vulkan: enable fp16 for gcn 3 and 4 chips

[netrunnereve]

Basically GCN 3 and 4 chips support FP16, but it's unable to process two values at once like GCN 5 can.. Since there's apparently no performance benefit shaderFloat16 is disabled in the drivers, but the chip fully supports it and RADV is able to generate those instructions.

While the actual FMAs won't run any faster having FP16 means that I can use four times less shared memory for mul mat and save a little bit of memory bandwidth when reading the B matrix. As a result I get a little improvement in prompt processing on my RX 470.

PR:

model	size	params	backend	ngl	test	t/s
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	99	pp512	195.34 ± 0.70
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	99	tg128	33.48 ± 0.53

Master:

model	size	params	backend	ngl	test	t/s
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	99	pp512	188.32 ± 0.38
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	99	tg128	33.62 ± 0.28

I'm leaving this as a draft for now as it's a bit hacky and I'm not sure if the proprietary drivers support this. The good thing here though is that it'll let me work on and test the FP16 shaders using my old card.

[0cc4m]

This can only be merged if it passes validation, which I doubt. If it does not, I would require disabling it by default and hiding behind an environment variable. The backend has to follow the Vulkan specification.

[0cc4m]

It's against the spec to enable this if support is not indicated with the shaderFloat16 feature, which it is not for these GPUs. Did you try running this with validation layers enabled? I'm relatively sure this should throw validation issues.

[netrunnereve]

I finally got the validation layers working and it fails with vkCreateComputePipelines(): pCreateInfos[0].stage SPIR-V Capability Float16 was declared, but one of the following requirements is required (VkPhysicalDeviceVulkan12Features::shaderFloat16 OR VK_AMD_gpu_shader_half_float).. So yeah Vulkan isn't happy with how I'm using FP16 when the driver says it doesn't support it.

Considering how small the improvement is I don't think it's worth having a special environment variable and all that, and I'm just going to close this.