GS/HW: Reduce number of copies for HDR

[refractionpcsx2]

THIS WILL ONLY AFFECT SPECIFIC GAMES!

Description of Changes

Tries to reduce the number of copies for HDR draws.

Rationale behind Changes

Master would copy the render target to a new texture, draw, then copy back, every single time, but a lot of games do a lot of separate draws in a row to the same target, so it's better to keep it in one place until the end of the chain. This PR will keep it HDR as long as possible before converting back, hugely reducing copies and render passes in some cases.

TL;DR: Makes some games go brr more

Suggested Testing Steps

Test the games listed below (sorry for the names)

Slightly Cherry Picked Reduction Results from Vulkan:

Ace Combat - Squadron Leader_SCES-52424 ['Draw Calls: -1 [620=>619]', 'Render Passes: -1 [44=>43]', 'Copies: -1 [8=>7]']
Big.Mutha.Truckers ['Draw Calls: -352 [903=>551]', 'Render Passes: -423 [440=>17]', 'Barriers: -127 [132=>5]', 'Copies: -352 [355=>3]']
DT_Carnage_SLUS-21793_20230908174547 ['Draw Calls: -29 [936=>907]', 'Render Passes: -32 [344=>312]', 'Copies: -30 [82=>52]']
Echo_Night_-_Beyond_SLUS-20928_20221019163851 ['Draw Calls: -2745 [6965=>4220]', 'Render Passes: -3384 [3456=>72]', 'Barriers: -1104 [1133=>29]', 'Copies: -2745 [2787=>42]']
evolution_colcliphdr ['Draw Calls: -142 [464=>322]', 'Render Passes: -179 [195=>16]', 'Barriers: -46 [50=>4]', 'Copies: -143 [149=>6]']
Full spectrum warrior ['Draw Calls: -3 [189=>186]', 'Render Passes: -3 [23=>20]', 'Copies: -3 [8=>5]']
JonnyMoseleydebug ['Draw Calls: -115 [211=>96]', 'Render Passes: -116 [128=>12]', 'Copies: -116 [118=>2]']
Kunoichi ['Draw Calls: -48 [1156=>1108]', 'Render Passes: -53 [168=>115]', 'Barriers: -1 [1=>0]', 'Copies: -48 [146=>98]']
Malice_SLES-52413_20240617171243 ['Draw Calls: -4 [453=>449]', 'Render Passes: -4 [232=>228]', 'Copies: -4 [23=>19]']
Mercenaries_-_Playground_of_Destruction_SLUS-20932 ['Draw Calls: -22 [780=>758]', 'Render Passes: -22 [86=>64]', 'Copies: -22 [51=>29]']
Nightshade_SLUS-20810_20231001162431 ['Draw Calls: -42 [1117=>1075]', 'Render Passes: -44 [131=>87]', 'Copies: -42 [119=>77]']
Pac-Man World Rally_SLUS-21328_20230127222927 ['Draw Calls: -57 [189=>132]', 'Render Passes: -64 [78=>14]', 'Barriers: -4 [4=>0]', 'Copies: -59 [64=>5]']
Pac-Man_World_2_SLUS-20224_20220816132032 ['Draw Calls: -2 [82=>80]', 'Render Passes: -2 [14=>12]', 'Copies: -2 [3=>1]']
sakura taisen shadows 1 (hw) ['Draw Calls: -96 [349=>253]', 'Render Passes: -111 [138=>27]', 'Barriers: -7 [7=>0]', 'Copies: -96 [116=>20]']
Sly 2 - Band of Thieves_SCUS-97316_20230422201628 ['Draw Calls: -173 [5962=>5789]', 'Render Passes: -152 [214=>62]', 'Barriers: -21 [546=>525]', 'Copies: -173 [178=>5]']
Snoopy vs Red Baron (snoopy) ['Draw Calls: -40 [198=>158]', 'Render Passes: -44 [60=>16]', 'Barriers: -2 [2=>0]', 'Copies: -40 [47=>7]']
Soul_Calibur_III_SLUS-21216_20231007100145 ['Draw Calls: -138 [526=>388]', 'Render Passes: -138 [156=>18]', 'Copies: -138 [143=>5]']
State_of_Emergency_2_SLUS-20966_20230921232225 ['Draw Calls: -2 [484=>482]', 'Render Passes: -2 [28=>26]', 'Copies: -2 [6=>4]']
Superman Returns - The Video Game_SLUS-21434_20221208170947 ['Render Passes: -1 [9=>8]']
Tekken_5_SLUS-21059_20240108195029 ['Draw Calls: -168 [641=>473]', 'Render Passes: -194 [244=>50]', 'Barriers: -45 [45=>0]', 'Copies: -168 [175=>7]']
tom_jerry_clear ['Draw Calls: -2 [93=>91]', 'Render Passes: -2 [13=>11]', 'Copies: -2 [3=>1]']
Virtua Quest _NTSC-U__SLUS-20977 ['Draw Calls: -2 [444=>442]', 'Render Passes: -2 [11=>9]', 'Copies: -2 [4=>2]']

