Merge branch 'master' into tex_unbind_issue_fix

[ROCm/clr commit: 9332a39838]

projects/clr/hipamd/include/hip/hcc_detail/hip_atomic.h

@@ -58,15 +58,15 @@ inline
 float atomicAdd(float* address, float val)
 {
     unsigned int* uaddr{reinterpret_cast<unsigned int*>(address)};
-    unsigned int old{__atomic_load_n(uaddr, __ATOMIC_RELAXED)};
+    unsigned int r{__atomic_load_n(uaddr, __ATOMIC_RELAXED)};
 
-    unsigned int r;
+    unsigned int old;
     do {
-        r = __atomic_load_n(uaddr, __ATOMIC_RELAXED);
+        old = __atomic_load_n(uaddr, __ATOMIC_RELAXED);
 
         if (r != old) { r = old; continue; }
 
-        old = atomicCAS(uaddr, r, __float_as_uint(val + __uint_as_float(r)));
+        r = atomicCAS(uaddr, r, __float_as_uint(val + __uint_as_float(r)));
 
         if (r == old) break;
     } while (true);
@@ -78,15 +78,15 @@ inline
 double atomicAdd(double* address, double val)
 {
     unsigned long long* uaddr{reinterpret_cast<unsigned long long*>(address)};
-    unsigned long long old{__atomic_load_n(uaddr, __ATOMIC_RELAXED)};
+    unsigned long long r{__atomic_load_n(uaddr, __ATOMIC_RELAXED)};
 
-    unsigned long long r;
+    unsigned long long old;
     do {
-        r = __atomic_load_n(uaddr, __ATOMIC_RELAXED);
+        old = __atomic_load_n(uaddr, __ATOMIC_RELAXED);
 
         if (r != old) { r = old; continue; }
 
-        old = atomicCAS(
+        r = atomicCAS(
             uaddr, r, __double_as_longlong(val + __longlong_as_double(r)));
 
         if (r == old) break;
@@ -140,13 +140,13 @@ __device__
 inline
 int atomicMin(int* address, int val)
 {
-    return __sync_fetch_and_min(address, val);
+    return __atomic_fetch_min(address, val, __ATOMIC_RELAXED);
 }
 __device__
 inline
 unsigned int atomicMin(unsigned int* address, unsigned int val)
 {
-    return __sync_fetch_and_umin(address, val);
+    return __atomic_fetch_min(address, val, __ATOMIC_RELAXED);
 }
 __device__
 inline
@@ -169,13 +169,13 @@ __device__
 inline
 int atomicMax(int* address, int val)
 {
-    return __sync_fetch_and_max(address, val);
+    return __atomic_fetch_max(address, val, __ATOMIC_RELAXED);
 }
 __device__
 inline
 unsigned int atomicMax(unsigned int* address, unsigned int val)
 {
-    return __sync_fetch_and_umax(address, val);
+    return __atomic_fetch_max(address, val, __ATOMIC_RELAXED);
 }
 __device__
 inline

projects/clr/hipamd/include/hip/hcc_detail/hip_runtime_api.h

@@ -2063,6 +2063,45 @@ hipError_t hipMemcpyToArray(hipArray* dst, size_t wOffset, size_t hOffset, const
 hipError_t hipMemcpyFromArray(void* dst, hipArray_const_t srcArray, size_t wOffset, size_t hOffset,
                               size_t count, hipMemcpyKind kind);
 
