CAD emulated float64 #114
CAD emulated float64 #114
Conversation
…Examples-and-Tests into emulated_float64_t
…Examples-and-Tests into emulated_float64_t
| const emulated::float64_t a = _static_cast<emulated::float64_t>(6.9f); | ||
| const emulated::float64_t b = _static_cast<emulated::float64_t>(4.5f); | ||
|
|
||
| vk::RawBufferStore<emulated::float64_t::storage_t>(0,(a*b).data); |
There was a problem hiding this comment.
umm this is unsafe, you're writing to gpu virtual address of 0u, you might crash. you create a buffer and pass the address via push constant
There was a problem hiding this comment.
it kinda looks like a file someone forgot to remove (no chanes in 03 main.cpp)
62_CAD/vertex_shader.hlsl
Outdated
| @@ -7,6 +7,11 @@ | |||
| #include <nbl/builtin/hlsl/algorithm.hlsl> | |||
| #include <nbl/builtin/hlsl/jit/device_capabilities.hlsl> | |||
|
|
|||
| using ActualFloat64 = portable_float64_t<jit::device_capabilities>; | |||
There was a problem hiding this comment.
can you add a mode/macro "ForceEmulatedFloat64" for us to test on GPUs that already have Float64 feature?
or alternatively tell me how you usually test emulated vs non-emulated rapidly?
There was a problem hiding this comment.
// force emulated_float64<true, true>
using portable_float64_t = typename conditional<device_capabilities_traits<device_caps>::shaderFloat64, emulated_float64_t<true, true>, emulated_float64_t<true, true> >::type;
// emulated_float64<true, true> only when shaderFloat64 not supported
using portable_float64_t = typename conditional<device_capabilities_traits<device_caps>::shaderFloat64, float64_t, emulated_float64_t<true, true> >::type;
I actually forgot to change that back, so emulated float is forced
There was a problem hiding this comment.
then I think you can do this ForceEmulatedFloat64 option by using a fake caps struct
There was a problem hiding this comment.
#ifdef NBL_FORCE_EMULATED_FLOAT_64
using portable_float64_t = emulated_float64_t<true, true>;
#else
using portable_float64_t = typename conditional<device_capabilities_traits<device_caps>::shaderFloat64, float64_t, emulated_float64_t<true, true> >::type;
#endifdid this, so in order to force ef64 one has to define NBL_FORCE_EMULATED_FLOAT_64 macro before including portable/float64.hlsl
62_CAD/main.cpp
Outdated
| @@ -68,7 +68,7 @@ constexpr std::array<float, (uint32_t)ExampleMode::CASE_COUNT> cameraExtents = | |||
| 600.0, // CASE_8 | |||
| }; | |||
|
|
|||
| constexpr ExampleMode mode = ExampleMode::CASE_7; | |||
There was a problem hiding this comment.
revert these changes please, I changed shader compilation a lot, it will introduce conflicts
…g/Nabla-Examples-and-Tests into emulated_float64_t
# Conflicts: # 62_CAD/main.cpp # 62_CAD/shaders/main_pipeline/common.hlsl # 62_CAD/shaders/main_pipeline/vertex_shader.hlsl
| // TODO[Przemok]: Fix/Remove your todo here?! | ||
| float2 transformPointScreenSpace(pfloat64_t3x3 transformation, uint32_t2 resolution, pfloat64_t2 point2d) | ||
| { | ||
| float64_t2 ndc = transformPointNdc(transformation, point2d); | ||
| return (float32_t2)((ndc + 1.0) * 0.5 * resolution); | ||
| pfloat64_t2 ndc = transformPointNdc(transformation, point2d); | ||
| // TODO: "ndc + 1" doesn't work, probably because of `emulated_float64_t::create(float val)` | ||
| pfloat64_t2 result = (ndc + 1.0f) * 0.5f * _static_cast<pfloat64_t2>(resolution); |
There was a problem hiding this comment.
- int shouldn't work, I don't like it
and still because HLSL doesn't do conversion operators you're unable to get an implicit conversion to satisfy an argument type.
In C++ double+int compiles because double has an implicit constructor from int and only operator+(const double)
NOT because an overload of + exists for every other type (float, int, etc.) as that would wreck mayhem everywhere else (every function that expects a double would throw a fit if you try to feed an integer into it.
What I want to avoid is double+int working and then people asking why f(double) doesn't work when you call with f(int)
| // for initial seed of 69, the polynomial is: | ||
| // f(x) = x^15 * 187.804 + x^14 * 11.6964 + x^13 * 2450.9 + x^12 * 88.6756 + x^11 * 3.62408 + x^10 * 11.4605 + x^9 * 53276.3 + x^8 * 16045.4 + x^7 * 2260.61 + x^6 * 8162.57 | ||
| // + x^5 * 20.674 + x^4 * 13918.6 + x^3 * 2.36093 + x^2 * 8.72536 + x^1 * 2335.63 + 176.719 | ||
| // f(1) = 98961.74987 | ||
| // int from 0 to 69 = 3.11133�10^30 | ||
|
|
||
| template<typename F64> | ||
| struct Random16thPolynomial | ||
| { | ||
| void randomizeCoefficients() | ||
| { | ||
| Xoroshiro64Star rng = Xoroshiro64Star::construct(69); | ||
|
|
||
| // can't just do `coefficients[i] = rng()` this will create retarded numbers or special value exponents |
There was a problem hiding this comment.
language and innuendo son.
Get rid of the 69 and "retarded"
There was a problem hiding this comment.
P.S. I'm sure the numbers would be "fine" if you wiped the position 62-60 bits to 0
|
Has my seal of approval |
…phics-Programming/Nabla-Examples-and-Tests into emulated_float64_t
…Examples-and-Tests into cpp_compat_intrinsics_refactor
…phics-Programming/Nabla-Examples-and-Tests into emulated_float64_t
Devsh-Graphics-Programming/Nabla#696