Broken real time lights on ZX C-1190 (integrated graphics from Zhaoxin KX-7000)

[illwieckz]

Broken real time lights on ZX C-1190 (integrated graphics from Zhaoxin KX-7000).

This integrated GPU is supposedly an S3 derivative which is not surprising given Zhaoxin CPUs are themselves derivatives of Via CPUs.

With the medium preset, the render is buggy:

As soon as I set r_realtimeLighting off, the bug disappears:

[illwieckz]

Some gfxinfo excerpt:

GL_VENDOR: Shanghai Zhaoxin Semiconductor Co., Ltd.
GL_RENDERER: ZX C-1190
GL_VERSION: 4.6 Mesa 26.00.48
GL_MAX_TEXTURE_SIZE: 16384
GL_MAX_3D_TEXTURE_SIZE: 2048
GL_MAX_CUBE_MAP_TEXTURE_SIZE: 16384
GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS: 80
GL_SHADING_LANGUAGE_VERSION: 4.60
GL_MAX_VERTEX_UNIFORM_COMPONENTS 1024
GL_MAX_VERTEX_ATTRIBS 16
GL_MAX_DRAW_BUFFERS: 8
GL_TEXTURE_MAX_ANISOTROPY_EXT: 16.000000
GL_MAX_COLOR_ATTACHMENTS: 8
PIXELFORMAT: color(24-bits)
MODE: -2, 1920 x 1080 windowed
Using OpenGL version 4.6, requested: 3.2
Using an OpenGL core profile.
OpenGL context is not forward compatible.
Using OpenGL extensions: GL_ARB_half_float_pixel GL_ARB_texture_float GL_EXT_gpu_shader4 GL_ARB_gpu_shader5 GL_EXT_texture_integer GL_ARB_texture_rg GL_EXT_texture_compression_s3tc GL_ARB_texture_compression_rgtc GL_EXT_texture_filter_anisotropic GL_ARB_vertex_array_object GL_ARB_vertex_buffer_object GL_ARB_half_float_vertex GL_EXT_framebuffer_object GL_EXT_framebuffer_blit GL_ARB_buffer_storage GL_ARB_uniform_buffer_object GL_ARB_map_buffer_range GL_ARB_sampler_objects GL_ARB_sync GL_ARB_texture_barrier GL_ARB_compute_shader GL_ARB_shader_draw_parameters GL_ARB_program_interface_query GL_ARB_multi_draw_indirect GL_ARB_shader_image_load_store GL_ARB_shader_atomic_counters GL_ARB_shader_atomic_counter_ops GL_ARB_indirect_parameters GL_ARB_direct_state_access GL_ARB_vertex_attrib_binding
Missing OpenGL extensions: GL_ARB_debug_output GL_ARB_texture_gather GL_ARB_get_program_binary GL_ARB_bindless_texture GL_ARB_shader_storage_buffer_object GL_ARB_shading_language_420pack GL_ARB_explicit_uniform_location GL_KHR_shader_subgroup
Using half-float vertex format.
Using S3TC (DXTC) texture compression.
Using GPU vertex skinning with max 41 bones in a single pass, some models may not be hardware accelerated.
Using dual processor acceleration.
Using OpenMP parallelism with 4 threads.

[illwieckz]

Just for our own knowledge, this GPU also benefits from OpenMP accelerated models:

Using GPU vertex skinning with max 41 bones in a single pass, some models may not be hardware accelerated.
Using dual processor acceleration.
Using OpenMP parallelism with 4 threads.

With common.ompThreads 0:

With common.ompThreads 1:

[illwieckz]

More knowledge about this GPU…

For unknown reasons, it claims to implement OpenGL 4.6, but also don't implement ARB_framebuffer_object like if it was a 2.1 device… ⁉️

Missing ARB_framebuffer_object, using EXT_framebuffer_object with EXT_framebuffer_blit instead.

