[ESIMD] Fix inconsistencies in the ESIMD API signatures. (#4800)

This is API breaking patch.

- Remove 'esimd_' prefix from public API names. This prefix is redundant, because all APIs are in ...esimd:: namespace already. API names follow C-for-Metal naming.

- esimd_sat -> saturate

Memory access API changes:

- Un-deprecate block_load/store
- rename slm_atomic -> slm_atomic_update, flat_atomic -> atomic_update. atomic is not
  used to avoid conflict with C++ atomic class
- Offset is now expected in bytes(was in elements).
  (in scatter, gather, scalar_load, scalar_store)
- Make 'offsets' argument preceed the 'val' argument to be consistent with
  other memory APIs.
  (in scatter, scatter_rgba, slm_scatter)
- Remove unused L1 / L3 CachHint template parameters.
  (in scatter, gather, scalar_load, scalar_store, scatter_rgba, gather_rgba)
- Rename enum EsimdFenceMask -> enum fence_mask and its elements (deprecate old ones)

  Old behavior is preserved in deprecated APIs :
scatter->scatter1
gather->gather1
scalar_load->scalar_load1
scalar_store->scalar_store1
gather_rgba->gather4
scatter_rgba->scatter4
slm_scatter->slm_store

Signed-off-by: Konstantin S Bobrovsky konstantin.s.bobrovsky@intel.com

Author: kbobrovs

sycl/include/sycl/ext/intel/experimental/esimd/common.hpp

@@ -17,23 +17,7 @@
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_ESIMD_KERNEL __attribute__((sycl_explicit_simd))
 #define SYCL_ESIMD_FUNCTION __attribute__((sycl_explicit_simd))
-#else
-#define SYCL_ESIMD_KERNEL
-#define SYCL_ESIMD_FUNCTION
-#endif
-
-__SYCL_INLINE_NAMESPACE(cl) {
-namespace sycl {
-namespace ext {
-namespace intel {
-namespace experimental {
-namespace esimd {
-
-using uchar = unsigned char;
-using ushort = unsigned short;
-using uint = unsigned int;
 
-#ifdef __SYCL_DEVICE_ONLY__
 // Mark a function being nodebug.
 #define ESIMD_NODEBUG __attribute__((nodebug))
 // Mark a "ESIMD global": accessible from all functions in current translation
@@ -43,14 +27,21 @@ using uint = unsigned int;
   __attribute__((opencl_private)) __attribute__((sycl_explicit_simd))
 // Bind a ESIMD global variable to a specific register.
 #define ESIMD_REGISTER(n) __attribute__((register_num(n)))
-#else
+
+#define __ESIMD_API ESIMD_NODEBUG ESIMD_INLINE
+#else // __SYCL_DEVICE_ONLY__
+#define SYCL_ESIMD_KERNEL
+#define SYCL_ESIMD_FUNCTION
+
 // TODO ESIMD define what this means on Windows host
 #define ESIMD_NODEBUG
 // On host device ESIMD global is a thread local static var. This assumes that
 // each work-item is mapped to a separate OS thread on host device.
 #define ESIMD_PRIVATE thread_local
 #define ESIMD_REGISTER(n)
-#endif
+
+#define __ESIMD_API ESIMD_INLINE
+#endif // __SYCL_DEVICE_ONLY__
 
 // Mark a function being noinline
 #define ESIMD_NOINLINE __attribute__((noinline))
@@ -70,6 +61,17 @@ using uint = unsigned int;
 #define __ESIMD_DEPR_ENUM_V(old, new, t)                                       \
   old __ESIMD_DEPRECATED(new) = static_cast<t>(new)
 
+__SYCL_INLINE_NAMESPACE(cl) {
+namespace sycl {
+namespace ext {
+namespace intel {
+namespace experimental {
+namespace esimd {
+
+using uchar = unsigned char;
+using ushort = unsigned short;
+using uint = unsigned int;
+
 /// Gen hardware supports applying saturation to results of some operation.
 /// This enum allows to control this behavior.
 enum class saturation : uint8_t { off, on };
@@ -237,8 +239,6 @@ enum class split_barrier_action : uint8_t {
 // For backward compatibility:
 using EsimdSbarrierType = split_barrier_action;
 
