Switched to using the hip_fp8 header instead of rccl_float8, resolving compatibility issues. (#1546)

* Revert "Revert "replacing rccl_float8 with hip_fp8 and address compatibility …"

This reverts commit 824b81c034.

* [UT] Modify max stack size to 496

* adding a check for OCP type and replacing ROCM_VERSION with HIP_VERSION

* addressing the ci failure

* Adding the device tag

---------

Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>

src/include/rccl_float8.h

@@ -24,8 +24,9 @@
 #define ROCBLAS_FLOAT8_H
 
 #include <stdint.h>
+#include <hip/hip_version.h>
 
-#if __cplusplus < 201103L || (!defined(__HCC__) && !defined(__HIPCC__))
+#if __cplusplus < 201103L || (!defined(__HIP_PLATFORM_AMD__) && !defined(__HIPCC__))
 /*! \brief Struct to represent a 8 bit floating-point number. */
 
 typedef struct
@@ -38,7 +39,55 @@ typedef struct
     uint8_t data;
 } rccl_bfloat8;
 
-#else // __cplusplus < 201103L || (!defined(__HCC__) && !defined(__HIPCC__))
+// __cplusplus < 201103L || (!defined(__HIP_PLATFORM_AMD__) && !defined(__HIPCC__))
+#elif HIP_VERSION >= 60200000
+
+#include <hip/hip_fp8.h>
+
+#if   __HIP_DEVICE_COMPILE__ && (defined(__gfx950__) || defined(__gfx1200__) || defined(__gfx1201__) ||  (defined(__gfx1100__) || defined(__gfx1101__)))//HIP_FP8_TYPE_OCP is enabled.
+typedef __hip_fp8_e4m3 rccl_float8;
+typedef __hip_fp8_e5m2 rccl_bfloat8;
+#elif __HIP_DEVICE_COMPILE__ && (defined(__gfx942__))
+typedef __hip_fp8_e4m3_fnuz rccl_float8;
+typedef __hip_fp8_e5m2_fnuz rccl_bfloat8;
+#else
+typedef __hip_fp8_e4m3 rccl_float8;
+typedef __hip_fp8_e5m2 rccl_bfloat8;
+#endif
+    
+inline std::ostream& operator<<(std::ostream& os, const rccl_float8& f8)
+{
+    return os << float(f8);
+}
+
+inline std::ostream& operator<<(std::ostream& os, const rccl_bfloat8& bf8)
+{
+    return os << float(bf8);
+}
+
+inline __host__ __device__ float operator*(rccl_float8 a, rccl_float8 b)
+{
+    return float(a) * float(b);
+}
+
+inline __host__ __device__ float operator*(rccl_bfloat8 a, rccl_bfloat8 b)
+{
+    return float(a) * float(b);
+}
+
+inline __host__ __device__ float operator*(rccl_float8 a, float b)
+{
+    return float(a) * float(b);
+}
+
+inline __host__ __device__ float operator*(rccl_bfloat8 a, float b)
+{
+    return float(a) * float(b);
+}
+
+// For older versions of ROCm that do not include hip_fp8.h,
+// we provide a local version of the header file as a fallback.
+#else
 
 #define HIP_HOST_DEVICE __host__ __device__
 #define HIP_HOST __host__
@@ -1016,6 +1065,6 @@ inline __host__ __device__ T explicit_downcast(Ta a, uint32_t rng)
 
 // =================================================================================================
 
-#endif // __cplusplus < 201103L || (!defined(__HCC__) && !defined(__HIPCC__))
+#endif
 
 #endif // ROCBLAS_FLOAT8_H

test/common/StandaloneUtils.hpp

@@ -26,7 +26,7 @@
 } while(0)
 
 // should be 112, temp fix to make CI pass
-#define MAX_STACK_SIZE 480
+#define MAX_STACK_SIZE 496
 
 #ifdef ENABLE_LL128
 #define MAX_STACK_SIZE_gfx90a 360