Merge branch 'develop' into doc2

src/targets/gpu/device/include/migraphx/gpu/device/launch.hpp

@@ -78,8 +78,9 @@ inline auto gs_launch(hipStream_t stream, index_int n, index_int local = 1024)
     index_int nglobal = std::min<index_int>(256, groups) * local;
 
     return [=](auto f) {
-        launch(stream, nglobal, local)(
-            [=](auto idx) { idx.global_stride(n, [&](auto i) { gs_invoke(f, i, idx); }); });
+        launch(stream, nglobal, local)([=](auto idx) __device__ {
+            idx.global_stride(n, [&](auto i) { gs_invoke(f, i, idx); });
+        });
     };
 }
 

src/targets/gpu/device/include/migraphx/gpu/device/multi_index.hpp

@@ -95,7 +95,7 @@ inline auto mi_launch(hipStream_t stream, const hip_shape<N>& global, index_int
     auto nglobal       = global.index(nglobal_multi);
 
     return [=](auto f) {
-        launch(stream, nglobal, nlocal)([=](auto idx) {
+        launch(stream, nglobal, nlocal)([=](auto idx) __device__ {
             auto midx = make_multi_index(global, idx.global, nglobal_multi);
             f(idx, midx.for_stride(global.lens));
         });

src/targets/gpu/device/include/migraphx/gpu/device/reduce.hpp

@@ -20,6 +20,15 @@ struct sum
     }
 };
 
+struct product
+{
+    template <class T, class U>
+    MIGRAPHX_DEVICE_CONSTEXPR auto operator()(T x, U y) const
+    {
+        return x * y;
+    }
+};
+
 struct id
 {
     template <class T>

src/targets/gpu/device/include/migraphx/gpu/device/visit.hpp

@@ -39,7 +39,12 @@ constexpr void visit_tensor_size(index_int n, F f)
         f(std::integral_constant<index_int, 5>{});
         break;
     }
-    default: throw std::runtime_error("Unknown tensor size");
+    case 6:
+    {
+        f(std::integral_constant<index_int, 6>{});
+        break;
+    }
+    default: throw std::runtime_error("Tensor size dim out of range");
     }
 }
 

src/targets/gpu/device/int8_gemm_pack.cpp

@@ -24,7 +24,7 @@ void int8_gemm_pack_a(hipStream_t stream, const argument& result, const argument
         auto* in_ptr          = device_cast(input.data());
         visit_tensor_size(out_lens.size(), [&](auto out_dim) {
             hip_tensor_descriptor<out_dim> desc(comp_shape);
-            gs_launch(stream, nelements, 256)([=](auto ii) {
+            gs_launch(stream, nelements, 256)([=](auto ii) __device__ {
                 const size_t nb    = 4;
                 auto idx           = desc.multi(ii);
                 std::size_t i_m    = idx[dim_1];
@@ -55,7 +55,7 @@ void int8_gemm_pack_b(hipStream_t stream, const argument& result, const argument
         auto* in_ptr          = device_cast(input.data());
         visit_tensor_size(out_lens.size(), [&](auto out_dim) {
             hip_tensor_descriptor<out_dim> desc(comp_shape);
-            gs_launch(stream, nelements, 256)([=](auto ii) {
+            gs_launch(stream, nelements, 256)([=](auto ii) __device__ {
                 const size_t nb    = 4;
                 auto idx           = desc.multi(ii);
                 std::size_t i_n    = idx[dim_1];

src/targets/gpu/device/log.cpp

@@ -9,7 +9,7 @@ namespace device {
 
 void log(hipStream_t stream, const argument& result, const argument& arg)
 {
-    nary(stream, result, arg)([](auto x) { return ::log(to_hip_type(x)); });
+    nary(stream, result, arg)([](auto x) __device__ { return ::log(to_hip_type(x)); });
 }
 
 } // namespace device