fix(simd.h): gather_mask was wrong for no-simd fallback (AcademySoftwareFoundation#4183)

The operation it's supposed to be performing is to merge in the gathered
value where the mask is set, and leave mask=0 lanes unchanged. But for
our "no SIMD" fallback (i.e., no AVX2 for 4- and 8-wide), we incorrectly
implemented it as setting the unmasked lanes to 0!

This was quite wrong, and unfortunately masked by the fact that our test
of the function initialized a variable to 0. Change to a better number.
The best number.

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

Author: lgritz

src/include/OpenImageIO/simd.h

@@ -4426,7 +4426,7 @@ vint4::gather_mask (const vbool_t& mask, const value_t *baseptr, const vint_t& v
 #if OIIO_SIMD_AVX >= 2
     m_simd = _mm_mask_i32gather_epi32 (m_simd, baseptr, vindex, _mm_cvtps_epi32(mask), scale);
 #else
-    SIMD_CONSTRUCT (mask[i] ? *(const value_t *)((const char *)baseptr + vindex[i]*scale) : 0);
+    SIMD_DO (if (mask[i]) m_val[i] = *(const value_t *)((const char *)baseptr + vindex[i]*scale));
 #endif
 }
 
@@ -5309,7 +5309,7 @@ vint8::gather_mask (const vbool_t& mask, const value_t *baseptr, const vint_t& v
 #if OIIO_SIMD_AVX >= 2
     m_simd = _mm256_mask_i32gather_epi32 (m_simd, baseptr, vindex, _mm256_cvtps_epi32(mask), scale);
 #else
-    SIMD_CONSTRUCT (mask[i] ? *(const value_t *)((const char *)baseptr + vindex[i]*scale) : 0);
+    SIMD_DO (if (mask[i]) m_val[i] = *(const value_t *)((const char *)baseptr + vindex[i]*scale));
 #endif
 }
 
@@ -7073,7 +7073,7 @@ vfloat4::gather_mask (const vbool_t& mask, const value_t *baseptr, const vint_t&
 #if OIIO_SIMD_AVX >= 2
     m_simd = _mm_mask_i32gather_ps (m_simd, baseptr, vindex, mask, scale);
 #else
-    SIMD_CONSTRUCT (mask[i] ? *(const value_t *)((const char *)baseptr + vindex[i]*scale) : 0);
+    SIMD_DO (if (mask[i]) m_val[i] = *(const value_t *)((const char *)baseptr + vindex[i]*scale));
 #endif
 }
 
@@ -8981,7 +8981,7 @@ vfloat8::gather_mask (const vbool_t& mask, const value_t *baseptr, const vint_t&
 #if OIIO_SIMD_AVX >= 2
     m_simd = _mm256_mask_i32gather_ps (m_simd, baseptr, vindex, mask, scale);
 #else
-    SIMD_CONSTRUCT (mask[i] ? *(const value_t *)((const char *)baseptr + vindex[i]*scale) : 0);
+    SIMD_DO (if (mask[i]) m_val[i] = *(const value_t *)((const char *)baseptr + vindex[i]*scale));
 #endif
 }
 

src/libutil/simd_test.cpp

@@ -607,9 +607,9 @@ test_gatherscatter()
     OIIO_CHECK_SIMD_EQUAL(g, VEC::Iota());
 
     BOOL mask = BOOL::from_bitmask(0x55555555);  // every other one
-    ELEM every_other_iota[] = { 0, 0, 2, 0, 4, 0, 6, 0, 8, 0, 10, 0, 12, 0, 14, 0 };
-    gm = 0;
+    gm = 42;
     gm.gather_mask (mask, gather_source.data(), indices);
+    ELEM every_other_iota[] = { 0, 42, 2, 42, 4, 42, 6, 42, 8, 42, 10, 42, 12, 42, 14, 42 };
     OIIO_CHECK_SIMD_EQUAL (gm, VEC(every_other_iota));
 
     std::vector<ELEM> scatter_out (bufsize, (ELEM)-1);
@@ -622,7 +622,7 @@ test_gatherscatter()
         OIIO_CHECK_EQUAL (scatter_out[i], ((i%3) == 1 && (i&1) ? i/3 : -1));
 
     benchmark ("gather", [&](const ELEM *d){ VEC v; v.gather (d, indices); return v; }, gather_source.data());
-    benchmark ("gather_mask", [&](const ELEM *d){ VEC v; v.gather_mask (mask, d, indices); return v; }, gather_source.data());
+    benchmark ("gather_mask", [&](const ELEM *d){ VEC v = ELEM(0); v.gather_mask (mask, d, indices); return v; }, gather_source.data());
     benchmark ("scatter", [&](ELEM *d){ g.scatter (d, indices); return g; }, scatter_out.data());
     benchmark ("scatter_mask", [&](ELEM *d){ g.scatter_mask (mask, d, indices); return g; }, scatter_out.data());
 }