Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into lds_bypass_spilling

include/ck/ck.hpp

@@ -168,6 +168,9 @@
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
+// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
+#define CK_WORKAROUND_SWDEV_383542 1
+
 // flag to enable (1) or disable (0) the debugging output in some kernels
 #define DEBUG_LOG 0
 

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

@@ -111,6 +111,7 @@ __global__ void
     ignore = p_b_grid;
     ignore = p_b1_grid;
     ignore = p_c_grid;
+    ignore = p_d0s_grid;
     ignore = a_element_op;
     ignore = b_element_op;
     ignore = c0de_element_op;

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

@@ -586,6 +586,11 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
             return false;
         }
 
+        if(ck::get_device_name() != "gfx90a" && std::is_same<ADataType, double>::value)
+        {
+            return false;
+        }
+
         if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
                                         arg.b_grid_desc_n_k_,
                                         arg.ds_grid_desc_m_n_,

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -280,43 +281,42 @@ struct AddHardswish
     };
 };
 
-// C = A * B
 // E = FastGelu(C + D)
 struct AddFastGelu
 {
-    // Fast GeLU
-    // https://paperswithcode.com/method/gelu
-    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
-    __host__ __device__ static constexpr float GetFastGeLU(float x)
-    {
-        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
-        const float emu = exp(-u);
-        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
-        return x * cdf;
-    }
-
-    template <typename T>
-    static inline constexpr bool is_valid_param_type_v =
-        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
-        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>;
-
     template <typename E, typename C, typename D>
-    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& e, const float& c, const float& d) const
     {
-        static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
-                      is_valid_param_type_v<D>);
+        const float x = c + d;
+
+        FastGelu{}.template operator()<float, float>(e, x);
+    }
 
-        const float y = GetFastGeLU(type_convert<float>(c) + type_convert<float>(d));
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
+    {
+        const half_t x = c + d;
 
-        e = type_convert<E>(y);
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
     }
 
-    template <typename D>
-    __host__ __device__ constexpr void operator()(float& e, const float& c, const D& d) const
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
     {
-        static_assert(is_valid_param_type_v<D>);
+        const float x0_f = c + d;
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
 
-        e = GetFastGeLU(c + type_convert<float>(d));
+        e = type_convert<half_t>(x1_f);
     }
 };
 

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

@@ -16,7 +16,7 @@ namespace element_wise {
 // Need to ensure compiler will fail if there is no matching candidate, instead of compiler
 // siliently do implicit type conversion
 //
-// Method 1:
+// Example:
 //
 // struct ExampleElementwiseOp
 // {
@@ -30,19 +30,6 @@ namespace element_wise {
 //     {
 //     }
 // };
-//
-// Method 2:
-//
-// template <typename Y, typename X>
-// struct ExampleElementwiseOp;
-//
-// template <>
-// struct ExampleElementwiseOp<float, ck::bhalf_t>
-// {
-//     __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
-//     {
-//     }
-// };
 
 struct AddReluAdd
 {
@@ -208,41 +195,74 @@ struct AddMultiply
     }
 };
 
-// C = A * B
 // E = FastGelu(C + D0 + D1)
 struct AddAddFastGelu
 {
-    // Fast GeLU
-    // https://paperswithcode.com/method/gelu
-    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
-    __host__ __device__ static constexpr float GetFastGeLU(float x)
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
+                                                                              const float& c,
+                                                                              const float& d0,
+                                                                              const float& d1) const
     {
-        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
-        const float emu = exp(-u);
-        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
-        return x * cdf;
+        const float x = c + d0 + d1;
+
+        FastGelu{}.template operator()<float, float>(e, x);
     }
 
-    template <typename T>
-    static inline constexpr bool is_valid_param_type_v =
-        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
-        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
-#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
-        || std::is_same_v<T, ck::int4_t>
-#endif
-        ;
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
+    {
+        const half_t x = c + d0 + d1;
 
