Added floating point types support for work_group_scan_exclusive tests#2533
Added floating point types support for work_group_scan_exclusive tests#2533shajder wants to merge 4 commits into
Conversation
-adapted from KhronosGroup#2525 due to code review
| constexpr cl_half g_half_min = 0xFC00; | ||
| constexpr cl_half g_half_max = 0x7C00; |
There was a problem hiding this comment.
Why do do not keep the maximum finite values?
There was a problem hiding this comment.
sorry, I've implemented that sep 2025, I don't recall the details right now and I will have to dive deeper. I will get back later.
| static void generate_input_values(Type *inptr, size_t n_elems, | ||
| size_t max_wg_size, | ||
| const Type *max_err = nullptr) | ||
| Type *const max_err = nullptr) |
There was a problem hiding this comment.
I understand why we want to add a const, but no need to remove the one already there, right?
There was a problem hiding this comment.
The two consts aren't the same: the old const Type * protects the pointee, but max_err is now written to (it's an out-param), so that const has to go; the new Type *const only locks the pointer, not the data.
| // to prevent overflow limit range of randomization | ||
| float max_range = 99.0; | ||
| float min_range = -99.0; | ||
| // generate reference values for one work group | ||
| for (size_t j = 0; j < max_wg_size; j++) | ||
| ref_vals[j] = cl_half_from_float( | ||
| get_random_float(min_range, max_range, d), | ||
| gHalfRoundingMode); |
There was a problem hiding this comment.
Why do we need to select in that specific range? we do not have any range for integers.
There was a problem hiding this comment.
Integer add wraps deterministically and is order-independent, so reference and device match even on overflow, no range needed. FP isn't: scan order is unspecified, and once a partial sum saturates to positive or negative INF the value is lost and INF (or INF-INF -> NaN) can appear at different positions on device vs reference, causing false mismatches. The range just keeps every partial sum finite.
| // populate reference data across all work groups | ||
| for (size_t i = 0; i < (size_t)n_elems; i += max_wg_size) | ||
| { | ||
| size_t wg_size = std::min(max_wg_size, n_elems - i); | ||
| memcpy(&inptr[i], ref_vals.data(), sizeof(Type) * wg_size); | ||
| } |
There was a problem hiding this comment.
Why do we duplicate the data? we generate random values for integers for each workgroup it seems.
3 participants
Related to #1364