Merge branch 'develop' into codegen_hiprtc

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_ck_fp8.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/random_gen.hpp"
+#include "ck/utility/type.hpp"
+
+#ifdef CK_USE_FNUZ_FP8
+#define CK_USE_FNUZ_FP8 1
+#else
+#define CK_USE_FNUZ_FP8 0
+#endif
+
+#ifdef CK_USE_OCP_FP8
+#define CK_USE_OCP_FP8 1
+#else
+#define CK_USE_OCP_FP8 0
+#endif
+
+namespace ck {
+
+using f8_fnuz_t  = _BitInt(8);
+using bf8_fnuz_t = unsigned _BitInt(8);
+
+#if(defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx1200__) || \
+    defined(__gfx1201__)) &&                                                                     \
+    __HIP_DEVICE_COMPILE__
+#define CK_FP8_CVT_FAST_PATH 1
+#else
+#define CK_FP8_CVT_FAST_PATH 0
+#endif
+
+#if(defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
+#define CK_OCP_FP8_CVT_FAST_PATH 1
+#else
+#define CK_OCP_FP8_CVT_FAST_PATH 0
+#endif
+
+typedef unsigned char fp8_storage_t;
+
+/**
+ * \brief Describes FP8 interpretation
+ */
+enum class ck_fp8_interpretation_t
+{
+    CK_E4M3_OCP  = 0, // OCP E4M3
+    CK_E5M2_OCP  = 1, // OCP E5M2
+    CK_E4M3_FNUZ = 2, // FP8
+    CK_E5M2_FNUZ = 3, // BF8
+};
+
+/**
+ * \brief Describes saturation behavior
+ */
+enum class ck_saturation_t
+{
+    CK_NOSAT     = 0, // No saturation - replace with NaN or Inf
+    CK_SATFINITE = 1, // Saturate to finite
+};
+
+namespace fp8_impl {
+
+typedef fp8_storage_t fp8x2_storage_t __attribute__((ext_vector_type(2)));
+typedef float float2_t __attribute__((ext_vector_type(2)));
+
+__host__ __device__ static inline constexpr bool fnuz_f8_is_nan(f8_fnuz_t a)
+{
+    return static_cast<unsigned char>(a) == 0x80;
+}
+__host__ __device__ static inline constexpr bool fnuz_bf8_is_nan(bf8_fnuz_t a)
+{
+    return static_cast<unsigned char>(a) == 0x80;
+}
+
+__host__ __device__ static inline constexpr bool ocp_f8_is_nan(fp8_storage_t a)
+{
+    return (a & 0x7f) == 0x7f;
+}
+__host__ __device__ static inline constexpr bool ocp_bf8_is_nan(fp8_storage_t a)
+{
+    return (a & 0x7f) > 0x7c;
+}
+
+// The conversion function is from rocblas
+// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_hip_f8_impl.h#L220
+// This has been modified to handle double types as well
+template <typename T, int wm, int we, bool is_fnuz, bool clip = false>
+__host__ __device__ static inline T cast_from_f8(fp8_storage_t x)
+{
+    constexpr bool is_half   = __hip_internal::is_same<T, _Float16>::value;
+    constexpr bool is_float  = __hip_internal::is_same<T, float>::value;
+    constexpr bool is_double = __hip_internal::is_same<T, double>::value;
+    static_assert(is_half || is_float || is_double, "only half, float and double are supported");
+
+    constexpr int weo = is_half ? 5 : (is_float ? 8 : 11);
+    constexpr int wmo = is_half ? 10 : (is_float ? 23 : 52);
+
+    T fInf, fNegInf, fNaN, fNeg0, fmax, fmin;
+    if constexpr(is_half)
+    {
+        const unsigned short int ihInf    = 0x7C00;
+        const unsigned short int ihNegInf = 0xFC00;
+        const unsigned short int ihNaN    = 0x7C01;
+        const unsigned short int ihNeg0   = 0x8000;
+        /* Max number in e5m2 57344*/
+        const unsigned short int ifmax = 0x7B00;
+        const unsigned short int ifmin = 0xFB00;
+
+        fInf    = bit_cast<_Float16>(ihInf);
+        fNegInf = bit_cast<_Float16>(ihNegInf);
+        fNaN    = bit_cast<_Float16>(ihNaN);
+        fNeg0   = bit_cast<_Float16>(ihNeg0);
+        fmax    = bit_cast<_Float16>(ifmax);
+        fmin    = bit_cast<_Float16>(ifmin);
+    }
+    else if constexpr(is_float)
+    {
+        const unsigned int ifInf    = 0x7F800000;
+        const unsigned int ifNegInf = 0xFF800000;
+        const unsigned int ifNaN    = 0x7F800001;
+        const unsigned int ifNeg0   = 0x80000000;
+        /* Max number in e5m2 57344*/
+        const unsigned int ifmax = 0x47600000;
+        const unsigned int ifmin = 0xC7600000;
+
+        fInf    = bit_cast<float>(ifInf);
+        fNegInf = bit_cast<float>(ifNegInf);
+        fNaN    = bit_cast<float>(ifNaN);
+        fNeg0   = bit_cast<float>(ifNeg0);
+        fmax    = bit_cast<float>(ifmax);
+        fmin    = bit_cast<float>(ifmin);
+    }
+    else if constexpr(is_double)
+    {
+        const unsigned long long ifInf    = 0x7FF0000000000000ull;
+        const unsigned long long ifNegInf = 0xFFF0000000000000ull;
+        const unsigned long long ifNaN    = 0x7FF0000000000001ull;
+        const unsigned long long ifNeg0   = 0x8000000000000000ull;
+        /* Max number in e5m2 57344*/
+        const unsigned long long ifmax = 0x40EC000000000000ull;
+        const unsigned long long ifmin = 0xC0EC000000000000ull;
+
+        fInf    = bit_cast<double>(ifInf);
+        fNegInf = bit_cast<double>(ifNegInf);
+        fNaN    = bit_cast<double>(ifNaN);
+        fNeg0   = bit_cast<double>(ifNeg0);
+        fmax    = bit_cast<double>(ifmax);
+        fmin    = bit_cast<double>(ifmin);
+    }
+
+    if(x == 0)
+    {
+        return 0;
+    }
+
+    unsigned long long sign     = x >> 7;
+    unsigned long long mantissa = x & ((1 << wm) - 1);
+    int exponent                = (x & 0x7F) >> wm;
+    if constexpr(is_fnuz)
+    {
+        if(x == 0x80)
+        {
+            return fNaN;
+        }
+    }
+    else
+    {
+        if(x == 0x80)
+        {
+            return fNeg0;
+        }
+        if constexpr(we == 4)
+        { // e4m3
+            if((x & 0x7F) == 0x7F)
+            {
+                return fNaN;
+            }
+        }
+        else if((x & 0x7C) == 0x7C)
+        { // e5m2
+            if((x & 0x3) == 0)
+            {
+                if constexpr(clip)
+                {
+                    return sign ? fmin : fmax;
+                }
+                return sign ? fNegInf : fInf;
+            }
+            return fNaN;
+        }
+    }
+
+    typename __hip_internal::conditional<
+        sizeof(T) == 2,
+        unsigned short int,
+        typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
+            type>::type retval;
+
+    if constexpr(we == 5 && is_half && !is_fnuz)
+    {
+        retval = x << 8;
+        return bit_cast<T>(retval);
+    }
+
+    const int exp_low_cutoff = (1 << (weo - 1)) - (1 << (we - 1)) + 1 - (is_fnuz ? 1 : 0);
+
+    // subnormal input
+    if(exponent == 0)
+    {
+#if defined(__HIP_DEVICE_COMPILE__) && __HIP_DEVICE_COMPILE__
+        // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
+        int sh = 1 + __clz(mantissa) - (32 - wm);
+#else
+        int sh = 1 + __builtin_clz(mantissa) - (32 - wm);
+#endif
+        mantissa <<= sh;
+        exponent += 1 - sh;
+        mantissa &= ((1ull << wm) - 1);
+    }
+    exponent += exp_low_cutoff - 1;
+    mantissa <<= wmo - wm;
+
+    // subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
+    if(exponent <= 0)
+    {
+        mantissa |= 1 << wmo;
+        mantissa >>= 1 - exponent;
+        exponent = 0;
+    }
+
+    if constexpr(sizeof(T) == 2)
+        retval = (sign << 15) | (exponent << 10) | mantissa;
+    else if constexpr(sizeof(T) == 4)
+        retval = (sign << 31) | (exponent << 23) | mantissa;
+    else
+        retval = (sign << 63) | (static_cast<unsigned long long>(exponent) << 52) | mantissa;
+
+    return bit_cast<T>(retval);
+}
+
+#if CK_FP8_CVT_FAST_PATH
+template <ck_fp8_interpretation_t interpret>
+static __device__ float cast_to_f32_from_f8(fp8_storage_t v)
+{
+    union
+    {
+        unsigned int i32val;
+        unsigned char i8val[4];
+    } val;
+    val.i8val[0] = v;
+
+    static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ ||
+                      interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
+                      interpret == ck_fp8_interpretation_t::CK_E5M2_FNUZ ||
+                      interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
+                  "Only FNUZ and OCP interpretations are supported");
+
+    if constexpr((interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
+                 (interpret == ck_fp8_interpretation_t::CK_E4M3_OCP))
+    {
+        return __builtin_amdgcn_cvt_f32_fp8(val.i32val, 0);
+    }
+    else
+    {
+        return __builtin_amdgcn_cvt_f32_bf8(val.i32val, 0);
+    }
+}
+
+template <ck_fp8_interpretation_t interpret>
+static __device__ float2_t cast_to_f32x2_from_f8x2(fp8x2_storage_t v)
+{
+    const auto i16val = bit_cast<uint16_t>(v);
+
+    static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ ||
+                      interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
+                      interpret == ck_fp8_interpretation_t::CK_E5M2_FNUZ ||
+                      interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
+                  "Only FNUZ and OCP interpretations are supported");
+
+    if constexpr((interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
+                 (interpret == ck_fp8_interpretation_t::CK_E4M3_OCP))
+    {
+        return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, false);
+    }
+    else
+    {
+        return __builtin_amdgcn_cvt_pk_f32_bf8(i16val, false);
+    }
+}
+
+#endif
+
+} // namespace fp8_impl
+
+struct f8_ocp_t
+{
+    using data_type = fp8_storage_t;
+    data_type data;
+
+    static constexpr ck_saturation_t default_saturation = ck_saturation_t::CK_SATFINITE;
+    static constexpr ck_fp8_interpretation_t default_interpret =
+        ck_fp8_interpretation_t::CK_E4M3_OCP;
+
+    static constexpr unsigned int we = 4; // exponent width
+    static constexpr unsigned int wm = 3; // mantissa width
+
+    __host__ __device__ constexpr bool operator==(const f8_ocp_t& other) const
+    {
+        return (data == other.data) && (fp8_impl::ocp_f8_is_nan(data) == false); // NaN != NaN
+    }
+
+#if CK_USE_OCP_FP8
+    __host__ __device__ explicit operator float() const
+#else
+    __host__ explicit operator float() const
+#endif
+    {
+#if CK_OCP_FP8_CVT_FAST_PATH
+        return fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data);
+#else
+        return fp8_impl::cast_from_f8<float, wm, we, false>(
+            this->data); // XXX: clip==false must be consistent with operator _Float16
+#endif
+    }
+
+#if CK_USE_OCP_FP8
+    __host__ __device__ explicit operator _Float16() const
+#else
+    __host__ explicit operator _Float16() const
+#endif
+    {
+#if CK_OCP_FP8_CVT_FAST_PATH
+        return static_cast<_Float16>(fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data));
+#else
+        return fp8_impl::cast_from_f8<_Float16, wm, we, false>(
+            this->data); // XXX: clip==false must be consistent with operator float
+#endif
+    }
+};
+
+struct bf8_ocp_t
+{
+    using data_type = fp8_storage_t;
+    data_type data;
+
+    static constexpr ck_saturation_t default_saturation = ck_saturation_t::CK_SATFINITE;
+    static constexpr ck_fp8_interpretation_t default_interpret =
+        ck_fp8_interpretation_t::CK_E5M2_OCP;
+
+    static constexpr unsigned int we = 5; // exponent width
