/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/utility/mxf8_utils.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/utility/mxf8_utils.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/utility/mxf8_utils.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #include "ck/utility/numeric_limits.hpp"
 #include "ck/utility/mxfp_utils.hpp"
  
 #if defined(__gfx950__) && __HIP_DEVICE_COMPILE__
 #define CK_MX_FP8_CVT_FAST_PATH 1
 #else
 #define CK_MX_FP8_CVT_FAST_PATH 0
 #endif
  
 namespace ck {
  
 namespace fp8_impl {
 #if CK_MX_FP8_CVT_FAST_PATH
 template <ck_fp8_interpretation_t interpret>
 static __device__ float cast_to_f32_from_f8_scaled(float scale, 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_OCP ||
                       interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
                   "Only OCP interpretations are supported");
  
     if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
     {
         return __builtin_amdgcn_cvt_scalef32_f32_fp8(val.i32val, scale, 0);
     }
     else
     {
         return __builtin_amdgcn_cvt_scalef32_f32_bf8(val.i32val, scale, 0);
     }
 }
  
 template <ck_fp8_interpretation_t interpret>
 static __device__ float2_t cast_to_f32_from_f8_scaled(float scale, fp8x2_storage_t v)
 {
     const auto i16val = bit_cast<uint16_t>(v);
  
     static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
                       interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
                   "Only OCP interpretations are supported");
  
     if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
     {
         return __builtin_amdgcn_cvt_scalef32_pk_f32_fp8(i16val, scale, 0);
     }
     else
     {
         return __builtin_amdgcn_cvt_scalef32_pk_f32_bf8(i16val, scale, 0);
     }
 }
  
 template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
 static __device__ fp8_storage_t cast_to_f8_from_f32_scaled(float v,
                                                            unsigned int rng = 0,
                                                            float scale      = 1.0f)
 {
     fp8_storage_t i8data;
     union
     {
         float fval;
         unsigned int i32val;
     } val;
  
     union
     {
         uint32_t ival;
         vector_type<int16_t, 2>::type v2i16;
         fp8_storage_t v4i8[4];
     } ret{};
  
     // unsigned int ival = 0;
     val.fval = v;
  
     if constexpr(stochastic_rounding)
     {
         ret.ival =
             (interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
                 ? __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, val.fval, rng, scale, 0)
                 : __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, val.fval, rng, scale, 0);
  
         i8data = ret.v4i8[0];
     }
     else
     {
         // RNE CVT
         // llvm.amdgcn.cvt.scalef32.pk.fp8.f32
         // v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
         if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
         {
             // If fval / scale > max fp8, returns Nan
             ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
                                                                  val.fval,
                                                                  val.fval,
                                                                  scale,
                                                                  /*dst_lo_hi_sel*/ false);
         }
         else
         {
             // If fval / scale > max bf8, returns Inf
             ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
                                                                  val.fval,
                                                                  val.fval,
                                                                  scale,
                                                                  /*dst_lo_hi_sel*/ false);
         }
  
         i8data = ret.v4i8[0];
     }
     return i8data;
 }
  
 template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
 static __device__ fp8x2_storage_t cast_to_f8_from_f32_scaled(float2_t v,
                                                              unsigned int rng = 0,
                                                              float scale      = 1.0f)
 {
  
     union
     {
         uint32_t ival;
         vector_type<int16_t, 2>::type v2i16;
         StaticallyIndexedArray<fp8x2_storage_t, 2> v2f8x2;
     } ret{};
  
     if constexpr(stochastic_rounding)
     {
         fp8x2_storage_t f8x2;
         if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
         {
             ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[0], rng, scale, 0);
             f8x2[0]  = ret.v2f8x2(Number<0>{})[0];
             ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[1], rng, scale, 0);
             f8x2[1]  = ret.v2f8x2(Number<0>{})[0];
         }
         else
         {
             ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[0], rng, scale, 0);
             f8x2[0]  = ret.v2f8x2(Number<0>{})[0];
             ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[1], rng, scale, 0);
             f8x2[1]  = ret.v2f8x2(Number<0>{})[0];
         }
         return f8x2;
     }
     else
     {
         // RNE CVT
         // llvm.amdgcn.cvt.scalef32.pk.fp8.f32
         // v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
         if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
         {
             // If fval / scale > max fp8, returns Nan
             ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
                                                                  v[0],
                                                                  v[1],
                                                                  scale,
                                                                  /*dst_lo_hi_sel*/ false);
         }
         else
         {
             // If fval / scale > max bf8, returns Inf
             ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
                                                                  v[0],
                                                                  v[1],
                                                                  scale,
                                                                  /*dst_lo_hi_sel*/ false);
         }
  
         return ret.v2f8x2(Number<0>{});
     }
 }
  
 #endif // CK_MX_FP8_CVT_FAST_PATH
  
 #if CK_MX_FP8_CVT_FAST_PATH
 template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
 __host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
 {
     __is_interpret_supported(interp);
     uint32_t rng = 0;
     if constexpr(stochastic_rounding)
     {
         // use HW clock for stochastic input multiply by incremented thread id
         rng = __builtin_amdgcn_prng_b32(__builtin_amdgcn_s_memrealtime() *
                                         (get_thread_global_1d_id() + 1));
     }
     return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
 }
  
 template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
 __host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
                                                                           float scale)
 {
     __is_interpret_supported(interp);
     uint32_t rng = 0;
     if constexpr(stochastic_rounding)
     {
         // use HW clock for stochastic input multiply by incremented thread id
         rng = __builtin_amdgcn_prng_b32(__builtin_amdgcn_s_memrealtime() *
                                         (get_thread_global_1d_id() + 1));
     }
     return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
 }
  