+/**
+ *  @brief Copies data between host and device.
+ *
+ *  @param[in]   dst       Destination memory address
+ *  @param[in]   dpitch    Pitch of destination memory
+ *  @param[in]   src       Source memory address
+ *  @param[in]   wOffset   Source starting X offset
+ *  @param[in]   hOffset   Source starting Y offset
+ *  @param[in]   width     Width of matrix transfer (columns in bytes)
+ *  @param[in]   height    Height of matrix transfer (rows)
+ *  @param[in]   kind      Type of transfer
+ *  @return      #hipSuccess, #hipErrorInvalidValue, #hipErrorInvalidPitchValue,
+ * #hipErrorInvalidDevicePointer, #hipErrorInvalidMemcpyDirection
+ *
+ *  @see hipMemcpy, hipMemcpy2DToArray, hipMemcpy2D, hipMemcpyFromArray, hipMemcpyToSymbol,
+ * hipMemcpyAsync
+ */
+hipError_t hipMemcpy2DFromArray( void* dst, size_t dpitch, hipArray_const_t src, size_t wOffset, size_t hOffset, size_t width, size_t height, hipMemcpyKind kind);
+
+/**
+ *  @brief Copies data between host and device asynchronously.
+ *
+ *  @param[in]   dst       Destination memory address
+ *  @param[in]   dpitch    Pitch of destination memory
+ *  @param[in]   src       Source memory address
+ *  @param[in]   wOffset   Source starting X offset
+ *  @param[in]   hOffset   Source starting Y offset
+ *  @param[in]   width     Width of matrix transfer (columns in bytes)
+ *  @param[in]   height    Height of matrix transfer (rows)
+ *  @param[in]   kind      Type of transfer
+ *  @param[in]   stream    Accelerator view which the copy is being enqueued
+ *  @return      #hipSuccess, #hipErrorInvalidValue, #hipErrorInvalidPitchValue,
+ * #hipErrorInvalidDevicePointer, #hipErrorInvalidMemcpyDirection
+ *
+ *  @see hipMemcpy, hipMemcpy2DToArray, hipMemcpy2D, hipMemcpyFromArray, hipMemcpyToSymbol,
+ * hipMemcpyAsync
+ */
+hipError_t hipMemcpy2DFromArrayAsync( void* dst, size_t dpitch, hipArray_const_t src, size_t wOffset, size_t hOffset, size_t width, size_t height, hipMemcpyKind kind, hipStream_t stream __dparm(0));
+
 /**
  *  @brief Copies data between host and device.
  *

projects/clr/hipamd/include/hip/hcc_detail/hip_vector_types.h

@@ -44,18 +44,35 @@ THE SOFTWARE.
         __attribute__((vector_size(__ROUND_UP_TO_NEXT_POT__(n) * sizeof(T))))
 #endif
 
-#if defined(__cplusplus)
+#if defined(__cplusplus) && defined(__clang__)
     #include <type_traits>
 
     namespace hip_impl {
         template<typename T, typename Vector, unsigned int idx>
         struct Scalar_accessor {
+            struct Address {
+                const Scalar_accessor* p;
+
+                __host__ __device__
+                operator const T*() const noexcept {
+                    return &reinterpret_cast<const T*>(p)[idx];
+                }
+                __host__ __device__
+                operator T*() noexcept {
+                    return &reinterpret_cast<T*>(
+                        const_cast<Scalar_accessor*>(p))[idx];
+                }
+            };
+
             // Idea from https://t0rakka.silvrback.com/simd-scalar-accessor
             Vector data;
 
             __host__ __device__
             operator T() const noexcept { return data[idx]; }
 
+            __host__ __device__
+            Address operator&() const noexcept { return Address{this}; }
+
             __host__ __device__
             Scalar_accessor& operator=(T x) noexcept {
                 data[idx] = x;
@@ -63,6 +80,29 @@ THE SOFTWARE.
                 return *this;
             }
 
+            __host__ __device__
+            Scalar_accessor& operator++() noexcept {
+                ++data[idx];
+                return *this;
+            }
+            __host__ __device__
+            T operator++(int) noexcept {
+                auto r{data[idx]};
+                ++data[idx];
+                return *this;
+            }
+            __host__ __device__
+            Scalar_accessor& operator--() noexcept {
+                --data[idx];
+                return *this;
+            }
+            __host__ __device__
+            T operator--(int) noexcept {
+                auto r{data[idx]};
+                --data[idx];
+                return *this;
+            }
+
             __host__ __device__
             Scalar_accessor& operator+=(T x) noexcept {
                 data[idx] += x;