+    static constexpr unsigned int wm = 2; // mantissa width
+
+    __host__ __device__ constexpr bool operator==(const bf8_ocp_t& other) const
+    {
+        return (data == other.data) && (fp8_impl::ocp_bf8_is_nan(data) == false); // NaN != NaN
+    }
+
+#if CK_USE_OCP_FP8
+    __host__ __device__ explicit operator float() const
+
+#else
+    __host__ explicit operator float() const
+#endif
+    {
+#if defined(__gfx1200__) || defined(__gfx1201__)
+        return fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data);
+#else
+        return fp8_impl::cast_from_f8<float, wm, we, false>(
+            this->data); // XXX: clip==false must be consistent with operator _Float16
+#endif
+    }
+
+#if CK_USE_OCP_FP8
+    __host__ __device__ explicit operator _Float16() const
+#else
+    __host__ explicit operator _Float16() const
+#endif
+    {
+#if defined(__gfx1200__) || defined(__gfx1201__)
+        return static_cast<_Float16>(fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data));
+#else
+        return fp8_impl::cast_from_f8<_Float16, wm, we, false>(
+            this->data); // XXX: clip==false must be consistent with operator float
+#endif
+    }
+};
+
+template <typename T>
+__host__ __device__ static inline constexpr bool fp8_is_nan(T);
+
+template <>
+__host__ __device__ inline constexpr bool fp8_is_nan(f8_ocp_t a)
+{
+    return fp8_impl::ocp_f8_is_nan(a.data);
+}
+template <>
+__host__ __device__ inline constexpr bool fp8_is_nan(bf8_ocp_t a)
+{
+    return fp8_impl::ocp_bf8_is_nan(a.data);
+}
+template <>
+__host__ __device__ inline constexpr bool fp8_is_nan(f8_fnuz_t a)
+{
+    return fp8_impl::fnuz_f8_is_nan(a);
+}
+template <>
+__host__ __device__ inline constexpr bool fp8_is_nan(bf8_fnuz_t a)
+{
+    return fp8_impl::fnuz_bf8_is_nan(a);
+}
+
+template <typename T,
+          std::enable_if_t<std::is_same_v<T, bf8_ocp_t> || std::is_same_v<T, f8_ocp_t> ||
+                               std::is_same_v<T, bf8_fnuz_t> || std::is_same_v<T, f8_fnuz_t>,
+                           bool> = true>
+__host__ __device__ static inline constexpr bool fp8_is_inf(T)
+{
+    return false;
+}
+template <>
+__host__ __device__ inline constexpr bool fp8_is_inf(bf8_ocp_t a)
+{
+    return (a.data & 0x7f) == 0x7c;
+}
+
+namespace fp8_impl {
+
+// Assertions to check for supported conversion types
+#define __assert_ocp_support(interp)                                               \
+    {                                                                              \
+        if(interp != ck_fp8_interpretation_t::CK_E4M3_OCP &&                       \
+           interp != ck_fp8_interpretation_t::CK_E5M2_OCP)                         \
+        {                                                                          \
+            __hip_assert(false && "type is unsupported by current target device"); \
+        }                                                                          \
+    }
+#define __assert_fnuz_support(interp)                                              \
+    {                                                                              \
+        if(interp != ck_fp8_interpretation_t::CK_E4M3_FNUZ &&                      \
+           interp != ck_fp8_interpretation_t::CK_E5M2_FNUZ)                        \
+        {                                                                          \
+            __hip_assert(false && "type is unsupported by current target device"); \
+        }                                                                          \
+    }
+
+__host__ __device__ static inline void
+__is_interpret_supported([[maybe_unused]] ck_fp8_interpretation_t interp)
+{
+#if defined(__HIP_DEVICE_COMPILE__) && __HIP_DEVICE_COMPILE__
+#if CK_USE_OCP_FP8
+    __assert_ocp_support(interp);
+#endif
+#if CK_USE_FNUZ_FP8
+    __assert_fnuz_support(interp);
+#endif
+#endif
+}
+
+#if CK_FP8_CVT_FAST_PATH
+// The conversion function is from rocblas
+// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_float8.h#L79
+template <ck_fp8_interpretation_t interpret, bool saturate, bool stochastic_rounding = false>
+static __device__ fp8_storage_t cast_to_f8_from_f32(float v, unsigned int rng = 0)
+{
+    fp8_storage_t i8data;
+    union
+    {
+        float fval;
+        unsigned int i32val;
+        unsigned char i8val[4]; // NOTE: not endian independent
+    } val;
+
+    unsigned int ival = 0;
+    val.fval          = v;
+
+    if constexpr(saturate)
+    {
+        if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ)
+        {
+            if((val.i32val & 0x7F800000) != 0x7F800000)
+            { /// propagate NAN/INF, no clipping
+                val.fval = __builtin_amdgcn_fmed3f(val.fval, 240.0, -240.0);
+            }
+        }
+        else if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
+        { // OCP type
+            if((val.i32val & 0x7F800000) != 0x7F800000)
+            { /// propagate NAN/INF, no clipping
+                val.fval = __builtin_amdgcn_fmed3f(val.fval, 448.0, -448.0);
+            }
+        }
+        else
+        {
+            if((val.i32val & 0x7F800000) != 0x7F800000)
+            { /// propagate NAN/INF, no clipping
+                val.fval = __builtin_amdgcn_fmed3f(val.fval, 57344.0, -57344.0);
+            }
+        }
+    }
+
+    if constexpr(stochastic_rounding)
+    {
+        ival       = (interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
+                       (interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
+                         ? __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0)
+                         : __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
+        val.i32val = ival;
+        i8data     = val.i8val[0]; // little endian
+    }
+    else
+    { // RNE CVT
+        ival       = (interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
+                       (interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
+                         ? __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false)
+                         : __builtin_amdgcn_cvt_pk_bf8_f32(val.fval,
+                                                     val.fval,
+                                                     ival,
+                                                     false); // false -> WORD0
+        val.i32val = ival;
+        i8data     = val.i8val[0];
+    }
+    return i8data;
+}
+#endif // CK_FP8_CVT_FAST_PATH
+
+// The conversion function is from rocblas
+// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_hip_f8_impl.h#L39
+// This has been modified to add double types conversion as well
+template <typename T, int wm, int we, bool is_fnuz, bool clip = false, bool stoch = false>
+__host__ __device__ static inline fp8_storage_t cast_to_f8(T _x, unsigned int rng = 0)
+{
+    constexpr bool is_half   = __hip_internal::is_same<T, _Float16>::value;
+    constexpr bool is_float  = __hip_internal::is_same<T, float>::value;
+    constexpr bool is_double = __hip_internal::is_same<T, double>::value;
+    static_assert(is_half || is_float || is_double,
+                  "Only half, float and double can be cast to f8");
+
+    constexpr int mfmt = (sizeof(T) == 8) ? 52 : ((sizeof(T) == 4) ? 23 : 10);
+
+    using T_bitwise = typename __hip_internal::conditional<
+        sizeof(T) == 2,
+        unsigned short int,
+        typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
+            type>::type;
+    T_bitwise x_bitwise = bit_cast<T_bitwise>(_x);
+
+    unsigned long long x{x_bitwise};
+
+    unsigned long long head, mantissa;
+    int exponent, bias;
+    unsigned int sign;
+    unsigned long long fInf, mask;
+
+    if constexpr(sizeof(T) == 8)
+    {
+        head     = x & 0xFFF0000000000000ull;
+        mantissa = x & 0xFFFFFFFFFFFFFull;
+        exponent = (head >> 52) & 0x7FF;
+        sign     = head >> 63;
+        bias     = 1023;
+        fInf     = 0x7FF0000000000000ull;
+        mask     = 0x7FFFFFFFFFFFFFFFull;
+    }
+    else if constexpr(sizeof(T) == 4)
+    {
+        head     = x & 0xFF800000;
+        mantissa = x & 0x7FFFFF;
+        exponent = (head >> 23) & 0xFF;
+        sign     = head >> 31;
+        bias     = 127;
+        fInf     = 0x7F800000;
+        mask     = 0x7FFFFFFF;
+    }
+    else
+    {
+        head     = x & 0xFC00;
+        mantissa = x & 0x3FF;
+        exponent = (head >> 10) & 0x1F;
+        sign     = head >> 15;
+        bias     = 15;
+        fInf     = 0x7C00;
+        mask     = 0x7FFF;
+    }
+    unsigned int signed_inf = 0;
+    unsigned int nan        = 0;
+    if constexpr(is_fnuz)
+    {
+        signed_inf = clip ? ((sign << 7) + 0x7f) : 0x80;
+        nan        = 0x80;
+    }
+    else
+    {
+        if constexpr(we == 4)
+        { // e4m3
+            signed_inf = (sign << 7) + (clip ? 0x7e : 0x7f);
+        }
+        else
+        { // e5m2
+            signed_inf = (sign << 7) + (clip ? 0x7b : 0x7c);
+        }
+        nan = (sign << 7) + 0x7f;
+    }
+    // Max values
+    unsigned long long ifmax = 0;
+    if constexpr(sizeof(T) == 8)
+    {
+        if constexpr(we == 5)
+        { // 57344
+            ifmax = 0x40EC000000000000ull;
+        }
+        else
+        {
+            if constexpr(is_fnuz)
+            { // 240
+                ifmax = 0x406E000000000000ull;
+            }
+            else
+            { // 448
+                ifmax = 0x407C000000000000ull;
+            }
+        }
+    }
+    else if(sizeof(T) == 4)
+    {
+        if constexpr(we == 5)
+        {
+            ifmax = 0x47600000;
+        }
+        else
+        {
+            if constexpr(is_fnuz)
+            {
+                ifmax = 0x43700000;
+            }
+            else
+            {
+                ifmax = 0x43E00000;
+            }
+        }
+    }
+    else
+    {
+        if constexpr(we == 5)
+        {
+            ifmax = 0x7B00;
+        }
+        else
+        {
+            if constexpr(is_fnuz)
+            {
+                ifmax = 0x5B80;
+            }
+            else
+            {
+                ifmax = 0x5F00;
+            }
+        }
+    }
+    // Deal with inf and NaNs
+    if((x & fInf) == fInf)
+    {
+        if constexpr(is_fnuz)
+            return signed_inf;
+
+        return mantissa != 0 ? nan : signed_inf;
+    }
+
+    if((x & mask) > ifmax)
+    {
+        return signed_inf;
+    }
+
+    if(x == 0)
+    {
+        return 0;
+    }
+
+    // First need to check if it is normal or denorm as there is a difference of
+    // implicit 1 Then need to adjust the exponent to align with the F8 exponent,
+    // in the meanwhile, shift The mantissa. Then for stochastic rounding, add rng
+    // to mantissa and truncate. And for RNE, no need to add rng. Then probably
+    // need to check whether there is carry and adjust exponent and mantissa again
+
+    // For IEEE bias mode, the bias is 2^(k-1) -1 where k is the width of exponent
