 
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
             {
-                b_k_split_offset = blockIdx.z * karg.KRead * karg.StrideB;
+                b_k_split_offset = k_id * karg.KRead * karg.StrideB;
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
             {
-                b_k_split_offset = blockIdx.z * karg.KRead;
+                b_k_split_offset = k_id * karg.KRead;
             }
 
-            if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
+            if(k_id < karg.KBatch - 1)
             {
                 karg.K = karg.KRead;
             }

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_ab_scale.hpp

@@ -38,8 +38,7 @@ __global__ void
     // __attribute__((amdgpu_waves_per_eu(1, 1)))
     kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(

include/ck/utility/amd_buffer_addressing.hpp

@@ -549,8 +549,10 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
             (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
             (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
             (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
-            (is_same<T, f8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
-            (is_same<T, bf8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
+            (is_same<T, f8_fnuz_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
+            (is_same<T, bf8_fnuz_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
+            (is_same<T, fp8_storage_t>::value &&
+             (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
             (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
         "wrong! not implemented");
 
@@ -843,8 +845,8 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave,
 
 #else
 
-    vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
-        src_wave_buffer_resource, src_thread_addr_offset, 0);
+    vector_t tmp{amd_buffer_load_impl<scalar_t, vector_size, coherence>(
+        src_wave_buffer_resource, src_thread_addr_offset, 0)};
     return src_thread_element_valid ? tmp : vector_t(0);
 #endif
 }
@@ -873,8 +875,8 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave,
 
     constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
 
-    vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
-        src_wave_buffer_resource, src_thread_addr_offset, 0);
+    vector_t tmp{amd_buffer_load_impl<scalar_t, vector_size, coherence>(
+        src_wave_buffer_resource, src_thread_addr_offset, 0)};
 
     return src_thread_element_valid ? tmp : vector_t(customized_value);
 }

include/ck/utility/amd_xdlops.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 namespace ck {
-// Define the common macro for gfx94x models
+// Define the common macro for MI300 models
 #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
 #define __gfx94__
 #endif

include/ck/utility/math_v2.hpp

@@ -80,7 +80,7 @@ static inline __host__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
-static inline __host__ bool isnan(f8_t x) { return (x & 0x80); };
+static inline __host__ bool isnan(f8_t x) { return ck::fp8_is_nan(x); };
 
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
 static inline __host__ bool isnan(int4_t x)
@@ -531,7 +531,7 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
-static inline __device__ bool isnan(f8_t x) { return (x & 0x80); };
+static inline __device__ bool isnan(f8_t x) { return ck::fp8_is_nan(x); };
 
 static inline __device__ half_t sqrt(half_t x)
 {
@@ -611,7 +611,7 @@ inline __device__ int8_t neg<int8_t>(int8_t x)
 template <>
 inline __device__ half_t neg<half_t>(half_t x)
 {
-    return __hneg(x);
+    return __hneg(static_cast<__half>(x));
 };
 
 template <typename T>

include/ck/utility/random_gen.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
+#include "ck/ck.hpp"
+
 namespace ck {
 
 // Pseudo random number generator
@@ -23,7 +25,7 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
 }
 
 // version for fp16
-template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<half_t, T>{}, bool> = false>
+template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<_Float16, T>{}, bool> = false>
 __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
 {
     uint16_t x         = *(reinterpret_cast<uint16_t*>(&val));
@@ -38,9 +40,10 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
 }
 
 // return 0 if data is not fp16 or fp32
-template <typename T,
-          uint32_t seed_t,
-          std::enable_if_t<!(std::is_same<float, T>{} || std::is_same<half_t, T>{}), bool> = false>
+template <
+    typename T,
+    uint32_t seed_t,
+    std::enable_if_t<!(std::is_same<float, T>{} || std::is_same<_Float16, T>{}), bool> = false>
 __host__ __device__ uint32_t prand_generator(int id, T val, uint32_t seed = seed_t)
 {
     std::ignore = id;

include/ck/utility/type_convert.hpp

@@ -9,7 +9,7 @@
 #include "ck/utility/array.hpp"
 
 namespace ck {
-// Define the common macro for gfx94x models
+// Define the common macro for MI300 models
 #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
 #define __gfx94__
 #endif
@@ -100,6 +100,18 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
     return type_convert<bhalf_t>(x_fp32);
 }
 