 #else
  
 template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
 __host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
 {
  
     static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
                       interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
                   "Only OCP interpretations are supported");
  
     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_OCP)
     {
         return cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f / scale, rng);
     }
     else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
     {
         return cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f / scale, rng);
     }
     else
     {
         __hip_assert(false && "FP8 type is not supported by current target device");
         return 0;
     }
 }
  
 template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
 __host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
                                                                           float scale)
 {
  
     static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
                       interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
                   "Only OCP interpretations are supported");
  
     uint32_t rng = 0;
     if constexpr(stochastic_rounding)
     {
         constexpr int seed = 1254739;
         rng                = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f[0]);
     }
  
     if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
     {
         return {cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[0] / scale, rng),
                 cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[1] / scale, rng)};
     }
     else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
     {
         return {cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[0] / scale, rng),
                 cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[1] / scale, rng)};
     }
     else
     {
         __hip_assert(false && "FP8 type is not supported by current target device");
         return 0;
     }
 }
  
 #endif // CK_MX_FP8_CVT_FAST_PATH
  
 } // namespace fp8_impl
  
 // Declare a template function for fp8 conversion using SR
 template <typename Y, typename X>
 __host__ __device__ constexpr Y mxf8_convert_sr(X x, float scale);
  
 // Declare a template function for fp8 conversion using RNE
 template <typename Y, typename X>
 __host__ __device__ constexpr Y mxf8_convert_rne(X x, float scale);
  
 // convert fp32 to fp8 with rounding to nearest even
 template <>
 inline __host__ __device__ f8_ocp_t mxf8_convert_rne<f8_ocp_t, float>(float x, float scale)
 {
     return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
 }
  
 // convert fp32 to bf8 with rounding to nearest even
 template <>
 inline __host__ __device__ bf8_ocp_t mxf8_convert_rne<bf8_ocp_t, float>(float x, float scale)
 {
     return bf8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
 }
  
 // convert fp32x2 to fp8x2 with rounding to nearest even
 template <>
 inline __host__ __device__ f8x2_ocp_t mxf8_convert_rne<f8x2_ocp_t, float2_t>(float2_t x,
                                                                              float scale)
 {
     return f8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
 }
  
 // convert fp32x2 to bf8x2 with rounding to nearest even
 template <>
 inline __host__ __device__ bf8x2_ocp_t mxf8_convert_rne<bf8x2_ocp_t, float2_t>(float2_t x,
                                                                                float scale)
 {
     return bf8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
 }
  
 // convert fp32x16 to fp8x16 with rounding to nearest even
 template <>
 inline __host__ __device__ f8x16_ocp_t mxf8_convert_rne<f8x16_ocp_t, float16_t>(float16_t x,
                                                                                 float scale)
 {
     union
     {
         float16_t float_1x16;
         float2_t float_2x8[8];
     } in{x};
  
     union
     {
         f8x16_ocp_t fp8_1x16;
         f8x2_ocp_t fp8_2x8[8];
     } out{};
  
     ck::static_for<0, 8, 1>{}(
         [&](auto i) { out.fp8_2x8[i] = mxf8_convert_rne<f8x2_ocp_t>(in.float_2x8[i], scale); });
  
     return out.fp8_1x16;
 }
  
 // convert fp32x16 to bf8x16 with rounding to nearest even
 template <>
 inline __host__ __device__ bf8x16_ocp_t mxf8_convert_rne<bf8x16_ocp_t, float16_t>(float16_t x,
                                                                                   float scale)
 {
     union
     {
         float16_t float_1x16;
         float2_t float_2x8[8];
     } in{x};
  
     union
     {
         bf8x16_ocp_t bf8_1x16;
         bf8x2_ocp_t bf8_2x8[8];
     } out{};
  
     ck::static_for<0, 8, 1>{}(
         [&](auto i) { out.bf8_2x8[i] = mxf8_convert_rne<bf8x2_ocp_t>(in.float_2x8[i], scale); });
  
     return out.bf8_1x16;
 }
  
 // convert fp32x32 to fp8x32 with rounding to nearest even
 template <>
 inline __host__ __device__ f8x32_ocp_t mxf8_convert_rne<f8x32_ocp_t, float32_t>(float32_t x,
                                                                                 float scale)
 {
     union
     {
         float32_t float_1x32;
         float16_t float_16x2[2];
     } in{x};
  
     union
     {
         f8x32_ocp_t fp8_1x32;
         f8x16_ocp_t fp8_16x2[2];
     } out{};
  
