batch: relax SIMD loop for certain operations

When building with a newer version of clang that has more OMP
functionality, I uncovered a few more SIMD pragma loops fail to
vectorize on clang, so need to use the OSL_OMP_COMPLEX_SIMD_LOOP
macro to avoid those errors.

Signed-off-by: Larry Gritz <[email protected]>

Author: lgritz

src/include/OSL/platform.h

@@ -291,6 +291,7 @@
 #define OSL_OMP_SIMD_LOOP(...) OSL_OMP_PRAGMA(omp simd __VA_ARGS__)
 
 #if (OSL_GNUC_VERSION || OSL_INTEL_CLASSIC_COMPILER_VERSION || OSL_INTEL_LLVM_COMPILER_VERSION)
+    // GCC, icc, icx: Use a simd loop for sure
 #   define OSL_OMP_COMPLEX_SIMD_LOOP(...) OSL_OMP_SIMD_LOOP(__VA_ARGS__)
 #else
     // Ignore requests to vectorize complex/nested SIMD loops for certain

src/liboslexec/wide/wide_opalgebraic.cpp

@@ -136,7 +136,7 @@ __OSL_OP2(length, Wf, Wv)(void* r_, void* V_)
         Wide<const Vec3> wV(V_);
         Wide<float> wr(r_);
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Vec3 V   = wV[lane];
             float r  = sfm::length(V);
@@ -155,7 +155,7 @@ __OSL_MASKED_OP2(length, Wf, Wv)(void* r_, void* V_, unsigned int mask_value)
         Wide<const Vec3> wV(V_);
         Masked<float> wr(r_, Mask(mask_value));
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Vec3 V = wV[lane];
             if (wr.mask()[lane]) {
@@ -217,7 +217,7 @@ __OSL_OP2(area, Wf, Wdv)(void* r_, void* DP_)
 
         Wide<float> wr(r_);
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Dual2<Vec3> DP = wDP[lane];
 
@@ -240,7 +240,7 @@ __OSL_MASKED_OP2(area, Wf, Wdv)(void* r_, void* DP_, unsigned int mask_value)
 
         Masked<float> wr(r_, Mask(mask_value));
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Dual2<Vec3> DP = wDP[lane];
             if (wr.mask()[lane]) {
@@ -455,7 +455,7 @@ __OSL_OP2(normalize, Wv, Wv)(void* r_, void* V_)
         Wide<const Vec3> wV(V_);
         Wide<Vec3> wr(r_);
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Vec3 V   = wV[lane];
             Vec3 N   = sfm::normalize(V);
@@ -473,7 +473,7 @@ __OSL_MASKED_OP2(normalize, Wv, Wv)(void* r_, void* V_, unsigned int mask_value)
         Wide<const Vec3> wV(V_);
         Masked<Vec3> wr(r_, Mask(mask_value));
 
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             Vec3 V = wV[lane];
             if (wr.mask()[lane]) {

src/liboslexec/wide/wide_opcolor.cpp

@@ -68,7 +68,7 @@ __OSL_MASKED_OP2(blackbody, Wv, Wf)(void* bsg_, void* wout_, void* wtemp_,
     Block<int> computeRequiredBlock;
     Wide<int> wcomputeRequired(computeRequiredBlock);
 
-    OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+    OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
     for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
         float temperature      = wL[lane];
         bool canNotLookup      = !cs.can_lookup_blackbody(temperature);

src/liboslexec/wide/wide_opspline.cpp

@@ -376,7 +376,7 @@ spline_evaluate_loop_over_wide(const MatrixT& M, RAccessorT wR, XAccessorT wX,
 
     OSL_FORCEINLINE_BLOCK
     {
-        OSL_OMP_PRAGMA(omp simd simdlen(vec_width))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(vec_width))
         for (int lane = 0; lane < vec_width; ++lane) {
             X_Type x   = wX[lane];
             auto knots = wK[lane];

src/liboslexec/wide/wide_opstring.cpp

@@ -98,7 +98,7 @@ __OSL_MASKED_OP2(hash, Wi, Ws)(void* wr_, void* ws_, unsigned int mask_value)
 
     OSL_FORCEINLINE_BLOCK
     {
-        OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
+        OSL_OMP_COMPLEX_SIMD_LOOP(simdlen(__OSL_WIDTH))
         for (int lane = 0; lane < __OSL_WIDTH; ++lane) {
             ustring s = wS[lane];
             if (wR.mask()[lane]) {