SpirvShader: Optimize SIMD loads of static, equal offsets, in-bound pointers
Bug: b/135609394
Change-Id: Ic566f2ef6d66e31b434d29b23aafd954a05958a4
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33709
Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
diff --git a/src/Pipeline/SpirvShader.cpp b/src/Pipeline/SpirvShader.cpp
index 316af3d..cf46147 100644
--- a/src/Pipeline/SpirvShader.cpp
+++ b/src/Pipeline/SpirvShader.cpp
@@ -291,22 +291,29 @@
{
using EL = typename Element<T>::type;
- if (ptr.hasStaticSequentialOffsets(sizeof(float)) &&
- ptr.isStaticAllInBounds(sizeof(float)))
+ if (ptr.isStaticAllInBounds(sizeof(float)))
{
- // All elements sequential and in bounds.
- // Perform regular load.
- auto load = rr::Load(rr::Pointer<SIMD::Int>(ptr.base + ptr.staticOffsets[0]), alignment, atomic, order);
- return As<T>(load & mask); // TODO: Mask here should be unnecessary, but keeps with MaskedLoad and Gather.
+ // All elements are statically known to be in-bounds.
+ // We can avoid costly conditional on masks.
+
+ if (ptr.hasStaticSequentialOffsets(sizeof(float)))
+ {
+ // Offsets are sequential. Perform regular load.
+ return rr::Load(rr::Pointer<T>(ptr.base + ptr.staticOffsets[0]), alignment, atomic, order);
+ }
+ if (ptr.hasStaticEqualOffsets())
+ {
+ // Load one, replicate.
+ return T(*rr::Pointer<EL>(ptr.base + ptr.staticOffsets[0], alignment));
+ }
}
-
- auto offsets = ptr.offsets();
-
- if(robust) // Disable OOB reads.
+ else if(robust) // Disable OOB reads.
{
mask &= ptr.isInBounds(sizeof(float));
}
+ auto offsets = ptr.offsets();
+
if (!atomic && order == std::memory_order_relaxed)
{
if (ptr.hasStaticEqualOffsets())