Edit: No, I had a stray r_arb_framebuffer_object 0.

The engine also prints some of those warnings from time to time:

Warn: Long wait for dynamic GL buffer: active segment: 3, timeout: 10000000ns

Edit: I got this with Wayland, with some stutters (despite the framerate display not reflecting it), it runs better on Xorg.

[illwieckz]

With #1968 the game runs with the default preset (medium) and it is playable:

• Detect Zhaoxin GPU, disable realtime lighting on it #1968

[illwieckz]

According to Wikipedia, the C-1190 is indeed an S3 Chrome graphic solution:

Zhaoxin KX-7000 series contains Chrome 1190 (ZX C-1190) engine for their integrated graphics.
-- https://en.wikipedia.org/wiki/S3_Chrome

[illwieckz]

Just for knowledge, an earlier model of this GPU is the C1080 which is known to not only be integrated in some earlier Zhaoxin GPUs like the KX600G, but also to be sold in a discrete card by Glenfly:

So coming back to the Zhaoxin KX-6000G CPUs which we may as well refer to as APUs since they utilize the same Zhaoxin-made ZX C1080 GPU that is featured on the Glenfly Arise-GT-10C0 dGPU that was launched in the Chinese domestic market just a few days ago.
-- https://wccftech.com/chinese-zhaoxin-kx-6000g-cpu-gt10c0-integrated-gpu-features-same-performance-as-nvidia-gt-630/

I have no idea how are reported the Glenfly GPUs, but maybe we would need to detect those cards and implement similar workarounds.

[illwieckz]

According to this page Glenfly is a subsidiary of Zhaoxin, and the screenshots shows that Glenfly also reports their GL vendor as Shanghai Zhaoxin Semiconductor Co., Ltd. so my patch would detect them as well.

[slipher]

It does that without even having rendered a single light? That's strange since enabling real-time lights changes so little as long as none are actually rendered.

Could it be due to violating the GL_MAX_UNIFORM_BLOCK_SIZE limit? The lights uniform array is 48K, and I found a Glenfly report with a limit of just 32K. If that's the case we can fix it by testing the uniform block size limit against sizeof(shaderLight_t) * MAX_REF_LIGHTS. (Or better, by adjusting the max number of lights down based on the GL limit.)

BTW it would be neat if when testing lesser-known GPUs you generated a https://opengl.gpuinfo.org/ report.

[slipher]

I took a stab at this: #1969

[illwieckz]

It does that without even having rendered a single light? That's strange since enabling real-time lights changes so little as long as none are actually rendered.

Yes. It's like if the whole lightMapping shader becomes invalid. It's black because it doesn't render anything, the forcefield renders correctly, but since it blends with background and there is no background, it blends with previous frames of itself and accumulate. Moving camera does the same as with the void.

Could it be due to violating the GL_MAX_UNIFORM_BLOCK_SIZE limit?

No idea, I'll try your patch.

and I found a Glenfly report with a limit of just 32K

Oh that's annoying they use a different GL_VENDOR there…

BTW it would be neat if when testing lesser-known GPUs you generated a https://opengl.gpuinfo.org/ report.

I'll do.

[illwieckz]

BTW it would be neat if when testing lesser-known GPUs you generated a https://opengl.gpuinfo.org/ report.

I'll do.

Done:

• https://opengl.gpuinfo.org/displayreport.php?id=16753
  ZX C-960, Zhaoxin KaiXian KX-U6780A, zx driver
• https://opengl.gpuinfo.org/displayreport.php?id=16754
  ZX C-1190, Zhaoxin KaiXian KX-7000, cx4 driver

[illwieckz]

I took a stab at this: #1969

It doesn't fix it, unfortunately. Anyway, it really looks like the whole lightMapping shader is skipped, probably because of being invalidated at build time because of a silent error.

• realtime lights, no clear:

• realtime lights, clear:

• no realtime lights, no clear: