Merge branch 'develop' into muozturk/complex_contraction_bilinear_test

client_example/18_groupnorm/CMakeLists.txt

 add_executable(client_groupnorm_swish groupnorm_swish.cpp)
-target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_other_operations)

client_example/19_pool/CMakeLists.txt

 add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
-target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
-target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
-target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
-target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_other_operations)

client_example/20_splitk_gemm/CMakeLists.txt

 if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
   add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
-  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
+  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
 endif()

client_example/21_grouped_gemm_bias/CMakeLists.txt

 add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_gemm_operations)

client_example/22_grouped_gemm/CMakeLists.txt

 add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_gemm_operations)
 
 add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_gemm_operations)
 
 add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)

client_example/22_im2col_col2im/CMakeLists.txt

 add_executable(client_image_to_column image_to_column.cpp)
-target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_column_to_image column_to_image.cpp)
-target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_other_operations)

client_example/23_elementwise_transpose/CMakeLists.txt

 add_executable(client_elementwise_transpose3d elementwise_transpose_3d.cpp)
-target_link_libraries(client_elementwise_transpose3d PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_elementwise_transpose3d PRIVATE composable_kernel::device_other_operations)

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/CMakeLists.txt

 add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_conv_operations)

client_example/24_grouped_convnd_fwd_scaleadd_ab/CMakeLists.txt

 add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp32 grouped_conv_fwd_scaleadd_ab_fp32.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp32 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp16 grouped_conv_fwd_scaleadd_ab_fp16.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp16 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_ab_bf16 grouped_conv_fwd_scaleadd_ab_bf16.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_bf16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_bf16 PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_convnd_fwd_scaleadd_ab_int8 grouped_conv_fwd_scaleadd_ab_int8.cpp)
-target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_conv_operations)

client_example/CMakeLists.txt

@@ -48,7 +48,7 @@ else()
     endif()
 endif()
 
-find_package(composable_kernel COMPONENTS device_operations)
+find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_contraction_operations device_reduction_operations)
 find_package(hip REQUIRED PATHS /opt/rocm)
 message(STATUS "Build with HIP ${hip_VERSION}")
 

include/ck/ck.hpp

@@ -134,6 +134,9 @@
 // inner product using V_DOT with DPP8 modifiers
 #define CK_USE_AMD_V_DOT_DPP8_INLINE_ASM 1
 
+// set stochastic rounding as default for f8 conversions
+#define CK_USE_SR_F8_CONVERSION 1
+
 // block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
 #define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
 

include/ck/stream_config.hpp

@@ -11,6 +11,6 @@ struct StreamConfig
     hipStream_t stream_id_ = nullptr;
     bool time_kernel_      = false;
     int log_level_         = 0;
-    int cold_niters_       = 50;
-    int nrepeat_           = 200;
+    int cold_niters_       = 1;
+    int nrepeat_           = 10;
 };

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

@@ -281,6 +281,24 @@ struct ConvertF8SR
     }
 };
 
+struct ConvertF8RNE
+{
+    // convert to fp8 using rounding to nearest even
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        // check Y datatype
+        static_assert(is_same<Y, f8_t>::value || is_same<Y, bf8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        // check X datatype
+        static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = f8_convert_rne<Y>(x);
+    }
+};
+
 struct Scale
 {
     __host__ __device__ Scale(float scale) : scale_(scale) {}

include/ck/utility/type_convert.hpp

@@ -95,9 +95,113 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
     return type_convert<bhalf_t>(x_fp32);
 }
 
