SWDEV-460834 - add unsafe atomic add for fp16 and bf16

Change-Id: I6de5c2c425c9f8ac7f6c4e5c83c8b8b7ac8fe4cb

hipamd/include/hip/amd_detail/amd_hip_bf16.h

@@ -102,6 +102,9 @@
 
 #include "amd_hip_vector_types.h"  // float2 etc
 #include "device_library_decls.h"  // ocml conversion functions
+#if defined(__clang__) && defined(__HIP__)
+#include "amd_hip_atomic.h"
+#endif // defined(__clang__) && defined(__HIP__)
 #include "math_fwd.h"              // ocml device functions
 
 #define __BF16_DEVICE__ __device__
@@ -1811,4 +1814,36 @@ __BF16_DEVICE_STATIC__ __hip_bfloat162 h2trunc(const __hip_bfloat162 h) {
   __hip_bfloat162_raw hr = h;
   return __hip_bfloat162(htrunc(__hip_bfloat16_raw{hr.x}), htrunc(__hip_bfloat16_raw{hr.y}));
 }
+
+#if defined(__clang__) && defined(__HIP__)
+/**
+ * \ingroup HIP_INTRINSIC_BFLOAT162_MATH
+ * \brief Atomic add bfloat162
+ */
+__BF16_DEVICE_STATIC__ __hip_bfloat162 unsafeAtomicAdd(__hip_bfloat162* address,
+                                                       __hip_bfloat162 value) {
+#if defined(__AMDGCN_UNSAFE_FP_ATOMICS__) && __has_builtin(__builtin_amdgcn_flat_atomic_fadd_v2bf16)
+  typedef short __attribute__((ext_vector_type(2))) vec_short2;
+  __hip_bfloat162_raw bf2_v = value;
+  vec_short2 s2_in{bf2_v.x, bf2_v.y};
+  vec_short2 s2_ret = __builtin_amdgcn_flat_atomic_fadd_v2bf16((vec_short2*)address, s2_in);
+  return __hip_bfloat162_raw{s2_ret[0], s2_ret[1]};
+#else
+  static_assert(sizeof(unsigned int) == sizeof(__hip_bfloat162_raw));
+  union u_hold {
+    __hip_bfloat162_raw h2r;
+    unsigned int u32;
+  };
+  u_hold old_val, new_val;
+  old_val.u32 =
+      __hip_atomic_load((unsigned int*)address, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);
+  do {
+    new_val.h2r = __hadd2(old_val.h2r, value);
+  } while (!__hip_atomic_compare_exchange_strong((unsigned int*)address, &old_val.u32, new_val.u32,
+                                                 __ATOMIC_RELAXED, __ATOMIC_RELAXED,
+                                                 __HIP_MEMORY_SCOPE_AGENT));
+  return old_val.h2r;
+#endif
+}
+#endif  // defined(__clang__) && defined(__HIP__)
 #endif

hipamd/include/hip/amd_detail/amd_hip_fp16.h

@@ -30,6 +30,9 @@ THE SOFTWARE.
   #define __HOST_DEVICE__ __host__ __device__
   #include <hip/amd_detail/amd_hip_common.h>
   #include "hip/amd_detail/host_defines.h"
+#if defined(__clang__) && defined(__HIP__)
+  #include "hip/amd_detail/amd_hip_atomic.h"
+#endif // defined(__clang__) && defined(__HIP__)
   #include <assert.h>
   #if defined(__cplusplus)
     #include <algorithm>
@@ -1503,6 +1506,39 @@ THE SOFTWARE.
                     static_cast<__half2_raw>(y).data};
             }
 
+            // Atomic
+            #if defined(__clang__) && defined(__HIP__)
+            inline __device__ __half2 unsafeAtomicAdd(__half2* address, __half2 value) {
+            #if defined(__AMDGCN_UNSAFE_FP_ATOMICS__) && __has_builtin(__builtin_amdgcn_flat_atomic_fadd_v2f16)
+                // The api expects an ext_vector_type of half
+                typedef __fp16 __attribute__((ext_vector_type(2))) vec_fp162;
+                static_assert(sizeof(vec_fp162) == sizeof(__half2_raw));
+                union {
+                    __half2_raw h2r;
+                    vec_fp162 fp16;
+                } u {value};
+                vec_fp162 ret =
+                    __builtin_amdgcn_flat_atomic_fadd_v2f16((vec_fp162*)address, u.fp16);
+                return __half2{ret[0], ret[1]};
+            #else
+                static_assert(sizeof(__half2_raw) == sizeof(unsigned int));
+                union u_hold {
+                    __half2_raw h2r;
+                    unsigned int u32;
+                };
+                u_hold old_val, new_val;
+                old_val.u32 = __hip_atomic_load((unsigned int*)address, __ATOMIC_RELAXED,
+                                                __HIP_MEMORY_SCOPE_AGENT);
+                do {
+                    new_val.h2r = __hadd2(old_val.h2r, value);
+                } while (!__hip_atomic_compare_exchange_strong(
+                    (unsigned int*)address, &old_val.u32, new_val.u32, __ATOMIC_RELAXED,
+                    __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT));
+                return old_val.h2r;
+            #endif
+            }
+            #endif // defined(__clang__) && defined(__HIP__)
+
             // Math functions
             #if defined(__clang__) && defined(__HIP__)
             inline