+    // bits
+    const int f8_bias                  = (1 << (we - 1)) - 1 + (is_fnuz ? 1 : 0);
+    const int f8_denormal_act_exponent = 1 - f8_bias; // actual exponent of f8 denormal
+    // act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
+    // f8_exponent is the converted f8 exponent with bias encoding
+    // exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
+    // the difference needs to be adjusted and mantissa shifted
+    int act_exponent, f8_exponent, exponent_diff;
+
+    if(exponent == 0)
+    { // fp32/fp16 is in denormal.
+        /* fp32 denormal is below 2^-127 so it is usually not a concern here, we
+    mostly concern fp16 here. In this case, f8 is usually in denormal. But there
+    could be exceptions. fp16 denormal has exponent bias 15 while bf8 with NANOO has
+    exponent bias 16. It means that there are some numbers in fp16 denormal but they
+    are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
+    where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8
+    (NANOO) normal. In this case, the fp16 mantissa should be shift left by 1  */
+        act_exponent  = exponent - bias + 1;
+        exponent_diff = f8_denormal_act_exponent -
+                        act_exponent; // actual exponent is exponent-bias+1 as it is denormal
+    }
+    else
+    { // fp32/fp16 is normal with implicit 1
+        act_exponent = exponent - bias;
+        if(act_exponent <= f8_denormal_act_exponent)
+        {
+            /* This is the case where fp32/fp16 is normal but it is in f8 denormal
+      range. For example fp8 nanoo mode, denormal exponent is -7, but if the fp32/fp16
+      actual exponent is -7, it is actually larger due to the implicit 1,
+      Therefore it needs to be adjust to -6 and mantissa shift right by 1.
+      So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
+            exponent_diff = f8_denormal_act_exponent - act_exponent;
+        }
+        else
+        {                      // both fp32/fp16 and f8 are in normal range
+            exponent_diff = 0; // exponent_diff=0 does not mean there is no difference
+                               // for this case, act_exponent could be larger. Just
+                               // that it does not need shift mantissa
+        }
+        mantissa += (1ull << mfmt); // Add the implicit 1 into mantissa
+    }
+
+    bool midpoint = (mantissa & ((1ull << (mfmt - wm + exponent_diff)) - 1)) ==
+                    (1ull << (mfmt - wm + exponent_diff - 1));
+    /* This part is a bit tricky. The judgment of whether it is a tie needs to be
+  done before we shift right as shift right could rip off some residual part and
+  make something not midpoint look like midpoint. For example, the fp16 number
+  0x1002 (0 00100 0000000010), it is larger than midpoint, but after shift right
+  by 4 bits, it would look like midpoint.
+  */
+
+    if(exponent_diff > 0)
+        mantissa >>= exponent_diff;
+    else if(exponent_diff == -1)
+        mantissa <<= -exponent_diff;
+    bool implicit_one = mantissa & (1ull << mfmt);
+    // if there is no implicit 1, it  means the f8 is denormal and need to adjust
+    // to denorm exponent
+    f8_exponent =
+        (act_exponent + exponent_diff) /*actual f8 exponent*/ + f8_bias - (implicit_one ? 0 : 1);
+
+    // Now we have the exponent and mantissa adjusted
+    unsigned long long drop_mask = (1ull << (mfmt - wm)) - 1;
+    bool odd =
+        mantissa & (1ull << (mfmt - wm)); // if the least significant bit that is not truncated is 1
+    mantissa +=
+        (stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1ull) : mantissa)) & drop_mask;
+
+    // Now we deal with overflow
+    if(f8_exponent == 0)
+    {
+        if((1ull << mfmt) & mantissa)
+        {
+            f8_exponent = 1; // denormal overflow to become normal, promote exponent
+        }
+    }
+    else
+    {
+        if((1ull << (mfmt + 1)) & mantissa)
+        {
+            mantissa >>= 1;
+            f8_exponent++;
+        }
+    }
+
+    mantissa >>= (mfmt - wm);
+
+    // above range: quantize to maximum possible float of the same sign
+    const int max_exp = (1 << we) - 1;
+    if(f8_exponent > max_exp)
+    {
+        if constexpr(clip)
+        {
+            mantissa    = (1 << wm) - 1;
+            f8_exponent = max_exp;
+        }
+        else
+        {
+            return signed_inf;
+        }
+    }
+
+    if(f8_exponent == 0 && mantissa == 0)
+        return is_fnuz ? 0 : (sign << 7);
+    mantissa &= (1 << wm) - 1;
+    return (sign << 7) | (f8_exponent << wm) | mantissa;
+}
+
+/**
+ * \brief convert float to @p fp8_storage_t
+ *
+ * \tparam interp interpretation of fp8
+ * \tparam sat saturation of fp8
+ * \param f float number
+ * \return fp8_storage_t
+ */
+template <ck_fp8_interpretation_t interp,
+          ck_saturation_t sat      = ck_saturation_t::CK_SATFINITE,
+          bool stochastic_rounding = false>
+#if CK_FP8_CVT_FAST_PATH
+__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
+{
+    __is_interpret_supported(interp);
+    uint32_t rng = 0;
+    if constexpr(stochastic_rounding)
+    {
+        constexpr int seed = 1254739;
+        rng                = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
+    }
+    return cast_to_f8_from_f32<interp, sat == ck_saturation_t::CK_SATFINITE, stochastic_rounding>(
+        f, rng);
+#else
+#if CK_USE_OCP_FP8
+__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
+{
+#else
+__host__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
+{
+#endif
+    uint32_t rng = 0;
+    if constexpr(stochastic_rounding)
+    {
+        constexpr int seed = 1254739;
+        rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
+    }
+
+    if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_FNUZ)
+    {
+        return cast_to_f8<float,
+                          3,
+                          4,
+                          true,
+                          sat == ck_saturation_t::CK_SATFINITE,
+                          stochastic_rounding>(f, rng);
+    }
+    else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_FNUZ)
+    {
+        return cast_to_f8<float,
+                          2,
+                          5,
+                          true,
+                          sat == ck_saturation_t::CK_SATFINITE,
+                          stochastic_rounding>(f, rng);
+    }
+    else if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
+    {
+        return cast_to_f8<float,
+                          3,
+                          4,
+                          false,
+                          sat == ck_saturation_t::CK_SATFINITE,
+                          stochastic_rounding>(f, rng);
+    }
+    else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
+    {
+        return cast_to_f8<float,
+                          2,
+                          5,
+                          false,
+                          sat == ck_saturation_t::CK_SATFINITE,
+                          stochastic_rounding>(f, rng);
+    }
+    else
+    {
+        __hip_assert(false && "FP8 type is not supported by current target device");
+        return 0;
+    }
+#endif // CK_FP8_CVT_FAST_PATH
+}
+
+/**
+ * \brief convert _Float16 to @p fp8_storage_t
+ *
+ * \tparam sat saturation of fp8
+ * \tparam interp interpretation of fp8
+ * \tparam stochastic_rounding switch between RNE and SR
+ * \param x _Float16 value
+ * \return fp8_storage_t
+ */
+template <ck_fp8_interpretation_t interp,
+          ck_saturation_t sat      = ck_saturation_t::CK_SATFINITE,
+          bool stochastic_rounding = false>
+#if CK_FP8_CVT_FAST_PATH || CK_USE_OCP_FP8
+__host__ __device__ static inline fp8_storage_t cvt_half_t_to_fp8(const _Float16 x)
+#else
+__host__ static inline fp8_storage_t cvt_half_t_to_fp8(const _Float16 x)
+#endif
+{
+    return cvt_float_to_fp8<interp, sat, stochastic_rounding>(static_cast<float>(x));
+}
+
+} // namespace fp8_impl
+
+// Declare a template function for fp8 conversion using RNE
+template <typename Y, typename X>
+__host__ __device__ constexpr Y f8_convert_rne(X x);
+
+// convert fp32 to fp8 with rounding to nearest even
+template <>
+inline __host__ __device__ f8_ocp_t f8_convert_rne<f8_ocp_t, float>(float x)
+{
+    return f8_ocp_t{
+        fp8_impl::cvt_float_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation>(x)};
+}
+
+// convert fp32 to bf8 with rounding to nearest even
+template <>
+inline __host__ __device__ bf8_ocp_t f8_convert_rne<bf8_ocp_t, float>(float x)
+{
+    return bf8_ocp_t{
+        fp8_impl::cvt_float_to_fp8<bf8_ocp_t::default_interpret, bf8_ocp_t::default_saturation>(x)};
+}
+
+// convert _Float16 to fp8 with rounding to nearest even
+template <>
+inline __host__ __device__ f8_ocp_t f8_convert_rne<f8_ocp_t, _Float16>(_Float16 x)
+{
+    return f8_ocp_t{
+        fp8_impl::cvt_half_t_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation>(x)};
+}
+
+template <>
+inline __host__ __device__ bf8_ocp_t f8_convert_rne<bf8_ocp_t, _Float16>(_Float16 x)
+{
+    return bf8_ocp_t{
+        fp8_impl::cvt_half_t_to_fp8<bf8_ocp_t::default_interpret, bf8_ocp_t::default_saturation>(
+            x)};
+}
+
+// Declare a template function for fp8 conversion using RNE
+template <typename Y, typename X>
+__host__ __device__ constexpr Y f8_convert_sr(X x);
+
+// convert fp32 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_ocp_t f8_convert_sr<f8_ocp_t, float>(float x)
+{
+    return f8_ocp_t{
+        fp8_impl::cvt_float_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation, true>(
+            x)};
+}
+
+// convert fp32 to bf8 with stochastic rounding
+template <>
+inline __host__ __device__ bf8_ocp_t f8_convert_sr<bf8_ocp_t, float>(float x)
+{
+    return bf8_ocp_t{fp8_impl::cvt_float_to_fp8<bf8_ocp_t::default_interpret,
+                                                bf8_ocp_t::default_saturation,
+                                                true>(x)};
+}
+
+// convert _Float16 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_ocp_t f8_convert_sr<f8_ocp_t, _Float16>(_Float16 x)
+{
+    return f8_ocp_t{fp8_impl::cvt_half_t_to_fp8<f8_ocp_t::default_interpret,
+                                                f8_ocp_t::default_saturation,
+                                                true>(x)};
+}
+
+// convert _Float16 to bf8 with stochastic rounding
+template <>
+inline __host__ __device__ bf8_ocp_t f8_convert_sr<bf8_ocp_t, _Float16>(_Float16 x)
+{
+    return bf8_ocp_t{fp8_impl::cvt_half_t_to_fp8<bf8_ocp_t::default_interpret,
+                                                 bf8_ocp_t::default_saturation,
+                                                 true>(x)};
+}
+
+#if CK_USE_OCP_FP8
+using f8_t  = f8_ocp_t;
+using bf8_t = bf8_ocp_t;
+#define CK_FP8_TYPE_FNUZ 0
+#define CK_FP8_TYPE_OCP 1
+#else
+using f8_t = f8_fnuz_t;
+using bf8_t = bf8_fnuz_t;
+#define CK_FP8_TYPE_FNUZ 1
+#define CK_FP8_TYPE_OCP 0
+#endif
+
+} // namespace ck