-// convert fp32 to fp8
+// Declare a template function for fp8 conversion using SR
+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_t type_convert<f8_t, float>(float x)
+inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
+{
+    constexpr int seed = 42;
+    uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    float max_fp8 = 240.0f;
+    x             = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
+    union
+    {
+        float fval;
+        uint32_t i32val;
+        uint8_t i8val[4]; // not endian independent
+    } val;
+    val.fval      = x;
+    uint32_t ival = 0;
+    ival          = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
+    val.i32val    = ival;
+    return val.i8val[0]; // little endian
+#else
+    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, f8_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)
+{
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    // convert to float and use native converion
+    return f8_convert_sr<f8_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               = 42;
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::
+        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
+#endif
+}
+
+// convert fp32 to bf8 with stochastic rounding
+template <>
+inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
+{
+    constexpr int seed = 42;
+    uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    union
+    {
+        float fval;
+        uint32_t i32val;
+        uint8_t i8val[4]; // not endian independent
+    } val;
+    val.fval      = x;
+    uint32_t ival = 0;
+    ival          = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
+    val.i32val    = ival;
+    return val.i8val[0]; // little endian
+#else
+    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);
+#endif
+}
+
+// convert fp16 to bf8 with stochastic rounding
+template <>
+inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
+{
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    // convert to float and use native converion
+    return f8_convert_sr<f8_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               = 42;
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::
+        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
+#endif
+}
+
+// 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_t f8_convert_rne<f8_t, float>(float x)
 {
 #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
     float max_fp8 = 240.0f;
@@ -124,6 +228,80 @@ inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
 #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)
+{
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    // convert to float and use native converion
+    return f8_convert_rne<f8_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);
+#endif
+}
+
+// convert fp32 to bf8 with rounding to nearest even
+template <>
+inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
+{
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    union
+    {
+        float fval;
+        uint32_t i32val;
+        uint8_t i8val[4]; // not endian independent
+    } val;
+    val.fval      = x;
+    uint32_t ival = 0;
+    ival       = __builtin_amdgcn_cvt_pk_bf8_f32(val.fval, val.fval, ival, false); // false -> WORD0
+    val.i32val = ival;
+    return val.i8val[0];
+#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<float, bf8_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)
+{
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+    // convert to float and use native converion
+    return f8_convert_rne<bf8_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);
+#endif
+}
+
+// convert fp32 to fp8
+template <>
+inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
+{
+#if defined CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_t>(x);
+#else
+    return f8_convert_rne<f8_t>(x);
+#endif
+}
+
 // convert fp8 to fp32
 template <>
 inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
@@ -174,17 +352,10 @@ inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
 template <>
 inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
 {
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    // convert to float and use native converion
-    return type_convert<f8_t>(type_convert<float>(x));
+#if defined CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_t>(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 f8_convert_nre<f8_t>(x);
 #endif
 }
 
@@ -205,26 +376,10 @@ inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
 template <>
 inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
 {
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    union
-    {
-        float fval;
-        uint32_t i32val;
-        uint8_t i8val[4]; // not endian independent
-    } val;
-    val.fval      = x;
-    uint32_t ival = 0;
-    ival       = __builtin_amdgcn_cvt_pk_bf8_f32(val.fval, val.fval, ival, false); // false -> WORD0
-    val.i32val = ival;
-    return val.i8val[0];
+#if defined CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_t>(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<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return f8_convert_rne<bf8_t>(x);
 #endif
 }
 
@@ -248,17 +403,10 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
 template <>
 inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
 {
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    // convert to float and use native converion
-    return type_convert<bf8_t>(type_convert<float>(x));
+#if defined CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_t>(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 f8_convert_rne<bf8_t>(x);
 #endif
 }
 
@@ -331,104 +479,4 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h
 
     return bf16_convert_rtn<bhalf_t>(x_fp32);
 }
-
-// Declare a template function for fp8 conversion using SR
-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_t f8_convert_sr<f8_t, float>(float x)
-{
-    constexpr int seed = 42;
-    uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    union
-    {
-        float fval;
-        uint32_t i32val;
-        uint8_t i8val[4]; // not endian independent
-    } val;
-    val.fval      = x;
-    uint32_t ival = 0;
-    ival          = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
-    val.i32val    = ival;
-    return val.i8val[0]; // little endian
-#else
-    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, f8_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)
-{
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    // convert to float and use native converion
-    return f8_convert_sr<f8_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               = 42;
-    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
-#endif
-}
-
-// convert fp32 to bf8 with stochastic rounding
-template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
-{
-    constexpr int seed = 42;
-    uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    union
-    {
-        float fval;
-        uint32_t i32val;
-        uint8_t i8val[4]; // not endian independent
-    } val;
-    val.fval      = x;
-    uint32_t ival = 0;
-    ival          = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
-    val.i32val    = ival;
-    return val.i8val[0]; // little endian
-#else
-    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);
-#endif
-}
-
-// convert fp16 to bf8 with stochastic rounding
-template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
-{
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
-    // convert to float and use native converion
-    return f8_convert_sr<f8_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               = 42;
-    // as thread id is not available on host, use 0 for prn generation
-    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
-#endif
-}
-
 } // namespace ck

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -58,7 +58,12 @@ endfunction(add_instance_library INSTANCE_NAME)
 
 
 file(GLOB dir_list LIST_DIRECTORIES true *)
-set(CK_DEVICE_INSTANCES)
+set(CK_DEVICE_OTHER_INSTANCES)
+set(CK_DEVICE_GEMM_INSTANCES)
+set(CK_DEVICE_CONV_INSTANCES)
+set(CK_DEVICE_MHA_INSTANCES)
+set(CK_DEVICE_CONTRACTION_INSTANCES)
+set(CK_DEVICE_REDUCTION_INSTANCES)
 FOREACH(subdir_path ${dir_list})
     set(target_dir)
     IF(IS_DIRECTORY "${subdir_path}")
@@ -122,7 +127,19 @@ FOREACH(subdir_path ${dir_list})
         if((add_inst EQUAL 1))
             get_filename_component(target_dir ${subdir_path} NAME)
             add_subdirectory(${target_dir})
-            list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            if("${cmake_instance}" MATCHES "gemm")
+                list(APPEND CK_DEVICE_GEMM_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            elseif("${cmake_instance}" MATCHES "conv")
+                 list(APPEND CK_DEVICE_CONV_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            elseif("${cmake_instance}" MATCHES "mha")
+                 list(APPEND CK_DEVICE_MHA_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            elseif("${cmake_instance}" MATCHES "contr")
+                 list(APPEND CK_DEVICE_CONTRACTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            elseif("${cmake_instance}" MATCHES "reduce")
+                 list(APPEND CK_DEVICE_REDUCTION_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            else()
+                 list(APPEND CK_DEVICE_OTHER_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            endif()
             message("add_instance_directory ${subdir_path}")
         else()
             message("skip_instance_directory ${subdir_path}")
@@ -130,18 +147,14 @@ FOREACH(subdir_path ${dir_list})
     ENDIF()
 ENDFOREACH()
 
-add_library(device_operations STATIC ${CK_DEVICE_INSTANCES})
-add_library(composablekernels::device_operations ALIAS device_operations)
 