-#undef __ESIMD_DEPR_ENUM_V
-
 // Since EsimdSbarrierType values are deprecated, these macros will generate
 // deprecation message.
 #define ESIMD_SBARRIER_WAIT EsimdSbarrierType::WAIT

sycl/include/sycl/ext/intel/experimental/esimd/detail/memory_intrin.hpp

@@ -76,9 +76,7 @@ constexpr unsigned int ElemsPerAddrDecoding(unsigned int ElemsPerAddrEncoded) {
 } // __SYCL_INLINE_NAMESPACE(cl)
 
 // flat_read does flat-address gather
-template <typename Ty, int N, int NumBlk = 0,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N, int NumBlk = 0>
 __ESIMD_INTRIN
     __SEIEED::vector_type_t<Ty, N * __SEIEED::ElemsPerAddrDecoding(NumBlk)>
     __esimd_svm_gather(__SEIEED::vector_type_t<uint64_t, N> addrs,
@@ -111,9 +109,7 @@ __ESIMD_INTRIN
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_write does flat-address scatter
-template <typename Ty, int N, int NumBlk = 0,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N, int NumBlk = 0>
 __ESIMD_INTRIN void __esimd_svm_scatter(
     __SEIEED::vector_type_t<uint64_t, N> addrs,
     __SEIEED::vector_type_t<Ty, N * __SEIEED::ElemsPerAddrDecoding(NumBlk)>
@@ -144,8 +140,7 @@ __ESIMD_INTRIN void __esimd_svm_scatter(
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_block_read reads a block of data from one flat address
-template <typename Ty, int N, __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N>
 __ESIMD_INTRIN __SEIEED::vector_type_t<Ty, N>
 __esimd_svm_block_ld_unaligned(uint64_t addr)
 #ifdef __SYCL_DEVICE_ONLY__
@@ -181,8 +176,7 @@ __esimd_svm_block_ld(uint64_t addr)
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_block_write writes a block of data using one flat address
-template <typename Ty, int N, __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N>
 __ESIMD_INTRIN void __esimd_svm_block_st(uint64_t addr,
                                          __SEIEED::vector_type_t<Ty, N> vals)
 #ifdef __SYCL_DEVICE_ONLY__
@@ -221,9 +215,7 @@ __ESIMD_INTRIN void __esimd_oword_st(SurfIndAliasTy surf_ind, uint32_t offset,
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_read4 does flat-address gather4
-template <typename Ty, int N, __SEIEE::rgba_channel_mask Mask,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N, __SEIEE::rgba_channel_mask Mask>
 __SEIEED::vector_type_t<Ty, N * get_num_channels_enabled(Mask)> __ESIMD_INTRIN
 __esimd_svm_gather4_scaled(__SEIEED::vector_type_t<uint64_t, N> addrs,
                            __SEIEED::simd_mask_storage_t<N> pred = 1)
@@ -276,9 +268,7 @@ __esimd_svm_gather4_scaled(__SEIEED::vector_type_t<uint64_t, N> addrs,
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_write does flat-address scatter
-template <typename Ty, int N, __SEIEE::rgba_channel_mask Mask,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <typename Ty, int N, __SEIEE::rgba_channel_mask Mask>
 __ESIMD_INTRIN void __esimd_svm_scatter4_scaled(
     __SEIEED::vector_type_t<uint64_t, N> addrs,
     __SEIEED::vector_type_t<Ty, N * get_num_channels_enabled(Mask)> vals,
@@ -351,8 +341,7 @@ __ESIMD_INTRIN void __esimd_svm_scatter4_scaled(
 // @param elem_offsets - per-element offsets
 //
 template <typename Ty, int N, typename SurfIndAliasTy, int TySizeLog2,
-          int16_t Scale = 0, __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+          int16_t Scale = 0>
 __ESIMD_INTRIN __SEIEED::vector_type_t<Ty, N>
 __esimd_gather_scaled2(SurfIndAliasTy surf_ind, uint32_t global_offset,
                        __SEIEED::vector_type_t<uint32_t, N> elem_offsets)
@@ -392,8 +381,7 @@ __esimd_gather_scaled2(SurfIndAliasTy surf_ind, uint32_t global_offset,
 // @param vals - values to write
 //
 template <typename Ty, int N, typename SurfIndAliasTy, int TySizeLog2,
-          int16_t Scale = 0, __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+          int16_t Scale = 0>
 __ESIMD_INTRIN void
 __esimd_scatter_scaled(__SEIEED::simd_mask_storage_t<N> pred,
                        SurfIndAliasTy surf_ind, uint32_t global_offset,
@@ -411,9 +399,7 @@ __esimd_scatter_scaled(__SEIEED::simd_mask_storage_t<N> pred,
 #endif // __SYCL_DEVICE_ONLY__
 