     ck::static_for<0, 2, 1>{}(
         [&](auto i) { out.fp8_16x2[i] = mxf8_convert_rne<f8x16_ocp_t>(in.float_16x2[i], scale); });
  
     return out.fp8_1x32;
 }
  
 // convert fp32x32 to bf8x32 with rounding to nearest even
 template <>
 inline __host__ __device__ bf8x32_ocp_t mxf8_convert_rne<bf8x32_ocp_t, float32_t>(float32_t x,
                                                                                   float scale)
 {
     union
     {
         float32_t float_1x32;
         float16_t float_16x2[2];
     } in{x};
  
     union
     {
         bf8x32_ocp_t bf8_1x32;
         bf8x16_ocp_t bf8_16x2[2];
     } out{};
  
     ck::static_for<0, 2, 1>{}(
         [&](auto i) { out.bf8_16x2[i] = mxf8_convert_rne<bf8x16_ocp_t>(in.float_16x2[i], scale); });
  
     return out.bf8_1x32;
 }
  
 // convert fp32 to fp8 with stochastic rounding
 template <>
 inline __host__ __device__ f8_ocp_t mxf8_convert_sr<f8_ocp_t, float>(float x, float scale)
 {
     return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
 }
  
 // convert fp32 to bf8 with stochastic rounding
 template <>
 inline __host__ __device__ bf8_ocp_t mxf8_convert_sr<bf8_ocp_t, float>(float x, float scale)
 {
     return bf8_ocp_t{
         fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
 }
  
 // convert fp32x2 to fp8x2 with stochastic rounding
 template <>
 inline __host__ __device__ f8x2_ocp_t mxf8_convert_sr<f8x2_ocp_t, float2_t>(float2_t x, float scale)
 {
     return f8x2_ocp_t{
         fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
 }
  
 // convert fp32x2 to bf8x2 with stochastic rounding
 template <>
 inline __host__ __device__ bf8x2_ocp_t mxf8_convert_sr<bf8x2_ocp_t, float2_t>(float2_t x,
                                                                               float scale)
 {
     return bf8x2_ocp_t{
         fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
 }
  
 // convert fp32x16 to fp8x16 with stochastic rounding
 template <>
 inline __host__ __device__ f8x16_ocp_t mxf8_convert_sr<f8x16_ocp_t, float16_t>(float16_t x,
                                                                                float scale)
 {
     union
     {
         float16_t float_1x16;
         float2_t float_2x8[8];
     } in{x};
  
     union
     {
         f8x16_ocp_t fp8_1x16;
         f8x2_ocp_t fp8_2x8[8];
     } out{};
  
     ck::static_for<0, 8, 1>{}(
         [&](auto i) { out.fp8_2x8[i] = mxf8_convert_sr<f8x2_ocp_t>(in.float_2x8[i], scale); });
  
     return out.fp8_1x16;
 }
  
 // convert fp32x16 to bf8x16 with stochastic rounding
 template <>
 inline __host__ __device__ bf8x16_ocp_t mxf8_convert_sr<bf8x16_ocp_t, float16_t>(float16_t x,
                                                                                  float scale)
 {
     union
     {
         float16_t float_1x16;
         float2_t float_2x8[8];
     } in{x};
  
     union
     {
         bf8x16_ocp_t bf8_1x16;
         bf8x2_ocp_t bf8_2x8[8];
     } out{};
  
     ck::static_for<0, 8, 1>{}(
         [&](auto i) { out.bf8_2x8[i] = mxf8_convert_sr<bf8x2_ocp_t>(in.float_2x8[i], scale); });
  
     return out.bf8_1x16;
 }
  
 // convert fp32x32 to fp8x32 with stochastic rounding
 template <>
 inline __host__ __device__ f8x32_ocp_t mxf8_convert_sr<f8x32_ocp_t, float32_t>(float32_t x,
                                                                                float scale)
 {
     union
     {
         float32_t float_1x32;
         float16_t float_16x2[2];
     } in{x};
  
     union
     {
         f8x32_ocp_t fp8_1x32;
         f8x16_ocp_t fp8_16x2[2];
     } out{};
  
     ck::static_for<0, 2, 1>{}(
         [&](auto i) { out.fp8_16x2[i] = mxf8_convert_sr<f8x16_ocp_t>(in.float_16x2[i], scale); });
  
     return out.fp8_1x32;
 }
  
 // convert fp32x32 to bf8x32 with stochastic rounding
 template <>
 inline __host__ __device__ bf8x32_ocp_t mxf8_convert_sr<bf8x32_ocp_t, float32_t>(float32_t x,
                                                                                  float scale)
 {
     union
     {
         float32_t float_1x32;
         float16_t float_16x2[2];
     } in{x};
  
     union
     {
         bf8x32_ocp_t bf8_1x32;
         bf8x16_ocp_t bf8_16x2[2];
     } out{};
  
     ck::static_for<0, 2, 1>{}(
         [&](auto i) { out.bf8_16x2[i] = mxf8_convert_sr<bf8x16_ocp_t>(in.float_16x2[i], scale); });
  
     return out.bf8_1x32;
 }
  
 } // namespace ck