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/data_type.hpp

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include "ck/utility/amd_ck_fp8.hpp"
 #include "ck/utility/statically_indexed_array.hpp"
 #ifdef CK_CODE_GEN_RTC
 using int8_t   = signed char;
@@ -22,8 +23,6 @@ using std::byte;
 using bhalf_t = ushort;
 using half_t  = _Float16;
 using int4_t  = _BitInt(4);
-using f8_t    = _BitInt(8);
-using bf8_t   = unsigned _BitInt(8);
 
 inline constexpr auto next_pow2(uint32_t x)
 {
@@ -31,14 +30,15 @@ inline constexpr auto next_pow2(uint32_t x)
     return x > 1u ? (1u << (32u - __builtin_clz(x - 1u))) : x;
 }
 
-// native types: double, float, _Float16, ushort, int32_t, int8_t, uint8_t, f8_t, bf8_t, bool
+// native types: double, float, _Float16, ushort, int32_t, int8_t, uint8_t, f8_fnuz_t, bf8_fnuz_t,
+// native types: bool
 template <typename T>
 inline constexpr bool is_native_type()
 {
     return is_same<T, double>::value || is_same<T, float>::value || is_same<T, half_t>::value ||
            is_same<T, bhalf_t>::value || is_same<T, int32_t>::value || is_same<T, int8_t>::value ||
-           is_same<T, uint8_t>::value || is_same<T, f8_t>::value || is_same<T, bf8_t>::value ||
-           is_same<T, bool>::value;
+           is_same<T, uint8_t>::value || is_same<T, f8_fnuz_t>::value ||
+           is_same<T, bf8_fnuz_t>::value || is_same<T, bool>::value;
 }
 
 // vector_type
@@ -178,16 +178,30 @@ struct scalar_type<int4_t>
 #endif
 
 template <>
-struct scalar_type<f8_t>
+struct scalar_type<f8_fnuz_t>
 {
-    using type                           = f8_t;
+    using type                           = f8_fnuz_t;
     static constexpr index_t vector_size = 1;
 };
 
 template <>
-struct scalar_type<bf8_t>
+struct scalar_type<bf8_fnuz_t>
 {
-    using type                           = bf8_t;
+    using type                           = bf8_fnuz_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<f8_ocp_t>
+{
+    using type                           = f8_ocp_t::data_type;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<bf8_ocp_t>
+{
+    using type                           = bf8_ocp_t::data_type;
     static constexpr index_t vector_size = 1;
 };
 
@@ -1022,60 +1036,203 @@ struct vector_type<T, 256, typename ck::enable_if_t<is_native_type<T>()>>
     }
 };
 
+template <typename T, index_t N, typename Enable = void>
+struct non_native_vector_base;
+
+template <typename T>
+struct nnvb_data_t_selector
+{
+    using type = unsigned _BitInt(8 * sizeof(T));
+};
+
+template <>
+struct nnvb_data_t_selector<f8_ocp_t>
+{
+    using type = f8_ocp_t::data_type;
+};
+template <>
+struct nnvb_data_t_selector<bf8_ocp_t>
+{
+    using type = bf8_ocp_t::data_type;
+};
+
+template <typename T, index_t N>
+struct non_native_vector_base<
+    T,
+    N,
+    std::enable_if_t<sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4 || sizeof(T) == 8>>
+{
+    using data_t = typename nnvb_data_t_selector<T>::type; // select data_t based on the size of T
+    static_assert(sizeof(T) == sizeof(data_t), "non_native_vector_base storage size mismatch");
+    using data_v = data_t __attribute__((ext_vector_type(N)));
+    using type   = non_native_vector_base<T, N>;
+
+    union alignas(next_pow2(N * sizeof(T)))
+    {
+        data_v dN; // storage vector;
+        StaticallyIndexedArray<data_t, N> dxN;
+        StaticallyIndexedArray<T, N> dTxN;
+        StaticallyIndexedArray<data_v, 1> dNx1;
+    } data_;
+
+    __host__ __device__ constexpr non_native_vector_base(data_t a) : data_{data_v(a)} {}
+    __host__ __device__ constexpr non_native_vector_base(T f)
+        : non_native_vector_base(bit_cast<data_t>(f))
+    {
+    }
+    __host__ __device__ constexpr non_native_vector_base() : non_native_vector_base(T{}){};
+    __host__ __device__ constexpr non_native_vector_base(data_v v) : data_{v} {}
+
+    __host__ __device__ constexpr operator data_v() const { return data_.dN; }
+    __host__ __device__ constexpr operator data_t() const
+    {
+        if constexpr(N == 1)
+        {
+            return data_.dxN[Number<0>{}];
+        }
+        else
+        {
+            return data_.dxN; // XXX this should cause an error
+        }
+    }
+    __host__ __device__ constexpr operator T() const
+    {
+        if constexpr(N == 1)
+        {
+            return data_.dTxN[Number<0>{}];
+        }
+        else
+        {
+            return data_.dTxN; // XXX this should cause an error
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same_v<X, data_t> || is_same_v<X, T> || is_same_v<X, data_v>,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same_v<X, data_t>)
+        {
+            return data_.dxN;
+        }
+        else if constexpr(is_same_v<X, T>)
+        {
+            return data_.dTxN;
+        }
+        else if constexpr(is_same_v<X, data_v>)
+        {
+            return data_.dNx1;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same_v<X, data_t> || is_same_v<X, T> || is_same_v<X, data_v>,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same_v<X, data_t>)
+        {
+            return data_.dxN;
+        }
+        else if constexpr(is_same_v<X, T>)
+        {
+            return data_.dTxN;
+        }
+        else if constexpr(is_same_v<X, data_v>)
+        {
+            return data_.dNx1;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
 template <typename T, index_t N>
-struct non_native_vector_base
+struct scalar_type<non_native_vector_base<T, N>>;
+
+template <index_t N>
+struct scalar_type<non_native_vector_base<f8_ocp_t, N>>
 {
-    using type = non_native_vector_base<T, N>;
+    using type = typename non_native_vector_base<f8_ocp_t, N>::data_t;
+
+    static constexpr index_t vector_size = N;
+};
 
-    __host__ __device__ non_native_vector_base()            = default;
-    __host__ __device__ non_native_vector_base(const type&) = default;
-    __host__ __device__ non_native_vector_base(type&&)      = default;
-    __host__ __device__ ~non_native_vector_base()           = default;
+template <index_t N>
+struct scalar_type<non_native_vector_base<bf8_ocp_t, N>>
+{
+    using type = typename non_native_vector_base<bf8_ocp_t, N>::data_t;
 
-    T d[N];
+    static constexpr index_t vector_size = N;
 };
 
 // non-native vector_type implementation
 template <typename T>
 struct vector_type<T, 1, typename ck::enable_if_t<!is_native_type<T>()>>
 {
-    using d1_t = T;
-    using type = d1_t;
+    using d1_t     = T;
+    using d1_nnv_t = non_native_vector_base<T, 1>;
+    using type     = d1_nnv_t;
 
     union alignas(next_pow2(1 * sizeof(T)))
     {
         d1_t d1_;
         StaticallyIndexedArray<d1_t, 1> d1x1_;
+        d1_nnv_t d1_nnv_;
     } data_;
 
-    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+    __host__ __device__ constexpr vector_type() : data_{d1_t{}} {}
 
     __host__ __device__ constexpr vector_type(type v) : data_{v} {}
 
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        return data_.d1x1_;
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
+        {
+            return data_.d1x1_;
+        }
+        else
+        {
+            return err;
+        }
     }
 
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        return data_.d1x1_;
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
+        {
+            return data_.d1x1_;
+        }
+        else
+        {
+            return err;
+        }
     }
 };
 
 template <typename T>
 struct vector_type<T, 2, typename ck::enable_if_t<!is_native_type<T>()>>
 {
-    using d1_t = T;
-    using d2_t = non_native_vector_base<T, 2>;
+    using d1_t     = T;
+    using d1_nnv_t = non_native_vector_base<T, 1>;
+    using d2_t     = non_native_vector_base<T, 2>;
 
     using type = d2_t;
 
@@ -1093,10 +1250,11 @@ struct vector_type<T, 2, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x2_;
         }
@@ -1113,10 +1271,11 @@ struct vector_type<T, 2, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x2_;
         }
@@ -1134,9 +1293,10 @@ struct vector_type<T, 2, typename ck::enable_if_t<!is_native_type<T>()>>
 template <typename T>
 struct vector_type<T, 4, typename ck::enable_if_t<!is_native_type<T>()>>
 {
-    using d1_t = T;
-    using d2_t = non_native_vector_base<T, 2>;
-    using d4_t = non_native_vector_base<T, 4>;
+    using d1_t     = T;
+    using d1_nnv_t = non_native_vector_base<T, 1>;
+    using d2_t     = non_native_vector_base<T, 2>;
+    using d4_t     = non_native_vector_base<T, 4>;
 
     using type = d4_t;
 