[JordanTheToaster]

Some benchmarks Sly 2 at 12x gains 30.3% Echo Night Beyond at 6x gains 333.6% Big Mutha Trucka at 6x gains 294.7%.

[crashGG]

OS：Win10 22H2
HW: AMD 7840HS with 780M Core Graphics (RDNA 3)
Test game: Soulcalibur III SLUS-21216
Test scenario: Museum - Character Profiles - XINGHUA
A relatively fixed scene, the heaviest load in the entire game, with the most HDR highlights and the lowest frame rate.

Use 5x internal resolution to test GPU full load

Conclusion: The most significant efficiency improvement in the past two years.

[refractionpcsx2]

Awesome, thanks for the tests @crashGG !

Unfortunately the performance uplift is limited to a few games, but it's nice where it is :) HDR/COLCLIP has been a bit of a sticking point for performance over the years.

[VGkav]

Nice! By the way, does this affect native resolution? Also, does this affect software rendering?

[lightningterror]

Nice! By the way, does this affect native resolution? Also, does this affect software rendering?

It affects all resolutions, only hardware renderers.

[refractionpcsx2]

It will depend on what you're limited by, however.

if you're GPU limited, it will have the most significant effect, if you're CPU limited (on the GS thread) then you might notice a small performance increase.. But as mentioned it will be limited to what games are affected, mainly the ones listed, but there might be some other games affected, I just don't have GS dumps of them :)

[crashGG]

This is why I do not agree with this PR. According to the test data, if you want to experience high rendering (4K) under limited GPU conditions (core graphics), the best choice for RDNA3 GPU is still DX11.

[refractionpcsx2]

It's not really relevant to this PR, so please keep the discussion on-topic.

[TheTechnician27]

OS: Linux 6.12; KDE 6.2.5; Manjaro
CPU: Ryzen 7 5800X
GPU: GTX 1070
Rendering: Vulkan; Basic blend; Native res; default
Game: Sly 3
Benchmark: First mission from start to finish (not including binocucomm cutscene) taken across two seconds

Master:
0.1%: 36.4
1%: 83.1
97th percentile: 615.1
Average: 170
GPU Load: 30
CPU Load: 19.9
Avg Frame Time: 5.9

PR:
0.1%: 86.5
1%: 108.1
97th percentile: 407.4
Average: 197.8
GPU Load: 32
CPU Load: 21.7
Avg Frame Time: 5.1

NOTE: Seemingly because PR runs faster than master and because I was using F4, Sly got to the unwinnable boss fight at the end of the level faster, and so I had to run around in the same area for longer than on master, thus potentially skewing the results somewhat. But Sly getting to the boss fight so much faster was probably due in large part to the faster framerate (partially to getting better at completing the level in fast-forward).

[JordanTheToaster]

Some DX11 and OpenGL tests on a 2070 Super.

[kamfretoz]

Moar benchmarks, this time on team red!

CPU: Ryzen 7 5800X
RAM: 32GB
GPU: Radeon RX 6600 XT

Tested with Vulkan at 6x Internal Resolution and Basic Blending