 
-set(DEV_OPS_INC_DIRS
-    ${PROJECT_SOURCE_DIR}/include/ck/
-    ${PROJECT_SOURCE_DIR}/library/include/ck/
-)
-
-target_compile_features(device_operations PUBLIC)
-set_target_properties(device_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
-target_include_directories(device_operations PUBLIC
+if(CK_DEVICE_OTHER_INSTANCES)
+        add_library(device_other_operations STATIC ${CK_DEVICE_OTHER_INSTANCES})
+        add_library(composablekernels::device_other_operations ALIAS device_other_operations)
+        target_compile_features(device_other_operations PUBLIC)
+        set_target_properties(device_other_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_other_operations PUBLIC
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_description>
@@ -157,23 +170,115 @@ target_include_directories(device_operations PUBLIC
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance>
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/quantization>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/softmax>
+        )
+        rocm_install(TARGETS device_other_operations
+            EXPORT device_other_operationsTargets)
+        rocm_install(EXPORT device_other_operationsTargets
+            FILE composable_kerneldevice_other_operationsTargets.cmake
+            NAMESPACE composable_kernel::
+            DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+if(CK_DEVICE_GEMM_INSTANCES)
+        add_library(device_gemm_operations STATIC ${CK_DEVICE_GEMM_INSTANCES})
+        add_library(composablekernels::device_gemm_operations ALIAS device_gemm_operations)
+        target_compile_features(device_gemm_operations PUBLIC)
+        set_target_properties(device_gemm_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_gemm_operations PUBLIC
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+        )
+        rocm_install(TARGETS device_gemm_operations
+            EXPORT device_gemm_operationsTargets)
+        rocm_install(EXPORT device_gemm_operationsTargets
+            FILE composable_kerneldevice_gemm_operationsTargets.cmake
+            NAMESPACE composable_kernel::
+            DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+if(CK_DEVICE_CONV_INSTANCES)
+        add_library(device_conv_operations STATIC ${CK_DEVICE_CONV_INSTANCES})
+        add_library(composablekernels::device_conv_operations ALIAS device_conv_operations)
+        target_compile_features(device_conv_operations PUBLIC)
+        set_target_properties(device_conv_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_conv_operations PUBLIC
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd>
+        )
+        rocm_install(TARGETS device_conv_operations
+            EXPORT device_conv_operationsTargets)
+        rocm_install(EXPORT device_conv_operationsTargets
+            FILE composable_kerneldevice_conv_operationsTargets.cmake
+            NAMESPACE composable_kernel::
+            DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+if(CK_DEVICE_MHA_INSTANCES)
+        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)
+        set_target_properties(device_mha_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_mha_operations PUBLIC
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/mha>
+        )
+        rocm_install(TARGETS device_mha_operations
+            EXPORT device_mha_operationsTargets)
+        rocm_install(EXPORT device_mha_operationsTargets
+            FILE composable_kerneldevice_mha_operationsTargets.cmake
+            NAMESPACE composable_kernel::
+            DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+if(CK_DEVICE_CONTRACTION_INSTANCES)
+        add_library(device_contraction_operations STATIC ${CK_DEVICE_CONTRACTION_INSTANCES})
+        add_library(composablekernels::device_contraction_operations ALIAS device_contraction_operations)
+        target_compile_features(device_contraction_operations PUBLIC)
+        set_target_properties(device_contraction_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_contraction_operations PUBLIC
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/contraction>
+        )
+        rocm_install(TARGETS device_contraction_operations
+            EXPORT device_contraction_operationsTargets)
+        rocm_install(EXPORT device_contraction_operationsTargets
+            FILE composable_kerneldevice_contraction_operationsTargets.cmake
+            NAMESPACE composable_kernel::
+            DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+if(CK_DEVICE_REDUCTION_INSTANCES)
+        add_library(device_reduction_operations STATIC ${CK_DEVICE_REDUCTION_INSTANCES})
+        add_library(composablekernels::device_reduction_operations ALIAS device_reduction_operations)
+        target_compile_features(device_reduction_operations PUBLIC)
+        set_target_properties(device_reduction_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_include_directories(device_reduction_operations PUBLIC
             $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/reduce>
-)
-
-#once new arches are enabled make this an option on the main cmake file
-# and pass down here to be exported
-target_compile_options(device_operations PRIVATE
-    --offload-arch=gfx908
-    --offload-arch=gfx90a
-)
-
-# install(TARGETS device_operations LIBRARY DESTINATION lib)
-rocm_install(TARGETS device_operations
-    EXPORT device_operationsTargets)
-
-rocm_install(DIRECTORY ${DEV_OPS_INC_DIRS} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck)
-rocm_install(EXPORT device_operationsTargets
-    FILE composable_kerneldevice_operationsTargets.cmake
+        )
+        rocm_install(TARGETS device_reduction_operations
+            EXPORT device_reduction_operationsTargets)
+        rocm_install(EXPORT device_reduction_operationsTargets
+            FILE composable_kerneldevice_reduction_operationsTargets.cmake
             NAMESPACE composable_kernel::
             DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+        )
+endif()
+
+add_library(device_operations INTERFACE)
+target_link_libraries(device_operations INTERFACE
+    device_contraction_operations
+    device_conv_operations
+    device_gemm_operations
+    device_other_operations
+    device_reduction_operations
+    utility)
+
+set(DEV_OPS_INC_DIRS
+    ${PROJECT_SOURCE_DIR}/include/ck/
+    ${PROJECT_SOURCE_DIR}/library/include/ck/
 )
+rocm_install(DIRECTORY ${DEV_OPS_INC_DIRS} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck)
+

profiler/include/profiler/profile_gemm_impl.hpp

@@ -166,7 +166,7 @@ int profile_gemm_impl(int do_verification,
             std::string op_name = op_ptr->GetTypeString();
 
             float avg_time =
-                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, 10, 50});
 
             std::size_t flop = std::size_t(2) * M * N * K;
 

test/transpose/test_transpose_ut_cases.inc

@@ -14,7 +14,6 @@ TYPED_TEST(TestTranspose, Test1)
     this->Run();
 }
 
-
 TYPED_TEST(TestTranpose, Test2)
 {
     std::vector<int> Ms{127, 255, 312, 799, 1573};
@@ -27,4 +26,3 @@ TYPED_TEST(TestTranpose, Test2)
 
     this->Run();
 }
-