@@ -1155,10 +1315,11 @@ struct vector_type<T, 4, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x4_;
         }
@@ -1179,10 +1340,11 @@ struct vector_type<T, 4, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x4_;
         }
@@ -1204,10 +1366,11 @@ struct vector_type<T, 4, typename ck::enable_if_t<!is_native_type<T>()>>
 template <typename T>
 struct vector_type<T, 8, typename ck::enable_if_t<!is_native_type<T>()>>
 {
-    using d1_t = T;
-    using d2_t = non_native_vector_base<T, 2>;
-    using d4_t = non_native_vector_base<T, 4>;
-    using d8_t = non_native_vector_base<T, 8>;
+    using d1_t     = T;
+    using d1_nnv_t = non_native_vector_base<T, 1>;
+    using d2_t     = non_native_vector_base<T, 2>;
+    using d4_t     = non_native_vector_base<T, 4>;
+    using d8_t     = non_native_vector_base<T, 8>;
 
     using type = d8_t;
 
@@ -1227,11 +1390,12 @@ struct vector_type<T, 8, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
-                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x8_;
         }
@@ -1256,11 +1420,12 @@ struct vector_type<T, 8, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
-                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x8_;
         }
@@ -1286,11 +1451,12 @@ struct vector_type<T, 8, typename ck::enable_if_t<!is_native_type<T>()>>
 template <typename T>
 struct vector_type<T, 16, typename ck::enable_if_t<!is_native_type<T>()>>
 {
-    using d1_t  = T;
-    using d2_t  = non_native_vector_base<T, 2>;
-    using d4_t  = non_native_vector_base<T, 4>;
-    using d8_t  = non_native_vector_base<T, 8>;
-    using d16_t = non_native_vector_base<T, 16>;
+    using d1_t     = T;
+    using d1_nnv_t = non_native_vector_base<T, 1>;
+    using d2_t     = non_native_vector_base<T, 2>;
+    using d4_t     = non_native_vector_base<T, 4>;
+    using d8_t     = non_native_vector_base<T, 8>;
+    using d16_t    = non_native_vector_base<T, 16>;
 
     using type = d16_t;
 
@@ -1311,12 +1477,12 @@ struct vector_type<T, 16, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
-                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
-                          is_same<X, d16_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value || is_same<X, d16_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x16_;
         }
@@ -1345,12 +1511,12 @@ struct vector_type<T, 16, typename ck::enable_if_t<!is_native_type<T>()>>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
-                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
-                          is_same<X, d16_t>::value,
+        static_assert(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value ||
+                          is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value || is_same<X, d16_t>::value,
                       "Something went wrong, please check src and dst types.");
 
-        if constexpr(is_same<X, d1_t>::value)
+        if constexpr(is_same<X, d1_t>::value || is_same<X, d1_nnv_t>::value)
         {
             return data_.d1x16_;
         }
@@ -1644,20 +1810,70 @@ using int8x32_t = typename vector_type<int8_t, 32>::type;
 using int8x64_t = typename vector_type<int8_t, 64>::type;
 
 // f8
-using f8x2_t  = typename vector_type<f8_t, 2>::type;
-using f8x4_t  = typename vector_type<f8_t, 4>::type;
-using f8x8_t  = typename vector_type<f8_t, 8>::type;
-using f8x16_t = typename vector_type<f8_t, 16>::type;
-using f8x32_t = typename vector_type<f8_t, 32>::type;
-using f8x64_t = typename vector_type<f8_t, 64>::type;
+using f8x2_fnuz_t  = typename vector_type<f8_fnuz_t, 2>::type;
+using f8x4_fnuz_t  = typename vector_type<f8_fnuz_t, 4>::type;
+using f8x8_fnuz_t  = typename vector_type<f8_fnuz_t, 8>::type;
+using f8x16_fnuz_t = typename vector_type<f8_fnuz_t, 16>::type;
+using f8x32_fnuz_t = typename vector_type<f8_fnuz_t, 32>::type;
+using f8x64_fnuz_t = typename vector_type<f8_fnuz_t, 64>::type;
 
 // bf8
-using bf8x2_t  = typename vector_type<bf8_t, 2>::type;
-using bf8x4_t  = typename vector_type<bf8_t, 4>::type;
-using bf8x8_t  = typename vector_type<bf8_t, 8>::type;
-using bf8x16_t = typename vector_type<bf8_t, 16>::type;
-using bf8x32_t = typename vector_type<bf8_t, 32>::type;
-using bf8x64_t = typename vector_type<bf8_t, 64>::type;
+using bf8x2_fnuz_t  = typename vector_type<bf8_fnuz_t, 2>::type;
+using bf8x4_fnuz_t  = typename vector_type<bf8_fnuz_t, 4>::type;
+using bf8x8_fnuz_t  = typename vector_type<bf8_fnuz_t, 8>::type;
+using bf8x16_fnuz_t = typename vector_type<bf8_fnuz_t, 16>::type;
+using bf8x32_fnuz_t = typename vector_type<bf8_fnuz_t, 32>::type;
+using bf8x64_fnuz_t = typename vector_type<bf8_fnuz_t, 64>::type;
+
+// f8
+using f8x2_ocp_t  = typename vector_type<f8_ocp_t, 2>::type;
+using f8x4_ocp_t  = typename vector_type<f8_ocp_t, 4>::type;
+using f8x8_ocp_t  = typename vector_type<f8_ocp_t, 8>::type;
+using f8x16_ocp_t = typename vector_type<f8_ocp_t, 16>::type;
+using f8x32_ocp_t = typename vector_type<f8_ocp_t, 32>::type;
+using f8x64_ocp_t = typename vector_type<f8_ocp_t, 64>::type;
+
+// bf8
+using bf8x2_ocp_t  = typename vector_type<bf8_ocp_t, 2>::type;
+using bf8x4_ocp_t  = typename vector_type<bf8_ocp_t, 4>::type;
+using bf8x8_ocp_t  = typename vector_type<bf8_ocp_t, 8>::type;
+using bf8x16_ocp_t = typename vector_type<bf8_ocp_t, 16>::type;
+using bf8x32_ocp_t = typename vector_type<bf8_ocp_t, 32>::type;
+using bf8x64_ocp_t = typename vector_type<bf8_ocp_t, 64>::type;
+
+#if CK_FP8_TYPE_OCP
+// f8
+using f8x2_t  = f8x2_ocp_t;
+using f8x4_t  = f8x4_ocp_t;
+using f8x8_t  = f8x8_ocp_t;
+using f8x16_t = f8x16_ocp_t;
+using f8x32_t = f8x32_ocp_t;
+using f8x64_t = f8x64_ocp_t;
+
+// bf8
+using bf8x2_t  = bf8x2_ocp_t;
+using bf8x4_t  = bf8x4_ocp_t;
+using bf8x8_t  = bf8x8_ocp_t;
+using bf8x16_t = bf8x16_ocp_t;
+using bf8x32_t = bf8x32_ocp_t;
+using bf8x64_t = bf8x64_ocp_t;
+#elif CK_FP8_TYPE_FNUZ
+// f8
+using f8x2_t  = f8x2_fnuz_t;
+using f8x4_t  = f8x4_fnuz_t;
+using f8x8_t  = f8x8_fnuz_t;
+using f8x16_t = f8x16_fnuz_t;
+using f8x32_t = f8x32_fnuz_t;
+using f8x64_t = f8x64_fnuz_t;
+
+// bf8
+using bf8x2_t  = bf8x2_fnuz_t;
+using bf8x4_t  = bf8x4_fnuz_t;
+using bf8x8_t  = bf8x8_fnuz_t;
+using bf8x16_t = bf8x16_fnuz_t;
+using bf8x32_t = bf8x32_fnuz_t;
+using bf8x64_t = bf8x64_fnuz_t;
+#endif
 
 // u8
 using uint8x2_t  = typename vector_type<uint8_t, 2>::type;
@@ -1804,7 +2020,7 @@ struct NumericLimits<int4_t>
 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
 
 template <>
-struct NumericLimits<f8_t>
+struct NumericLimits<f8_fnuz_t>
 {
     // negative zero nan mode with exp bias = 8
     static constexpr uint8_t binary_min    = 0x08; // 0b00001000
@@ -1817,17 +2033,17 @@ struct NumericLimits<f8_t>
     // static constexpr uint8_t binary_lowest = 0xF7; // 0b11110111
     // static constexpr uint8_t binary_qnan   = 0x79; // any sign, exp=1111, mant!=0
 
-    __host__ __device__ static constexpr f8_t Min() { return f8_t(binary_min); }
+    __host__ __device__ static constexpr f8_fnuz_t Min() { return f8_fnuz_t(binary_min); }
 
-    __host__ __device__ static constexpr f8_t Max() { return f8_t(binary_max); }
+    __host__ __device__ static constexpr f8_fnuz_t Max() { return f8_fnuz_t(binary_max); }
 
-    __host__ __device__ static constexpr f8_t Lowest() { return f8_t(binary_lowest); }
+    __host__ __device__ static constexpr f8_fnuz_t Lowest() { return f8_fnuz_t(binary_lowest); }
 
-    __host__ __device__ static constexpr f8_t QuietNaN() { return f8_t(binary_qnan); }
+    __host__ __device__ static constexpr f8_fnuz_t QuietNaN() { return f8_fnuz_t(binary_qnan); }
 };
 
 template <>
-struct NumericLimits<bf8_t>
+struct NumericLimits<bf8_fnuz_t>
 {
     // negative zero nan mode with exp bias = 16
     static constexpr uint8_t binary_min    = 0x04; // 0b00000100
@@ -1840,13 +2056,59 @@ struct NumericLimits<bf8_t>
     // static constexpr uint8_t binary_lowest = 0xFB; // 0b11111011
     // static constexpr uint8_t binary_qnan   = 0x79; // any sign, exp=1111, mant!=
 
-    __host__ __device__ static constexpr bf8_t Min() { return bf8_t(binary_min); }
+    __host__ __device__ static constexpr bf8_fnuz_t Min() { return bf8_fnuz_t(binary_min); }
 
-    __host__ __device__ static constexpr bf8_t Max() { return bf8_t(binary_max); }
+    __host__ __device__ static constexpr bf8_fnuz_t Max() { return bf8_fnuz_t(binary_max); }
 
-    __host__ __device__ static constexpr bf8_t Lowest() { return bf8_t(binary_lowest); }
+    __host__ __device__ static constexpr bf8_fnuz_t Lowest() { return bf8_fnuz_t(binary_lowest); }
 