+template <>
+inline __host__ __device__ constexpr f8_ocp_t type_convert<f8_ocp_t, int>(int x)
+{
+    return f8_ocp_t{type_convert<f8_ocp_t::data_type>(x)};
+}
+
+template <>
+inline __host__ __device__ constexpr bf8_ocp_t type_convert<bf8_ocp_t, int>(int x)
+{
+    return bf8_ocp_t{type_convert<bf8_ocp_t::data_type>(x)};
+}
+
 // Convert X to Y
 template <typename Y, typename X>
 __host__ __device__ constexpr Y type_convert_sp(X x)
@@ -163,7 +175,7 @@ __host__ __device__ constexpr Y f8_convert_sr(X x);
 
 // convert fp32 to fp8 with stochastic rounding
 template <>
-inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
+inline __host__ __device__ f8_fnuz_t f8_convert_sr<f8_fnuz_t, float>(float x)
 {
     constexpr int seed = 1254739;
     uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
@@ -189,33 +201,35 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     return utils::
-        cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
-                                                                                               rng);
+        cast_to_f8<float, f8_fnuz_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
 #endif
 }
 
 // convert fp16 to fp8 with stochastic rounding
 template <>
-inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t f8_convert_sr<f8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_sr<f8_t>(type_convert<float>(x));
+    return f8_convert_sr<f8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     constexpr int seed               = 1254739;
     uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             f8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp32 to bf8 with stochastic rounding
 template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_sr<bf8_fnuz_t, float>(float x)
 {
     constexpr int seed = 1254739;
     uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
@@ -240,28 +254,32 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
-    return utils::
-        cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<float,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp16 to bf8 with stochastic rounding
 template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_sr<bf8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_sr<bf8_t>(type_convert<float>(x));
+    return f8_convert_sr<bf8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     constexpr int seed               = 1254739;
     uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
@@ -271,7 +289,7 @@ __host__ __device__ constexpr Y f8_convert_rne(X x);
 
 // convert fp32 to fp8 with rounding to nearest even
 template <>
-inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
+inline __host__ __device__ f8_fnuz_t f8_convert_rne<f8_fnuz_t, float>(float x)
 {
 #if defined(__gfx94__)
     union
@@ -296,32 +314,34 @@ inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
     return utils::
-        cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
-                                                                                               rng);
+        cast_to_f8<float, f8_fnuz_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
 #endif
 }
 
 // convert fp16 to fp8 with rounding to nearest even
 template <>
-inline __host__ __device__ f8_t f8_convert_rne<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t f8_convert_rne<f8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_rne<f8_t>(type_convert<float>(x));
+    return f8_convert_rne<f8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             f8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp32 to bf8 with rounding to nearest even
 template <>
-inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_rne<bf8_fnuz_t, float>(float x)
 {
 #if defined(__gfx94__)
     union
@@ -345,44 +365,59 @@ inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<float,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp16 to bf8 with rounding to nearest even
 template <>
-inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_rne<bf8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_rne<bf8_t>(type_convert<float>(x));
+    return f8_convert_rne<bf8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
+#endif
+}
+
+// convert fp32 to fp8
+template <>
+inline __host__ __device__ f8_fnuz_t type_convert<f8_fnuz_t, float>(float x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_fnuz_t>(x);
+#else
+    return f8_convert_rne<f8_fnuz_t>(x);
 #endif
 }
 
 // convert fp32 to fp8
 template <>
-inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
+inline __host__ __device__ f8_ocp_t type_convert<f8_ocp_t, float>(float x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<f8_t>(x);
+    return f8_convert_sr<f8_ocp_t>(x);
 #else
-    return f8_convert_rne<f8_t>(x);
+    return f8_convert_rne<f8_ocp_t>(x);
 #endif
 }
 
