build: refactor icc/icx specific pragma macros (#1756)

Added OSL_INTEL_CLASSIC_PRAGMA(...) which only emits for Intel(r) Classic C++ Compiler, previously OSL_INTEL_PRAGMA did that.
As many Intel compiler specicif pragmas are now supported by icx and icc,
redefined OSL_INTEL_PRAGMA to emit for the classic and llvm based Intel compilers (icc and icx).
Updated some uses of OSL_INTEL_PRAGMA to OSL_INTEL_CLASSIC_PRAGMA as appropriate.

Fixed long icx 2023.1.0, 2023.2.0, 2023.2.1 build times for certain files by removing the "-mllvm -inline-threshold=100000" flag for all but 1 file avoiding the overhead of optimization passes being applied to inlined code that wasn't actually part of the SIMD loop that we truly needed inlined.  icx now supports "#pragma forceinline recursive" which allows us to target just the SIMD loops to be inlined (as we did for icc).

Signed-off-by: Alex M. Wells <[email protected]>

src/include/OSL/Imathx/Imathx.h

@@ -265,7 +265,7 @@ nonAffineInverse(const Matrix44 &source);
 
 Matrix44 OSL_HOSTDEVICE nonAffineInverse(const Matrix44 &source)
 {
-    OSL_INTEL_PRAGMA(float_control(strict,on,push))
+    OSL_INTEL_CLASSIC_PRAGMA(float_control(strict,on,push))
     OSL_CLANG_PRAGMA(clang fp contract(off))
 
 	using ScalarT = typename Matrix44::BaseType;

src/include/OSL/platform.h

@@ -206,9 +206,14 @@
 #        define OSL_CLANG_PRAGMA(UnQuotedPragma)
 #    endif
 #    if defined(__INTEL_COMPILER)
-#        define OSL_INTEL_PRAGMA(UnQuotedPragma) OSL_PRAGMA(UnQuotedPragma)
+#        define OSL_INTEL_CLASSIC_PRAGMA(UnQuotedPragma) OSL_PRAGMA(UnQuotedPragma)
 #    else
-#        define OSL_INTEL_PRAGMA(UnQuotedPragma)
+#        define OSL_INTEL_CLASSIC_PRAGMA(UnQuotedPragma)
+#    endif
+#    if defined(__INTEL_LLVM_COMPILER)
+#        define OSL_INTEL_LLVM_PRAGMA(UnQuotedPragma) OSL_PRAGMA(UnQuotedPragma)
+#    else
+#        define OSL_INTEL_LLVM_PRAGMA(UnQuotedPragma)
 #    endif
 #    define OSL_MSVS_PRAGMA(UnQuotedPragma)
 #elif defined(_MSC_VER)
@@ -218,7 +223,8 @@
 #    define OSL_PRAGMA_VISIBILITY_POP  /* N/A on MSVS */
 #    define OSL_GCC_PRAGMA(UnQuotedPragma)
 #    define OSL_CLANG_PRAGMA(UnQuotedPragma)
-#    define OSL_INTEL_PRAGMA(UnQuotedPragma)
+#    define OSL_INTEL_CLASSIC_PRAGMA(UnQuotedPragma)
+#    define OSL_INTEL_LLVM_PRAGMA(UnQuotedPragma)
 #    define OSL_MSVS_PRAGMA(UnQuotedPragma) OSL_PRAGMA(UnQuotedPragma)
 #else
 #    define OSL_PRAGMA_WARNING_PUSH
@@ -227,10 +233,14 @@
 #    define OSL_PRAGMA_VISIBILITY_POP
 #    define OSL_GCC_PRAGMA(UnQuotedPragma)
 #    define OSL_CLANG_PRAGMA(UnQuotedPragma)
-#    define OSL_INTEL_PRAGMA(UnQuotedPragma)
+#    define OSL_INTEL_CLASSIC_PRAGMA(UnQuotedPragma)
+#    define OSL_INTEL_LLVM_PRAGMA(UnQuotedPragma)
 #    define OSL_MSVS_PRAGMA(UnQuotedPragma)
 #endif
 
+// A pragma that applies to both icc and icx
+#define OSL_INTEL_PRAGMA(UnQuotedPragma) OSL_INTEL_CLASSIC_PRAGMA(UnQuotedPragma) OSL_INTEL_LLVM_PRAGMA(UnQuotedPragma)
+
 #ifdef __clang__
     #define OSL_CLANG_ATTRIBUTE(value) __attribute__((value))
 #else

src/liboslexec/CMakeLists.txt

@@ -103,7 +103,11 @@ set ( liboslexec_require_INF_NaN
     wide/wide_shadingsys
     wide/wide_optest_float
     )
-
+
+set ( liboslexec_exceeds_INTELCLANG_inlining_limits
+    wide/wide_opspline
+    )
+
 set (local_lib oslexec)
 set (lib_src
           shadingsys.cpp closure.cpp
@@ -390,11 +394,15 @@ foreach(batched_target ${BATCHED_TARGET_LIST})
             list (APPEND TARGET_CXX_OPTS "-ffp-contract=off")
         endif ()
 
-        # large SIMD function loops will exceed llvm's -inline-threshold default of 225.
-        # remark: loop not vectorized: call instruction cannot be vectorized [-Rpass-analysis]     
-        # choose to increase that limit via compiler flags vs. 
-        # workaround with __attribute__((flatten))     
-        list (APPEND TARGET_CXX_OPTS "-mllvm" "-inline-threshold=100000")
+        if (NOT CMAKE_COMPILER_IS_INTELCLANG)
+            # large SIMD function loops will exceed llvm's -inline-threshold default of 225.
+            # remark: loop not vectorized: call instruction cannot be vectorized [-Rpass-analysis]     
+            # choose to increase that limit via compiler flags vs. 
+            # workaround with __attribute__((flatten))     
+            list (APPEND TARGET_CXX_OPTS "-mllvm" "-inline-threshold=100000")
+            # NOTE: INTELCLANG we use icx/icc specific "#pragma force inline recursive"
+            #       to only inline the SIMD loops contents vs. everything.
+        endif ()
 
         # For loops with small loop bodies, clang was unrolling the loop before
         # #pragma omp simd
@@ -489,7 +497,18 @@ foreach(batched_target ${BATCHED_TARGET_LIST})
         if (${target_src} IN_LIST liboslexec_require_INF_NaN)
             REQUIRE_INF_NAN ( ${TARGET_SRC} ) 
         endif()
-
+
+        if (CMAKE_COMPILER_IS_INTELCLANG)
+            if (${target_src} IN_LIST liboslexec_exceeds_INTELCLANG_inlining_limits)
+                # NOTE: Some files still require increasing
+                #       inlining threshold to successfully vectorize.
+                #       Retest with future compiler versions
+                set_property(SOURCE ${TARGET_SRC} APPEND PROPERTY COMPILE_OPTIONS 
+                    "-mllvm" "-inline-threshold=100000")
+            endif()
+        endif()
+
+
     endforeach(target_src)
 
     add_library ( ${batched_target_lib} MODULE ${TARGET_LIB_SOURCES} )

src/liboslexec/batched_analysis.cpp

@@ -448,7 +448,7 @@ could_op_return_be_logically_boolean(ustring opname)
 // comparison operators even if the are not "currently"
 // being referenced.  May need to disable some warnings
 // error #177: function "..." was declared but never referenced
-OSL_INTEL_PRAGMA(warning(disable : 177))
+OSL_INTEL_CLASSIC_PRAGMA(warning(disable : 177))
 
 
 // The Position returned by top_pos changes and symbols are pushed and popped.