-    __host__ __device__ static constexpr bf8_t QuietNaN() { return bf8_t(binary_qnan); }
+    __host__ __device__ static constexpr bf8_fnuz_t QuietNaN() { return bf8_fnuz_t(binary_qnan); }
+};
+
+template <>
+struct NumericLimits<f8_ocp_t>
+{
+    static constexpr uint8_t binary_min    = 0x08; // 0b00001000 = 2^-6
+    static constexpr uint8_t binary_max    = 0x7E; // 0b01111110 = 448
+    static constexpr uint8_t binary_lowest = 0xFE; // 0b11111110 = -448
+    static constexpr uint8_t binary_qnan   = 0x7F; // 0b01111111
+
+    __host__ __device__ static constexpr f8_ocp_t Min() { return bit_cast<f8_ocp_t>(binary_min); }
+
+    __host__ __device__ static constexpr f8_ocp_t Max() { return bit_cast<f8_ocp_t>(binary_max); }
+
+    __host__ __device__ static constexpr f8_ocp_t Lowest()
+    {
+        return bit_cast<f8_ocp_t>(binary_lowest);
+    }
+
+    __host__ __device__ static constexpr f8_ocp_t QuietNaN()
+    {
+        return bit_cast<f8_ocp_t>(binary_qnan);
+    }
+};
+
+template <>
+struct NumericLimits<bf8_ocp_t>
+{
+    static constexpr uint8_t binary_min    = 0x04; // 0b00000100 = 2^-14
+    static constexpr uint8_t binary_max    = 0x7B; // 0b01111011 = 57344
+    static constexpr uint8_t binary_lowest = 0xFB; // 0b11111011 = -57344
+    static constexpr uint8_t binary_qnan   = 0x7D; // 0b01111101
+
+    __host__ __device__ static constexpr bf8_ocp_t Min() { return bit_cast<bf8_ocp_t>(binary_min); }
+
+    __host__ __device__ static constexpr bf8_ocp_t Max() { return bit_cast<bf8_ocp_t>(binary_max); }
+
+    __host__ __device__ static constexpr bf8_ocp_t Lowest()
+    {
+        return bit_cast<bf8_ocp_t>(binary_lowest);
+    }
+
+    __host__ __device__ static constexpr bf8_ocp_t QuietNaN()
+    {
+        return bit_cast<bf8_ocp_t>(binary_qnan);
+    }
 };
 
 template <typename T>
@@ -1889,7 +2151,7 @@ struct NumericUtils<half_t>
 };
 
 template <>
-struct NumericUtils<f8_t>
+struct NumericUtils<f8_fnuz_t>
 {
     static constexpr int exp  = 4;
     static constexpr int mant = 3;
@@ -1898,13 +2160,28 @@ struct NumericUtils<f8_t>
 };
 
 template <>
-struct NumericUtils<bf8_t>
+struct NumericUtils<bf8_fnuz_t>
 {
     static constexpr int exp  = 5;
     static constexpr int mant = 2;
     static constexpr int bias = 16; // negative zero nan mode
     // static constexpr int bias = 15; // ieee mode
 };
+template <>
+struct NumericUtils<f8_ocp_t>
+{
+    static constexpr int exp  = 4;
+    static constexpr int mant = 3;
+    static constexpr int bias = 7;
+};
+
+template <>
+struct NumericUtils<bf8_ocp_t>
+{
+    static constexpr int exp  = 5;
+    static constexpr int mant = 2;
+    static constexpr int bias = 15;
+};
 
 template <>
 struct NumericUtils<bhalf_t>

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)
 {

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/utility/ignore.hpp>
 
+#include "ck/ck.hpp"
+
 namespace ck {
 
 // Pseudo random number generator
@@ -24,7 +26,7 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
 }
 
 // version for fp16
-template <typename T, uint32_t seed_t, ck::enable_if_t<std::is_same<half_t, T>{}, bool> = false>
+template <typename T, uint32_t seed_t, ck::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));
@@ -39,9 +41,10 @@ __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,
-          ck::enable_if_t<!(std::is_same<float, T>{} || std::is_same<half_t, T>{}), bool> = false>
+template <
+    typename T,
+    uint32_t seed_t,
+    ck::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)
 {
     ck::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<size_t>(&x), x);
@@ -189,33 +201,36 @@ 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<size_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    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<size_t>(&x), x);
@@ -240,28 +255,33 @@ 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<size_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::cast_to_f8<half_t,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
@@ -271,7 +291,7 @@ __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 +316,34 @@ 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 +367,59 @@ 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 +429,44 @@ 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 +479,64 @@ 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_fnuz_t>(x);
+#else
+    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_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 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_t type_convert<bf8_t, float>(float x)
+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_t>(x);
+    return f8_convert_sr<bf8_fnuz_t>(x);
 #else
-    return f8_convert_rne<bf8_t>(x);
+    return f8_convert_rne<bf8_fnuz_t>(x);
+#endif
+}
+
+// convert fp32 to bf8
+template <>
+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_ocp_t>(x);
+#else
+    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 +546,42 @@ 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.

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.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
 
@@ -62,9 +62,9 @@ struct ReferenceGemm : public device::BaseOperator
             auto f_mk_kn_mn = [&](auto m, auto n) {
                 const int K = arg.a_m_k_.mDesc.GetLengths()[1];
 
-                AccDataType v_acc = 0;
-                ComputeTypeA v_a  = 0;
-                ComputeTypeB v_b  = 0;
+                AccDataType v_acc{0};
+                ComputeTypeA v_a{0};
+                ComputeTypeB v_b{0};
 
                 for(int k = 0; k < K; ++k)
                 {
@@ -93,7 +93,7 @@ struct ReferenceGemm : public device::BaseOperator
                         ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
                 }
 
-                CDataType v_c = 0;
+                CDataType v_c{0};
 
                 arg.c_element_op_(v_c, v_acc);
 

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -62,7 +62,7 @@ function(add_instance_library INSTANCE_NAME)
     endforeach()
     # Do not build mha instances if gfx94 or gfx90a targets are not on the target list
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
+	    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
          message("removing mha instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
@@ -346,7 +346,7 @@ if(CK_DEVICE_CONV_INSTANCES)
 endif()
 if(CK_DEVICE_MHA_INSTANCES)
         set(gpu_list ${INST_TARGETS})
-        if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
+	if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
             add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
             add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
             target_compile_features(device_mha_operations PUBLIC)

library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f8_instance.cpp

@@ -15,7 +15,7 @@ void add_device_pool3d_fwd_ndhwc_f8_instances(
         instances)
 {
     add_device_operation_instances(
-        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F8, ReduceOpId, false>{});
+        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F32, ReduceOpId, false>{});
 }
 
 void add_device_pool3d_fwd_ndhwc_index_f8_instances(
@@ -23,7 +23,7 @@ void add_device_pool3d_fwd_ndhwc_index_f8_instances(
         instances)
 {
     add_device_operation_instances(
-        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F8, ReduceOpId, true>{});
+        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F32, ReduceOpId, true>{});
 }
 
 } // namespace instance

profiler/README.md

+[Back to the main page](../README.md)
+# Composable Kernel profiler
 ## Profile GEMM kernels
 ```bash
 #arg1: tensor operation (gemm=GEMM)
@@ -180,3 +182,13 @@ Note: Column to image kernel adds to the output memory, this will cause output b
 ################            op datatype  verify  init  log  time  dim0 dim1 dim2 in_stride0 in_stride1 in_stride2 out_stride0 out_stride1 out_stride2
 ./bin/ckProfiler permute_scale        0       1     1    0     1    64   64   64       4096         64          1           1          64        4096
 ```
+
+## Convert MIOpen driver command to CKProfiler
+
+```bash
+python3 ../script/convert_miopen_driver_to_profiler.py
+/opt/rocm/bin/MIOpenDriver conv -n 32 -c 64 -H 28 -W 28 -k 64 -y 3 -x 3
+-p 1 -q 1 -u 2 -v 2 -l 1 -j 1 -m conv -g 32 -F 1 -t 1 
+```
+
+Only convolution driver is supported.

profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.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
 
@@ -150,7 +150,7 @@ bool profile_batched_gemm_bias_softmax_gemm_permute_impl(bool do_verification,
         break;
     default:
         a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
         d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
     }

profiler/include/profiler/profile_batched_gemm_gemm_impl.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
 
@@ -157,7 +157,7 @@ bool profile_batched_gemm_gemm_impl(bool do_verification,
         break;
     default:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.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
 
@@ -174,7 +174,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
         break;
     default:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.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
 
@@ -140,7 +140,7 @@ bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
         break;
     default:
         a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_gemm_impl.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
 
@@ -74,8 +74,8 @@ int profile_gemm_impl(int do_verification,
     switch(init_method)
     {
     case 0:
-        ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
-        ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
+        ck::utils::FillConstant<ADataType>{type_convert<ADataType>(1.f)}(a_m_k);
+        ck::utils::FillConstant<BDataType>{type_convert<BDataType>(1.f)}(b_k_n);
         break;
     case 1:
         ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);

test/data_type/CMakeLists.txt

@@ -9,13 +9,38 @@ if (USE_BITINT_EXTENSION_INT4)
   endif()
 endif()
 
-add_gtest_executable(test_fp8 test_fp8.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_fp8 PRIVATE utility)
+
+
+add_custom_target(test_fp8)
+
+if (CK_USE_OCP_FP8)
+  add_gtest_executable(test_fp8_ocp test_fp8_ocp.cpp)
+  if(result EQUAL 0)
+    target_link_libraries(test_fp8_ocp PRIVATE utility)
+  endif()
+
+  add_gtest_executable(test_bf8_ocp test_bf8_ocp.cpp)
+  if(result EQUAL 0)
+    target_link_libraries(test_bf8_ocp PRIVATE utility)
+  endif()
+
+  add_dependencies(test_fp8 test_fp8_ocp)
+  add_dependencies(test_fp8 test_bf8_ocp)
 endif()
-add_gtest_executable(test_bf8 test_bf8.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_bf8 PRIVATE utility)
+
+if (CK_USE_FNUZ_FP8)
+  add_gtest_executable(test_fp8_fnuz test_fp8_fnuz.cpp)
+  if(result EQUAL 0)
+    target_link_libraries(test_fp8_fnuz PRIVATE utility)
+  endif()
+
+  add_gtest_executable(test_bf8_fnuz test_bf8_fnuz.cpp)
+  if(result EQUAL 0)
+    target_link_libraries(test_bf8_fnuz PRIVATE utility)
+  endif()
+
+  add_dependencies(test_fp8 test_fp8_fnuz)
+  add_dependencies(test_fp8 test_bf8_fnuz)
 endif()
 
 add_gtest_executable(test_custom_type test_custom_type.cpp)

test/data_type/test_bf8.cpp → test/data_type/test_bf8_fnuz.cpp

@@ -5,158 +5,169 @@
 #include "ck/utility/data_type.hpp"
 #include "ck/utility/type_convert.hpp"
 