 // convert fp8 to fp32
 template <>
-inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
+inline __host__ __device__ float type_convert<float, f8_fnuz_t>(f8_fnuz_t x)
 {
 #if defined(__gfx94__)
     float fval;
@@ -392,30 +427,44 @@ inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
     return fval;
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<f8_t, float, negative_zero_nan>(x);
+    return utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(x);
 #endif
 }
 
 template <>
-inline __host__ __device__ float2_t type_convert<float2_t, f8x2_t>(f8x2_t x)
+inline __host__ __device__ float2_t type_convert<float2_t, f8x2_fnuz_t>(f8x2_fnuz_t x)
 {
 #if defined(__gfx94__)
     const auto i16val = bit_cast<uint16_t>(x);
     return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, 0);
 #else
     constexpr bool negative_zero_nan = true;
-    const auto f8x2_v                = vector_type<f8_t, 2>(x);
+    const auto f8x2_v                = vector_type<f8_fnuz_t, 2>(x);
     vector_type<float, 2> f32x2_v;
     f32x2_v.template AsType<float>()(Number<0>{}) =
-        utils::cast_from_f8<f8_t, float, negative_zero_nan>(
-            f8x2_v.template AsType<f8_t>()[Number<0>{}]);
+        utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(
+            f8x2_v.template AsType<f8_fnuz_t>()[Number<0>{}]);
     f32x2_v.template AsType<float>()(Number<1>{}) =
-        utils::cast_from_f8<f8_t, float, negative_zero_nan>(
-            f8x2_v.template AsType<f8_t>()[Number<1>{}]);
+        utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(
+            f8x2_v.template AsType<f8_fnuz_t>()[Number<1>{}]);
     return f32x2_v.template AsType<float2_t>()[Number<0>{}];
 #endif
 }
 
+template <>
+inline __host__ __device__ float2_t type_convert<float2_t, f8x2_ocp_t>(f8x2_ocp_t x)
+{
+#if CK_OCP_FP8_CVT_FAST_PATH
+    return fp8_impl::cast_to_f32x2_from_f8x2<f8_ocp_t::default_interpret>(
+        x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
+#else
+    return float2_t{fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(
+                        x.AsType<fp8_storage_t>()[Number<0>{}]),
+                    fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(
+                        x.AsType<fp8_storage_t>()[Number<1>{}])};
+#endif
+}
+
 template <>
 inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
 {
@@ -428,42 +477,64 @@ inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
 
 // convert fp16 to fp8
 template <>
-inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t type_convert<f8_fnuz_t, half_t>(half_t x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<f8_t>(x);
+    return f8_convert_sr<f8_fnuz_t>(x);
 #else
-    return f8_convert_rne<f8_t>(x);
+    return f8_convert_rne<f8_fnuz_t>(x);
+#endif
+}
+
+// convert fp16 to fp8
+template <>
+inline __host__ __device__ f8_ocp_t type_convert<f8_ocp_t, half_t>(half_t x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_ocp_t>(x);
+#else
+    return f8_convert_rne<f8_ocp_t>(x);
 #endif
 }
 
 // convert fp8 to fp16
 template <>
-inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
+inline __host__ __device__ half_t type_convert<half_t, f8_fnuz_t>(f8_fnuz_t x)
 {
 #if defined(__gfx94__)
     // use native conversion to float and convert to fp16
     return type_convert<half_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x);
+    return utils::cast_from_f8<f8_fnuz_t, half_t, negative_zero_nan>(x);
+#endif
+}
+
+// convert fp32 to bf8
+template <>
+inline __host__ __device__ bf8_fnuz_t type_convert<bf8_fnuz_t, float>(float x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_fnuz_t>(x);
+#else
+    return f8_convert_rne<bf8_fnuz_t>(x);
 #endif
 }
 
