Made load a new format callable #2106

SadiinsoSnowfall · 2025-05-28T09:40:50Z

No description provided.

DenisYaroshevskiy

algo changes look OK. THe load itself - didn't look yet.

test/unit/memory/load/tuple.cpp

SadiinsoSnowfall · 2025-06-05T09:01:37Z

Notes:
I moved everything to detail/function. A part of the x86 and generic impls previously were in core/regular.

There's now the following functions :

load_ (iterator flavour): handles the iterators directly, special case for logical values with !is_wide_logical ABIs.
load_ (normal): entry point for the callable impl, handles things that should always be done first: unsafe and algorithm-based loads, as well as handling bundle values. Once we have a safe non-bundle non-alg load, load_common is called.
load_common: the "main" function which handles the following cases:
- if we are in aggregated or emulated ABI:
  - if there is no condition (ignore_none): call aggregate_load or piecewise_load
  - if there is: call load_cx_
- if there is a backend that can handle the operation with a condition directly, call it (load_impl(api, cx, src, tgt))
- if there is no condition, call the non-conditional form of the backend (load_impl(api, src, tgt)). Note: this must work or else it would mean that no load backends are defined for the current architecture.
- if there is, call load_cx_
load_cx_: handle the conditions when not handled by the architecture-specific backends. Special-cased for logical values with !is_wide_logical ABIs.

load_common takes an api parameter so that architecture-specific backends are able to "bail out" and, for example, ask for load_cx_ to handle specific forms of conditional load.

A generic load_impl for logical values is provided. This is because most backends either do not handle logical loads or handle them in a similar manner. This function is special-cased for values with !is_wide_logical ABIs.

DenisYaroshevskiy · 2025-06-06T10:08:33Z

include/eve/detail/function/load.hpp


-  template<> struct supports_optimized_conversion<tag::load_> : std::true_type {};
+    template<detail::data_source Ptr, std::ptrdiff_t N>


concepts using detail is ofc not awesome, but this is not for you.

DenisYaroshevskiy · 2025-06-06T10:15:39Z

include/eve/detail/function/load.hpp

+  //================================================================================================
+  //! @addtogroup simd
+  //! @{
+  //!   @var load


@jfalcou - docs in detail. Will they be rendered?

We need to move it to the non detail place like for bit_cast.

DenisYaroshevskiy · 2025-06-06T10:16:24Z

include/eve/module/core/regular/load.hpp

@@ -6,95 +6,4 @@
 //==================================================================================================
 #pragma once

-#include <eve/arch.hpp>


Maybe actual load should be in this file instead of detail.

It is this way cause of our adage of having the basics of wide, etc, not depending on the module/core. When everything settles on this front, we can revisit that. Also, the doc should be here.

include/eve/detail/function/simd/common/load.hpp

DenisYaroshevskiy · 2025-06-06T10:20:14Z

include/eve/detail/function/simd/common/load.hpp

+    }
+    else
+    {
+      return piecewise_load(src, tgt);


For my sanity - where is the assert that there is enough space?

You mean a check to ensure that every element in the [b, e) range fits in the logical register ? There was previously no checks for that. I also suspect that the second iterator was only used to differentiate the implementation for the standard one as it has never been used, or so it seems.

DenisYaroshevskiy · 2025-06-06T10:21:41Z

include/eve/detail/function/simd/common/load.hpp

  {
-    return load(ignore_none, safe, as<Wide>{}, b, e);
+    if constexpr (logical_simd_value<Wide> && !Wide::abi_type::is_wide_logical)


There is some confusion here.

This code is duplicated between this overload and the next overload.
Should be unified for sure.

Not really, the iterator path always end up calling piecewise_load whereas the logical path could have arch-specific backends.

DenisYaroshevskiy · 2025-06-06T10:22:59Z

include/eve/detail/function/simd/common/load.hpp

-    return aggregate_load(tgt,ptr);
+    if constexpr (!std::is_pointer_v<DS>)
+    {
+      constexpr bool is_aligned_enough = c_t() * sizeof(e_t) >= DS::alignment();


I'm so confused to what;'s happening.

DS is data_source. We use that in the API of load. Here we call ::alignmetn on it. We obv can't do that for pointers. What's happening.

well there's a !std::is_pointer_v<DS> check right before the call to alignment

SadiinsoSnowfall force-pushed the callable/load branch 4 times, most recently from a5ffc3c to c71d6fc Compare June 4, 2025 09:13

DenisYaroshevskiy reviewed Jun 4, 2025

View reviewed changes

test/unit/memory/load/tuple.cpp Show resolved Hide resolved

SadiinsoSnowfall marked this pull request as ready for review June 4, 2025 10:45

SadiinsoSnowfall force-pushed the callable/load branch from 9777c53 to 7e7c693 Compare June 5, 2025 11:56

DenisYaroshevskiy reviewed Jun 6, 2025

View reviewed changes

SadiinsoSnowfall force-pushed the callable/load branch 2 times, most recently from fe19266 to 975a766 Compare June 16, 2025 11:12

SadiinsoSnowfall added 13 commits June 16, 2025 13:33

stash

563d243

finished x86 refactor

a723c00

[no ci] stash

649c8eb

CI test

7aea14a

algo & x86 fixes

61e95dd

fix

e32549b

fix

6d9c015

move non-wide-logical logic

a2b8e1b

do not use unsafe2 where not required

2d65b64

updated doc

4ee0f39

review fixes

58c7ece

move documentation to core/regular

d12120f

fixed warning

4b6877f

SadiinsoSnowfall force-pushed the callable/load branch from e9e94be to 4b6877f Compare June 16, 2025 11:33


		template<> struct supports_optimized_conversion<tag::load_> : std::true_type {};
		template<detail::data_source Ptr, std::ptrdiff_t N>

Made load a new format callable #2106

Are you sure you want to change the base?

Made load a new format callable #2106

Conversation

SadiinsoSnowfall commented May 28, 2025

Uh oh!

DenisYaroshevskiy left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

SadiinsoSnowfall commented Jun 5, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!