-using ck::bf8_t;
+using ck::bf8_fnuz_t;
 using ck::f8_convert_rne;
 using ck::f8_convert_sr;
 using ck::half_t;
 using ck::type_convert;
 
-TEST(BF8, NumericLimits)
+TEST(BF8FNUZ, NumericLimits)
 {
     // constants given for negative zero nan mode
-    EXPECT_EQ(ck::NumericLimits<bf8_t>::Min(), type_convert<bf8_t>(0x04));
-    EXPECT_EQ(ck::NumericLimits<bf8_t>::Max(), type_convert<bf8_t>(0x7F));
-    EXPECT_EQ(ck::NumericLimits<bf8_t>::Lowest(), type_convert<bf8_t>(0xFF));
-    EXPECT_EQ(ck::NumericLimits<bf8_t>::QuietNaN(), type_convert<bf8_t>(0x80));
+    EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Min(), type_convert<bf8_fnuz_t>(0x04));
+    EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Max(), type_convert<bf8_fnuz_t>(0x7F));
+    EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Lowest(), type_convert<bf8_fnuz_t>(0xFF));
+    EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(), type_convert<bf8_fnuz_t>(0x80));
 }
 
-TEST(BF8, ConvertFP32Nearest)
+TEST(BF8FNUZ, ConvertFP32Nearest)
 {
     // fix the tolerance value
     float abs_tol = 1e-6;
     // convert 0 float to bf8 and back, check if holds
-    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_t>(0.0f)), abs_tol);
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(0.0f)), abs_tol);
     // don't run the next test on gfx11 devices
 #ifndef CK_SKIP_FLAKY_F8_TEST
     // convert minimal float to bf8 and back, check if holds
     ASSERT_NEAR(std::numeric_limits<float>::min(),
-                type_convert<float>(f8_convert_rne<bf8_t>(std::numeric_limits<float>::min())),
+                type_convert<float>(f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::min())),
                 abs_tol);
 #endif
-    // convert maximal bf8_t to float and check if equal to 57344.0
-    ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_rne<bf8_t>(57344.0f)), abs_tol);
+
+    const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_fnuz_t>::Max());
+    // convert maximal bf8_fnuz_t to float and check if equal to 57344.0
+    ASSERT_NEAR(
+        max_bf8_t_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(max_bf8_t_float)), abs_tol);
     // convert maximal float to bf8 and back, check if clipped to 57344.0
-    ASSERT_NEAR(57344.0f,
-                type_convert<float>(f8_convert_rne<bf8_t>(std::numeric_limits<float>::max())),
+    ASSERT_NEAR(max_bf8_t_float,
+                type_convert<float>(f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::max())),
                 abs_tol);
-    // convert inf float to bf8_t and check if it is qNan
-    ASSERT_NEAR(type_convert<bf8_t>(0x80),
-                f8_convert_rne<bf8_t>(std::numeric_limits<float>::infinity()),
+    // convert inf float to bf8_fnuz_t and check if it is qNan
+    ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
+                f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::infinity()),
                 abs_tol);
     // positive norm float value to bf8 and back, check if holds
     float pos_float = 0.0000762939f;
-    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_t>(pos_float)), abs_tol);
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(pos_float)), abs_tol);
     // negative norm float value to bf8 and back, check if holds
     float neg_float = -0.0000610351f;
-    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_t>(neg_float)), abs_tol);
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(neg_float)), abs_tol);
     // positive subnorm float value to bf8 and back, check if holds
     pos_float = 0.0000305175f;
-    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_t>(pos_float)), abs_tol);
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(pos_float)), abs_tol);
     // negative subnorm float value to bf8 and back, check if holds
     neg_float = -0.0000152587f;
-    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_t>(neg_float)), abs_tol);
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(neg_float)), abs_tol);
 }
 
-TEST(BF8, ConvertFP32Stochastic)
+TEST(BF8FNUZ, ConvertFP32Stochastic)
 {
     // fix the tolerance value
     float abs_tol = 1e-6;
     // convert 0 float to bf8 and back, check if holds
-    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_t>(0.0f)), abs_tol);
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(0.0f)), abs_tol);
     // convert minimal float to bf8 and back, check if holds
     ASSERT_NEAR(std::numeric_limits<float>::min(),
-                type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::min())),
+                type_convert<float>(f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::min())),
                 abs_tol);
-    // convert maximal bf8_t to float and check if equal to 57344.0
-    ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_sr<bf8_t>(57344.0f)), abs_tol);
+
+    const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_fnuz_t>::Max());
+    // convert maximal bf8_fnuz_t to float and check if equal to 57344.0
+    ASSERT_NEAR(
+        max_bf8_t_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(max_bf8_t_float)), abs_tol);
     // convert maximal float to bf8 and back, check if clipped to 57344.0
-    ASSERT_NEAR(57344.0f,
-                type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::max())),
+    ASSERT_NEAR(max_bf8_t_float,
+                type_convert<float>(f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::max())),
                 abs_tol);
-    // convert inf float to bf8_t and check if it is qNan
-    ASSERT_NEAR(type_convert<bf8_t>(0x80),
-                f8_convert_sr<bf8_t>(std::numeric_limits<float>::infinity()),
+    // convert inf float to bf8_fnuz_t and check if it is qNan
+    ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
+                f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::infinity()),
                 abs_tol);
     // positive norm float value to bf8 and back, check if holds
     float pos_float = 0.0000762939f;
-    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(pos_float)), abs_tol);
     // negative norm float value to bf8 and back, check if holds
     float neg_float = -0.0000610351f;
-    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(neg_float)), abs_tol);
     // positive subnorm float value to bf8 and back, check if holds
     pos_float = 0.0000305175f;
-    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(pos_float)), abs_tol);
     // negative subnorm float value to bf8 and back, check if holds
     neg_float = -0.0000152587f;
-    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(neg_float)), abs_tol);
 }
 
-TEST(BF8, ConvertFP16Nearest)
+TEST(BF8FNUZ, ConvertFP16Nearest)
 {
     // fix the tolerance value
     float abs_tol = 1e-3;
     // convert 0 fp16 to bf8 and back, check if holds
-    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<bf8_t>(half_t{0.0})), abs_tol);
+    ASSERT_NEAR(
+        half_t{0.0}, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(half_t{0.0})), abs_tol);
     // convert minimal fp16 to bf8 and back, check if holds
     ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
-                type_convert<half_t>(f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::Min())),
+                type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
                 abs_tol);
-    // convert maximal bf8_t to fp16 and check if equal to 57344.0
+
+    const auto max_bf8_t_half = type_convert<half_t>(ck::NumericLimits<bf8_fnuz_t>::Max());
+    // convert maximal bf8_fnuz_t to fp16 and check if equal to 57344.0
     ASSERT_NEAR(
-        half_t{57344.0}, type_convert<half_t>(f8_convert_rne<bf8_t>(half_t{57344.0})), abs_tol);
+        max_bf8_t_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(max_bf8_t_half)), abs_tol);
     // convert maximal fp16 to bf8 and back, check if clipped to 57344.0
-    ASSERT_NEAR(half_t{57344.0},
-                type_convert<half_t>(f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::Max())),
+    ASSERT_NEAR(max_bf8_t_half,
+                type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
                 abs_tol);
-    // convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
-    ASSERT_NEAR(type_convert<bf8_t>(0x80),
-                f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+    // convert QuietNaN fp16 to bf8_fnuz_t and check if it is QuietNaN
+    ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
+                f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
                 abs_tol);
     // positive norm fp16 value to bf8 and back, check if holds
     half_t pos_half = half_t{0.0000762939};
-    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_t>(pos_half)), abs_tol);
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(pos_half)), abs_tol);
     // negative norm fp16 value to bf8 and back, check if holds
     half_t neg_half = half_t{-0.0000610351};
-    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_t>(neg_half)), abs_tol);
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(neg_half)), abs_tol);
     // positive subnorm fp16 value to bf8 and back, check if holds
     pos_half = half_t{0.0000305175};
-    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_t>(pos_half)), abs_tol);
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(pos_half)), abs_tol);
     // negative subnorm fp16 value to bf8 and back, check if holds
     neg_half = half_t{-0.0000152587};
-    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_t>(neg_half)), abs_tol);
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(neg_half)), abs_tol);
 }
 
-TEST(BF8, ConvertFP16Stochastic)
+TEST(BF8FNUZ, ConvertFP16Stochastic)
 {
     // fix the tolerance value
     float abs_tol = 1e-3;
     // convert 0 fp16 to bf8 and back, check if holds
-    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{0.0})), abs_tol);
+    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(half_t{0.0})), abs_tol);
     // convert minimal fp16 to bf8 and back, check if holds
     ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
-                type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Min())),
+                type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
                 abs_tol);
-    // convert maximal bf8_t to fp16 and check if equal to 57344.0
+
+    const auto max_bf8_t_half = type_convert<half_t>(ck::NumericLimits<bf8_fnuz_t>::Max());
+    // convert maximal bf8_fnuz_t to fp16 and check if equal to 57344.0
     ASSERT_NEAR(
-        half_t{57344.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{57344.0})), abs_tol);
+        max_bf8_t_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(max_bf8_t_half)), abs_tol);
     // convert maximal fp16 to bf8 and back, check if clipped to 57344.0
-    ASSERT_NEAR(half_t{57344.0},
-                type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Max())),
+    ASSERT_NEAR(max_bf8_t_half,
+                type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
                 abs_tol);
-    // convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
-    ASSERT_NEAR(type_convert<bf8_t>(0x80),
-                f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+    // convert QuietNaN fp16 to bf8_fnuz_t and check if it is QuietNaN
+    ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
+                f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
                 abs_tol);
     // positive norm fp16 value to bf8 and back, check if holds
     half_t pos_half = half_t{0.0000762939};
-    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(pos_half)), abs_tol);
     // negative norm fp16 value to bf8 and back, check if holds
     half_t neg_half = half_t{-0.0000610351};
-    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(neg_half)), abs_tol);
     // positive subnorm fp16 value to bf8 and back, check if holds
     pos_half = half_t{0.0000305175};
-    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(pos_half)), abs_tol);
     // negative subnorm fp16 value to bf8 and back, check if holds
     neg_half = half_t{-0.0000152587};
-    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(neg_half)), abs_tol);
 }