 // convert fp32 to bf8
 template <>
-inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
+inline __host__ __device__ bf8_ocp_t type_convert<bf8_ocp_t, float>(float x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<bf8_t>(x);
+    return f8_convert_sr<bf8_ocp_t>(x);
 #else
-    return f8_convert_rne<bf8_t>(x);
+    return f8_convert_rne<bf8_ocp_t>(x);
 #endif
 }
 
 // convert bf8 to fp32
 template <>
-inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
+inline __host__ __device__ float type_convert<float, bf8_fnuz_t>(bf8_fnuz_t x)
 {
 #if defined(__gfx94__)
     float fval;
@@ -473,31 +544,42 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
     return fval;
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<bf8_t, float, negative_zero_nan>(x);
+    return utils::cast_from_f8<bf8_fnuz_t, float, negative_zero_nan>(x);
+#endif
+}
+
+// convert fp16 to bf8
+template <>
+inline __host__ __device__ bf8_fnuz_t type_convert<bf8_fnuz_t, half_t>(half_t x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_fnuz_t>(x);
+#else
+    return f8_convert_rne<bf8_fnuz_t>(x);
 #endif
 }
 
 // convert fp16 to bf8
 template <>
-inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_ocp_t type_convert<bf8_ocp_t, half_t>(half_t x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<bf8_t>(x);
+    return f8_convert_sr<bf8_ocp_t>(x);
 #else
-    return f8_convert_rne<bf8_t>(x);
+    return f8_convert_rne<bf8_ocp_t>(x);
 #endif
 }
 
 // convert bf8 to fp16
 template <>
-inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
+inline __host__ __device__ half_t type_convert<half_t, bf8_fnuz_t>(bf8_fnuz_t x)
 {
 #if defined(__gfx94__)
     // use native conversion to float and convert to fp16
     return type_convert<half_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<bf8_t, half_t, negative_zero_nan>(x);
+    return utils::cast_from_f8<bf8_fnuz_t, half_t, negative_zero_nan>(x);
 #endif
 }
 

include/ck_tile/README.md

-# ck_tile
+[Back to the main page](../../README.md)
+# Composable Kernel Tile
 ## concept
 `ck_tile` provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
  - tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.
@@ -44,5 +45,8 @@ our implementation of different device operators.
 **[ops/epilogue]**  
 epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
 
+**[ref]**  
+reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
+
 ## examples
 currently we put all ck_tile related example under [/example/ck_tile](/example/ck_tile/) folder. Please check each example's subfolder.

include/ck_tile/core.hpp

@@ -54,6 +54,7 @@
 #include "ck_tile/core/tensor/tile_window_linear.hpp"
 #include "ck_tile/core/tensor/tile_window_utils.hpp"
 #include "ck_tile/core/tensor/update_tile.hpp"
+#include "ck_tile/core/utility/amd_address_space.hpp"
 #include "ck_tile/core/utility/bit_cast.hpp"
 #include "ck_tile/core/utility/functional.hpp"
 #include "ck_tile/core/utility/functional_with_tuple.hpp"

include/ck_tile/core/utility/amd_address_space.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core/config.hpp"
+
+// Address Space for AMDGCN
+// https://llvm.org/docs/AMDGPUUsage.html#address-space
+
+namespace ck_tile {
+
+#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
+
+template <typename T>
+__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
+{
+    // cast a pointer in "Constant" address space (4) to "Generic" address space (0)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+template <typename T>
+__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
+{
+    // cast a pointer in "Generic" address space (0) to "Constant" address space (4)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck_tile

include/ck_tile/ops/flatmm.hpp

@@ -5,6 +5,7 @@
 
 #include "ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp"
 #include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp"
+#include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32_itl.hpp"
 #include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp

@@ -339,7 +339,7 @@ struct FmhaFwdSplitKVCombineKernel
                 number<FmhaPipeline::kAlignmentOacc>{},
                 number<1>{});
 
-            auto o_acc_dram_view = pad_tensor_view(
+            const auto o_acc_dram_view = pad_tensor_view(
                 o_acc_dram_naive,
                 make_tuple(number<1>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
                 sequence<false, kPadSeqLenQ, kPadHeadDimV>{});