Merge pull request #863 from aaronenyeshi/add-dot-instructions

Add Dot functions as amd_mixed_dot function

hipamd/include/hip/hcc_detail/hip_fp16.h

@@ -1403,6 +1403,15 @@ THE SOFTWARE.
             }
 
             // Math functions
+            #if (__hcc_workweek__ >= 19015) || __HIP_CLANG_ONLY__
+            inline
+            __device__
+            float amd_mixed_dot(__half2 a, __half2 b, float c, bool saturate) {
+                return __ockl_fdot2(static_cast<__half2_raw>(a).data,
+                                    static_cast<__half2_raw>(b).data,
+                                    c, saturate);
+            }
+            #endif
             inline
             __device__
             __half htrunc(__half x)

hipamd/include/hip/hcc_detail/hip_fp16_math_fwd.h

@@ -53,6 +53,10 @@ extern "C"
     typedef _Float16 __2f16 __attribute__((ext_vector_type(2)));
     typedef short __2i16 __attribute__((ext_vector_type(2)));
 
+    #if (__hcc_workweek__ >= 19015) || __HIP_CLANG_ONLY__
+    __device__ __attribute__((const)) float __ockl_fdot2(__2f16 a, __2f16 b, float c, bool s);
+    #endif
+
     __device__ __attribute__((const)) __2f16 __ocml_ceil_2f16(__2f16);
     __device__ __2f16 __ocml_cos_2f16(__2f16);
     __device__ __attribute__((pure)) __2f16 __ocml_exp_2f16(__2f16);
@@ -75,4 +79,4 @@ extern "C"
     __device__ __2f16 __ocml_sin_2f16(__2f16);
     __device__ __attribute__((const)) __2f16 __ocml_sqrt_2f16(__2f16);
     __device__ __attribute__((const)) __2f16 __ocml_trunc_2f16(__2f16);
-}
+}

hipamd/include/hip/hcc_detail/math_functions.h

@@ -23,6 +23,7 @@ THE SOFTWARE.
 #pragma once
 
 #include "hip_fp16_math_fwd.h"
+#include "hip_vector_types.h"
 #include "math_fwd.h"
 
 #include <hip/hcc_detail/host_defines.h>
@@ -119,6 +120,40 @@ uint64_t __make_mantissa(const char* tagp)
     return __make_mantissa_base10(tagp);
 }
 
+// DOT FUNCTIONS
+#if (__hcc_workweek__ >= 19015) || __HIP_CLANG_ONLY__
+__DEVICE__
+inline
+int amd_mixed_dot(short2 a, short2 b, int c, bool saturate) {
+    return __ockl_sdot2(a.data, b.data, c, saturate);
+}
+__DEVICE__
+inline
+uint amd_mixed_dot(ushort2 a, ushort2 b, uint c, bool saturate) {
+    return __ockl_udot2(a.data, b.data, c, saturate);
+}
+__DEVICE__
+inline
+int amd_mixed_dot(char4 a, char4 b, int c, bool saturate) {
+    return __ockl_sdot4(a.data, b.data, c, saturate);
+}
+__DEVICE__
+inline
+uint amd_mixed_dot(uchar4 a, uchar4 b, uint c, bool saturate) {
+    return __ockl_udot4(a.data, b.data, c, saturate);
+}
+__DEVICE__
+inline
+int amd_mixed_dot(int a, int b, int c, bool saturate) {
+    return __ockl_sdot8(a, b, c, saturate);
+}
+__DEVICE__
+inline
+uint amd_mixed_dot(uint a, uint b, uint c, bool saturate) {
+    return __ockl_udot8(a, b, c, saturate);
+}
+#endif
+
 // BEGIN FLOAT
 __DEVICE__
 inline

hipamd/include/hip/hcc_detail/math_fwd.h

@@ -28,6 +28,45 @@ THE SOFTWARE.
     extern "C" {
 #endif
 
+// DOT FUNCTIONS
+#if (__hcc_workweek__ >= 19015) || __HIP_CLANG_ONLY__
+__device__
+__attribute__((const))
+int __ockl_sdot2(
+    HIP_vector_base<short, 2>::Native_vec_,
+    HIP_vector_base<short, 2>::Native_vec_,
+    int, bool);
+
+__device__
+__attribute__((const))
+unsigned int __ockl_udot2(
+    HIP_vector_base<unsigned short, 2>::Native_vec_,
+    HIP_vector_base<unsigned short, 2>::Native_vec_,
+    unsigned int, bool);
+
+__device__
+__attribute__((const))
+int __ockl_sdot4(
+    HIP_vector_base<char, 4>::Native_vec_,
+    HIP_vector_base<char, 4>::Native_vec_,
+    int, bool);
+
+__device__
+__attribute__((const))
+unsigned int __ockl_udot4(
+    HIP_vector_base<unsigned char, 4>::Native_vec_,
+    HIP_vector_base<unsigned char, 4>::Native_vec_,
+    unsigned int, bool);
+
+__device__
+__attribute__((const))
+int __ockl_sdot8(int, int, int, bool);
+
+__device__
+__attribute__((const))
+unsigned int __ockl_udot8(unsigned int, unsigned int, unsigned int, bool);
+#endif
+
 // BEGIN FLOAT
 __device__
 __attribute__((const))