test/data_type/test_bf8_ocp.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type_convert.hpp"
+
+using ck::bf8_ocp_t;
+using ck::f8_convert_rne;
+using ck::f8_convert_sr;
+using ck::half_t;
+using ck::type_convert;
+
+TEST(BF8OCP, NumericLimits)
+{ // constants given for OCP FP8
+    EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Min(),
+              type_convert<bf8_ocp_t>(0x04)); // 0b00000100 = 2^-14
+    EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
+              type_convert<bf8_ocp_t>(0x7B)); // 0b01111011 = 57344
+    EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Lowest(),
+              type_convert<bf8_ocp_t>(0xFB)); // 0b11111011 = -57344
+    EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::QuietNaN().data,
+              type_convert<bf8_ocp_t>(0x7D).data); // 0b01111101
+    EXPECT_FALSE(ck::NumericLimits<bf8_ocp_t>::QuietNaN() ==
+                 ck::NumericLimits<bf8_ocp_t>::QuietNaN());
+    EXPECT_TRUE(ck::fp8_is_inf(type_convert<bf8_ocp_t>(0xFC)) &&
+                ck::fp8_is_inf(type_convert<bf8_ocp_t>(0x7C)));
+}
+
+TEST(BF8OCP, ConvertFP32Nearest)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+
+    // convert 0 float to bfp8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_ocp_t>(0.0f)), 0.0f);
+
+    // convert minimal float to bf8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+
+    const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_ocp_t>::Max());
+
+    // convert maximal bf8_ocp_t to float and check if equal to bf8 max
+    ASSERT_NEAR(
+        max_bf8_t_float, type_convert<float>(f8_convert_rne<bf8_ocp_t>(max_bf8_t_float)), 0.0f);
+
+    // convert maximal float to bf8 and back, check if clipped to bf8 max (saturation to finite)
+    ASSERT_NEAR(max_bf8_t_float,
+                type_convert<float>(f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::max())),
+                0.0f);
+
+    // convert float infinity to bf8_ocp_t and check if it is max value (saturation to finite)
+    ASSERT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
+              f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::infinity()));
+
+    // positive normal float value to bf8 and back, check if holds
+    float pos_float = 0.0000762939f; // 10*2^-17
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_ocp_t>(pos_float)), abs_tol);
+
+    // negative smallest normal bf8 value to bf8 and back, check if holds
+    constexpr auto neg_min_bf8 = -0.00006103515625f; //-2^-14
+    ASSERT_NEAR(neg_min_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(neg_min_bf8)), 0.0f);
+
+    // positive subnorm float value to bf8 and back, check if holds
+    constexpr auto pos_subnorm_bf8 = 0.000030517578125f; // 2^-15
+    ASSERT_NEAR(
+        pos_subnorm_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(pos_subnorm_bf8)), 0.0f);
+
+    // min subnorm bf8 value to bf8 and back, check if holds
+    constexpr auto min_subnorm_bf8 = -0.0000152587890625f; //-2^-16
+    ASSERT_NEAR(
+        min_subnorm_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(min_subnorm_bf8)), 0.0f);
+
+    // smaller than min subnorm bf8 value to bf8 must be zero
+    constexpr auto less_than_min_subnorm = 0.00000762939453125f; // 2^-17
+    ASSERT_EQ(0.0f, type_convert<float>(f8_convert_rne<bf8_ocp_t>(less_than_min_subnorm)));
+
+    // convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
+    const auto bf8_nan = f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
+    ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
+}
+
+TEST(BF8OCP, ConvertFP32Stochastic)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+
+    // convert 0 float to bfp8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_ocp_t>(0.0f)), 0.0f);
+
+    // convert minimal float to bf8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+
+    const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_ocp_t>::Max());
+
+    // convert maximal bf8_ocp_t to float and check if equal to bf8 max
+    ASSERT_NEAR(
+        max_bf8_t_float, type_convert<float>(f8_convert_sr<bf8_ocp_t>(max_bf8_t_float)), 0.0f);
+
+    // convert maximal float to bf8 and back, check if clipped to bf8 max (saturation to finite)
+    ASSERT_NEAR(max_bf8_t_float,
+                type_convert<float>(f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::max())),
+                0.0f);
+
+    // convert float infinity to bf8_ocp_t and check if it is max value (saturation to finite)
+    ASSERT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
+              f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::infinity()));
+
+    // positive normal float value to bf8 and back, check if holds
+    float pos_float = 0.0000762939f; // 10*2^-17
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_ocp_t>(pos_float)), abs_tol);
+
+    // negative smallest normal bf8 value to bf8 and back, check if holds
+    constexpr auto neg_min_bf8 = -0.00006103515625f; //-2^-14
+    ASSERT_NEAR(neg_min_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(neg_min_bf8)), 0.0f);
+
+    // positive subnorm float value to bf8 and back, check if holds
+    constexpr auto pos_subnorm_bf8 = 0.000030517578125f; // 2^-15
+    ASSERT_NEAR(
+        pos_subnorm_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(pos_subnorm_bf8)), 0.0f);
+
+    // min subnorm bf8 value to bf8 and back, check if holds
+    constexpr auto min_subnorm_bf8 = -0.0000152587890625f; //-2^-16
+    ASSERT_NEAR(
+        min_subnorm_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(min_subnorm_bf8)), 0.0f);
+
+    // smaller than min subnorm bf8 value to bf8  alternates between 0 and 2^-16
+    constexpr auto less_than_min_subnorm = 0.00000762939453125f; // 2^-17
+    ASSERT_NEAR(0.0f,
+                type_convert<float>(f8_convert_sr<bf8_ocp_t>(less_than_min_subnorm)),
+                0.0000152587890625f);
+
+    // convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
+    const auto bf8_nan = f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
+    ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
+}
+
+TEST(BF8OCP, ConvertFP16Nearest)
+{
+    // fix the tolerance value
+    constexpr half_t half_t_tol  = 1e-3;
+    constexpr half_t half_t_zero = 0.0;
+
+    // convert 0 half_t to bfp8 and back, check if holds
+    ASSERT_NEAR(
+        half_t_zero, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(half_t_zero)), half_t_zero);
+
+    // convert minimal half_t to bf8 and back, check if holds
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::Min())),
+                half_t_tol);
+
+    const auto max_bf8_t_half_t = type_convert<half_t>(ck::NumericLimits<bf8_ocp_t>::Max());
+
+    // convert maximal bf8_ocp_t to half_t and check if equal to bf8 max
+    ASSERT_NEAR(max_bf8_t_half_t,
+                type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(max_bf8_t_half_t)),
+                half_t_zero);
+
+    // convert maximal half_t to bf8 and back, check if clipped to bf8 max (saturation to finite)
+    ASSERT_NEAR(max_bf8_t_half_t,
+                type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::Max())),
+                half_t_zero);
+
+    // convert half_t infinity to bf8_ocp_t and check if it is max value (saturation to finite)
+    ASSERT_EQ(
+        ck::NumericLimits<bf8_ocp_t>::Max(),
+        f8_convert_rne<bf8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
+
+    // positive normal bf8 value to bf8 and back, check if holds
+    constexpr half_t pos_norm_bf8{0.0000762939f}; // 10*2^-17
+    ASSERT_NEAR(
+        pos_norm_bf8, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(pos_norm_bf8)), half_t_tol);
+
+    // negative smallest normal bf8 value to bf8 and back, check if holds
+    constexpr half_t neg_min_bf8{-0.00006103515625f}; //-2^-14
+    ASSERT_NEAR(
+        neg_min_bf8, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(neg_min_bf8)), half_t_zero);
+
+    // positive subnorm bf8 value to bf8 and back, check if holds
+    constexpr half_t pos_subnorm_bf8{0.000030517578125f}; // 2^-15
+    ASSERT_NEAR(pos_subnorm_bf8,
+                type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(pos_subnorm_bf8)),
+                half_t_zero);
+
+    // min subnorm bf8 value to bf8 and back, check if holds
+    constexpr half_t min_subnorm_bf8{-0.0000152587890625f}; //-2^-16
+    ASSERT_NEAR(min_subnorm_bf8,
+                type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(min_subnorm_bf8)),
+                half_t_zero);
+
+    // smaller than min subnorm bf8 value to bf8 must be zero
+    constexpr half_t less_than_min_subnorm{0.00000762939453125f}; // 2^-17
+    ASSERT_EQ(half_t_zero, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(less_than_min_subnorm)));
+
+    // convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
+    const auto bf8_nan = f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
+    ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
+}
+
+TEST(BF8OCP, ConvertFP16Stochastic)
+{
+    // fix the tolerance value
+    constexpr half_t half_t_tol    = 1e-3;
+    constexpr half_t half_t_zero   = 0.0;
+    constexpr auto min_subnorm_bf8 = 0.0000152587890625f; // 2^-16
+
+    // convert 0 half_t to bfp8 and back, check if holds
+    ASSERT_NEAR(
+        half_t_zero, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(half_t_zero)), half_t_zero);
+
+    // convert minimal half_t (6.103515625e-05) to fp8 and back
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::Min())),
+                half_t_zero);
+
+    const auto max_bf8_t_half_t = type_convert<half_t>(ck::NumericLimits<bf8_ocp_t>::Max());
+
+    // convert maximal bf8_ocp_t to half_t and check if equal to bf8 max
+    ASSERT_NEAR(max_bf8_t_half_t,
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(max_bf8_t_half_t)),
+                half_t_zero);
+
+    // convert maximal half_t to bf8 and back, check if clipped to bf8 max (saturation to finite)
+    ASSERT_NEAR(max_bf8_t_half_t,
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::Max())),
+                half_t_zero);
+
+    // convert half_t infinity to bf8_ocp_t and check if it is max value (saturation to finite)
+    ASSERT_EQ(
+        ck::NumericLimits<bf8_ocp_t>::Max(),
+        f8_convert_sr<bf8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
+
+    // positive normal bf8 value to bf8 and back, check if holds
+    constexpr half_t pos_norm_bf8{0.0000762939f}; // 10*2^-17
+    ASSERT_NEAR(
+        pos_norm_bf8, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(pos_norm_bf8)), half_t_tol);
+
+    // negative smallest normal bf8 value to bf8 and back, check if holds
+    constexpr half_t neg_min_bf8{-0.00006103515625f}; //-2^-14
+    ASSERT_NEAR(
+        neg_min_bf8, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(neg_min_bf8)), half_t_zero);
+
+    // positive subnorm bf8 value to bf8 and back, check if holds
+    constexpr half_t pos_subnorm_bf8{0.000030517578125f}; // 2^-15
+    ASSERT_NEAR(pos_subnorm_bf8,
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(pos_subnorm_bf8)),
+                half_t_zero);
+
+    // min subnorm bf8 value to bf8 and back, check if holds
+    ASSERT_NEAR(half_t{-min_subnorm_bf8},
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(half_t{-min_subnorm_bf8})),
+                half_t_zero);
+
+    // smaller than min subnorm bf8 value to bf8  alternates between 0 and 2^-16
+    constexpr half_t less_than_min_subnorm{0.00000762939453125f}; // 2^-17
+    ASSERT_NEAR(half_t_zero,
+                type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(less_than_min_subnorm)),
+                half_t{min_subnorm_bf8});
+
+    // convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
+    const auto bf8_nan = f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
+    ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
+}