 // flat_atomic: flat-address atomic
-template <__SEIEE::atomic_op Op, typename Ty, int N,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <__SEIEE::atomic_op Op, typename Ty, int N>
 __ESIMD_INTRIN __SEIEED::vector_type_t<Ty, N>
 __esimd_svm_atomic0(__SEIEED::vector_type_t<uint64_t, N> addrs,
                     __SEIEED::simd_mask_storage_t<N> pred)
@@ -425,9 +411,7 @@ __esimd_svm_atomic0(__SEIEED::vector_type_t<uint64_t, N> addrs,
 }
 #endif // __SYCL_DEVICE_ONLY__
 
-template <__SEIEE::atomic_op Op, typename Ty, int N,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <__SEIEE::atomic_op Op, typename Ty, int N>
 __ESIMD_INTRIN __SEIEED::vector_type_t<Ty, N>
 __esimd_svm_atomic1(__SEIEED::vector_type_t<uint64_t, N> addrs,
                     __SEIEED::vector_type_t<Ty, N> src0,
@@ -440,9 +424,7 @@ __esimd_svm_atomic1(__SEIEED::vector_type_t<uint64_t, N> addrs,
 }
 #endif // __SYCL_DEVICE_ONLY__
 
-template <__SEIEE::atomic_op Op, typename Ty, int N,
-          __SEIEE::CacheHint L1H = __SEIEE::CacheHint::None,
-          __SEIEE::CacheHint L3H = __SEIEE::CacheHint::None>
+template <__SEIEE::atomic_op Op, typename Ty, int N>
 __ESIMD_INTRIN __SEIEED::vector_type_t<Ty, N>
 __esimd_svm_atomic2(__SEIEED::vector_type_t<uint64_t, N> addrs,
                     __SEIEED::vector_type_t<Ty, N> src0,

sycl/include/sycl/ext/intel/experimental/esimd/detail/simd_obj_impl.hpp

@@ -648,9 +648,7 @@ void simd_obj_impl<T, N, T1, SFINAE>::copy_from(const T *Addr)
                 "block size must be at most 8 owords");
 
   uintptr_t AddrVal = reinterpret_cast<uintptr_t>(Addr);
-  *this =
-      __esimd_svm_block_ld_unaligned<T, N, CacheHint::None, CacheHint::None>(
-          AddrVal);
+  *this = __esimd_svm_block_ld_unaligned<T, N>(AddrVal);
 }
 
 template <typename T, int N, class T1, class SFINAE>
@@ -691,7 +689,7 @@ void simd_obj_impl<T, N, T1, SFINAE>::copy_to(T *addr) const
                 "block size must be at most 8 owords");
 
   uintptr_t AddrVal = reinterpret_cast<uintptr_t>(addr);
-  __esimd_svm_block_st<T, N, CacheHint::None, CacheHint::None>(AddrVal, data());
+  __esimd_svm_block_st<T, N>(AddrVal, data());
 }
 
 template <typename T, int N, class T1, class SFINAE>