-    template <typename E, typename C, typename D0, typename D1>
-    __host__ __device__ constexpr void
-    operator()(E& e, const C& c, const D0& d0, const D1& d1) const
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
+        half_t& e, const float& c, const half_t& d0, const half_t& d1) const
     {
-        static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
-                      is_valid_param_type_v<D0> && is_valid_param_type_v<D1>);
+        const float x0_f = c + d0 + d1;
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<half_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
+        bhalf_t& e, const float& c, const bhalf_t& d0, const bhalf_t& d1) const
+    {
+        const float x0_f = c + type_convert<float>(d0) + type_convert<float>(d1);
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<bhalf_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t, int8_t>(
+        int8_t& e, const int32_t& c, const int8_t& d0, const int8_t& d1) const
+    {
+        const float x0_f =
+            type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1);
+
+        float x1_f = 0;
 
-        const float y =
-            GetFastGeLU(type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1));
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
 
-        e = type_convert<E>(y);
+        e = type_convert<int8_t>(x1_f);
     }
 };
 

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -11,6 +11,10 @@ namespace ck {
 namespace tensor_operation {
 namespace element_wise {
 
+#if CK_WORKAROUND_SWDEV_383542
+extern "C" __device__ float __ocml_native_recip_f32(float);
+#endif
+
 struct PassThrough
 {
     template <typename Y, typename X>
@@ -200,36 +204,83 @@ struct Relu
     }
 };
 
-// Y = FastGelu(X)
+// Fast GeLU
+// https://paperswithcode.com/method/gelu
+// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+// host code use higher accuracy "exp" and "div"
+// gpu code use lower accuracy "__expf" and "rcp" function
 struct FastGelu
 {
-    // Fast GeLU
-    // https://paperswithcode.com/method/gelu
-    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
-    __host__ __device__ static constexpr float GetFastGeLU(float x)
+    template <typename Y, typename X>
+    __host__ void operator()(Y& y, const X& x) const;
+
+    template <typename Y, typename X>
+    __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ void operator()<float, float>(float& y, const float& x) const
     {
         const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
         const float emu = exp(-u);
         const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
-        return x * cdf;
+
+        y = x * cdf;
     }
 
-    template <typename T>
-    static inline constexpr bool is_valid_param_type_v =
-        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
-        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
-#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
-        || std::is_same_v<T, ck::int4_t>
+    // device code, use lower precision "__expf" and "rcp"
+    template <>
+    __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = __expf(-u);
+
+#if !CK_WORKAROUND_SWDEV_383542
+        const float cdf = 0.5f + 0.5f * (2.f * __frcp_rn(1.f + emu) - 1.f);
+#else
+        const float cdf = 0.5f + 0.5f * (2.f * __ocml_native_recip_f32(1.f + emu) - 1.f);
 #endif
-        ;
 
-    template <typename Y, typename X>
-    __host__ __device__ void operator()(Y& y, const X& x) const
+        y = x * cdf;
+    }
+
+    template <>
+    __host__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
     {
-        static_assert(is_valid_param_type_v<Y> && is_valid_param_type_v<X>);
+        float y_f;
+
+        this->operator()<float, float>(y_f, type_convert<float>(x));
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, type_convert<float>(x));
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __host__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, x);
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __device__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, x);
 
-        const float tmp_y = GetFastGeLU(type_convert<float>(x));
-        y                 = type_convert<Y>(tmp_y);
+        y = type_convert<half_t>(y_f);
     }
 };
 

script/cmake-ck-dev.sh

@@ -10,8 +10,8 @@ cmake
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
 -D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD"         \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
--D BUILD_DEV=OFF                                                                                   \
--D GPU_TARGETS="gfx90a"                                                                    \
+-D BUILD_DEV=ON                                                                                   \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
 -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
 -D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
 ${MY_PROJECT_SOURCE}