SWDEV-1 - Merge github PRs to amd-staging

Change-Id: I2944a63ddc2eec8dc1403d9790ffffbaec343385

catch/unit/warp/CMakeLists.txt

@@ -1,14 +1,19 @@
 # Common Tests - Test independent of all platforms
 set(TEST_SRC
-  warp_shfl_xor.cc
-  warp_shfl.cc
-  warp_shfl_up.cc
-  warp_shfl_down.cc
   warp_ballot.cc
   warp_any.cc
   warp_all.cc
 )
 
+if(HIP_PLATFORM MATCHES "amd")
+  set(TEST_SRC ${TEST_SRC}
+    warp_shfl.cc
+    warp_shfl_xor.cc
+    warp_shfl_up.cc
+    warp_shfl_down.cc
+  )
+endif()
+
 hip_add_exe_to_target(NAME WarpTest
                       TEST_SRC ${TEST_SRC}
                       TEST_TARGET_NAME build_tests)

catch/unit/warp/warp_common.hh

@@ -21,6 +21,7 @@ THE SOFTWARE.
 
 #include <hip_test_common.hh>
 #include <hip/hip_cooperative_groups.h>
+#include <hip/hip_fp16.h>
 
 static __device__ bool deactivate_thread(const uint64_t* const active_masks) {
   const auto warp =

catch/unit/warp/warp_shfl.cc

@@ -100,7 +100,7 @@ template <typename T> class WarpShfl : public WarpShflTest<WarpShfl<T>, T> {
  *  - Device supports warp shuffle
  */
 TEMPLATE_TEST_CASE("Unit_Warp_Shfl_Positive_Basic", "", int, unsigned int, long, unsigned long,
-                   long long, unsigned long long, float, double) {
+                   long long, unsigned long long, float, double, __half, __half2) {
   int device;
   hipDeviceProp_t device_properties;
   HIP_CHECK(hipGetDevice(&device));
@@ -111,11 +111,7 @@ TEMPLATE_TEST_CASE("Unit_Warp_Shfl_Positive_Basic", "", int, unsigned int, long,
     return;
   }
 
-  SECTION("Shfl with specified active mask and input values") {
-    WarpShfl<TestType>().run(false);
-  }
+  SECTION("Shfl with specified active mask and input values") { WarpShfl<TestType>().run(false); }
 
-  SECTION("Shfl with random active mask and input values") {
-    WarpShfl<TestType>().run(true);
-  }
+  SECTION("Shfl with random active mask and input values") { WarpShfl<TestType>().run(true); }
 }

catch/unit/warp/warp_shfl_common.hh

@@ -25,6 +25,15 @@ THE SOFTWARE.
 #include <resource_guards.hh>
 #include <utils.hh>
 
+static bool operator==(__half x, __half y) {
+  // __heq doesn't have a __host__ version
+  return static_cast<__half_raw>(x).x == static_cast<__half_raw>(y).x;
+}
+static bool operator!=(__half x, __half y) { return static_cast<__half_raw>(x).x != static_cast<__half_raw>(y).x; }
+
+static bool operator==(__half2 x, __half2 y) { return __hbeq2(x, y); }
+static bool operator!=(__half2 x, __half2 y) { return !(__hbeq2(x, y)); }
+
 template <typename Derived, typename T> class WarpShflTest {
  public:
   WarpShflTest() : warp_size_{get_warp_size()} {}
@@ -82,6 +91,16 @@ template <typename Derived, typename T> class WarpShflTest {
           return static_cast<T>(
               GenerateRandomReal(std::numeric_limits<T>().min(), std::numeric_limits<T>().max()));
         });
+      } else if constexpr (std::is_same_v<__half, T>) {
+        std::generate_n(input, grid_.thread_count_, [] {
+          return __float2half(GenerateRandomReal(std::numeric_limits<float>().min(),
+                                                 std::numeric_limits<float>().max()));
+        });
+      } else if constexpr (std::is_same_v<__half2, T>) {
+        std::generate_n(input, grid_.thread_count_, [] {
+          return __float2half2_rn(GenerateRandomReal(std::numeric_limits<float>().min(),
+                                                     std::numeric_limits<float>().max()));
+        });
       } else {
         std::generate_n(input, grid_.thread_count_, [] {
           return static_cast<T>(GenerateRandomInteger(std::numeric_limits<T>().min(),

catch/unit/warp/warp_shfl_down.cc

@@ -100,7 +100,7 @@ template <typename T> class WarpShflDown : public WarpShflTest<WarpShflDown<T>,
  *  - Device supports warp shuffle
  */
 TEMPLATE_TEST_CASE("Unit_Warp_Shfl_Down_Positive_Basic", "", int, unsigned int, long, unsigned long,
-                   long long, unsigned long long, float, double) {
+                   long long, unsigned long long, float, double, __half, __half2) {
   int device;
   hipDeviceProp_t device_properties;
   HIP_CHECK(hipGetDevice(&device));

catch/unit/warp/warp_shfl_up.cc

@@ -99,7 +99,7 @@ template <typename T> class WarpShflUp : public WarpShflTest<WarpShflUp<T>, T> {
  *  - Device supports warp shuffle
  */
 TEMPLATE_TEST_CASE("Unit_Warp_Shfl_Up_Positive_Basic", "", int, unsigned int, long, unsigned long,
-                   long long, unsigned long long, float, double) {
+                   long long, unsigned long long, float, double, __half, __half2) {
   int device;
   hipDeviceProp_t device_properties;
   HIP_CHECK(hipGetDevice(&device));

catch/unit/warp/warp_shfl_xor.cc

@@ -97,7 +97,7 @@ template <typename T> class WarpShflXOR : public WarpShflTest<WarpShflXOR<T>, T>
  *  - Device supports warp shuffle
  */
 TEMPLATE_TEST_CASE("Unit_Warp_Shfl_XOR_Positive_Basic", "", int, unsigned int, long, unsigned long,
-                   long long, unsigned long long, float, double) {
+                   long long, unsigned long long, float, double, __half, __half2) {
   int device;
   hipDeviceProp_t device_properties;
   HIP_CHECK(hipGetDevice(&device));