jacob/power_play
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

shader printf progress

Browse Source
This commit is contained in:
jacob 2025-12-08 16:33:44 -06:00
parent 859e1c5245
commit 3e722bb982
11 changed files with 285 additions and 105 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/base/base.h
View File

@ -435,8 +435,6 @@
//- countof
#if IsLanguageC
#define countof(a) (sizeof(a) / sizeof((a)[0]))
#elif IsLanguageHlsl
template<typename T, uint N> uint countof(T ints[N]) { return N; }
#endif
//- IsArray

62
src/base/base_gpu.hlsl
View File

@ -24,8 +24,9 @@ typedef float4x4 Mat4x4;
////////////////////////////////////////////////////////////
//~ Handle dereference wrappers
/* NOTE: Non-uniform resource access is assumed as the default behavior */
/* TODO: Add explicit "uniform" variants of handle deref operations for optimization on AMD devices */
/* TODO: Non-uniform resource access is assumed as the default behavior. We
* may want to add explicit "uniform" variants for optimization on AMD devices
* in the future. */
template<typename T> StructuredBuffer<T> StructuredBufferFromHandle(StructuredBufferHandle h) { return ResourceDescriptorHeap[NonUniformResourceIndex(h.v)]; }
template<typename T> RWStructuredBuffer<T> RWStructuredBufferFromHandle(RWStructuredBufferHandle h) { return ResourceDescriptorHeap[NonUniformResourceIndex(h.v)]; }
@ -43,52 +44,19 @@ template<typename T> RWTexture3D<T> RWTexture3DFromHandle(RWTexture3DHandle h) {
SamplerState SamplerStateFromHandle(SamplerStateHandle h) { return SamplerDescriptorHeap[NonUniformResourceIndex(h.v)]; }
////////////////////////////////////////////////////////////
//~ Dimension helpers
//~ Countof
u32 Count1D(Texture1D tex)
{
u32 result;
tex.GetDimensions(result.x);
return result;
}
template<typename T>
u32 Count1D(RWTexture1D<T> tex)
{
u32 result;
tex.GetDimensions(result.x);
return result;
}
Vec2U32 Count2D(Texture2D tex)
{
Vec2U32 result;
tex.GetDimensions(result.x, result.y);
return result;
}
template<typename T>
Vec2U32 Count2D(RWTexture2D<T> tex)
{
Vec2U32 result;
tex.GetDimensions(result.x, result.y);
return result;
}
Vec3U32 Count3D(Texture3D tex)
{
Vec3U32 result;
tex.GetDimensions(result.x, result.y, result.z);
return result;
}
template<typename T>
Vec3U32 Count3D(RWTexture3D<T> tex)
{
Vec3U32 result;
tex.GetDimensions(result.x, result.y, result.z);
return result;
}
template<typename T, u32 N> u32 countof(T arr[N]) { return N; }
template<typename T> u32 countof(StructuredBuffer<T> buff) { u32 result; buff.GetDimensions(result); return result; }
template<typename T> u32 countof(RWStructuredBuffer<T> buff) { u32 result; buff.GetDimensions(result); return result; }
u32 countof(ByteAddressBuffer buff) { u32 result; buff.GetDimensions(result); return result; }
u32 countof(RWByteAddressBuffer buff) { u32 result; buff.GetDimensions(result); return result; }
u32 countof(Texture1D tex) { u32 result; tex.GetDimensions(result); return result; }
template<typename T> u32 countof(RWTexture1D<T> tex) { u32 result; tex.GetDimensions(result); return result; }
Vec2U32 countof(Texture2D tex) { Vec2U32 result; tex.GetDimensions(result.x, result.y); return result; }
template<typename T> Vec2U32 countof(RWTexture2D<T> tex) { Vec2U32 result; tex.GetDimensions(result.x, result.y); return result; }
Vec3U32 countof(Texture3D tex) { Vec3U32 result; tex.GetDimensions(result.x, result.y, result.z); return result; }
template<typename T> Vec3U32 countof(RWTexture3D<T> tex) { Vec3U32 result; tex.GetDimensions(result.x, result.y, result.z); return result; }
////////////////////////////////////////////////////////////
//~ Color helpers

22
src/base/base_string.c
View File

@ -565,52 +565,52 @@ String FormatStringV(Arena *arena, String fmt, va_list args)
no_more_args = 1;
} break;
case FmtKind_Char:
case FmtArgKind_Char:
{
parsed_str = StringFromChar(arena, arg.value.c);
} break;
case FmtKind_String:
case FmtArgKind_String:
{
parsed_str = PushString(arena, arg.value.string);
} break;
case FmtKind_Uint:
case FmtArgKind_Uint:
{
parsed_str = StringFromU64(arena, arg.value.uint, 10, arg.z);
} break;
case FmtKind_Sint:
case FmtArgKind_Sint:
{
parsed_str = StringFromI64(arena, arg.value.sint, 10, arg.z);
} break;
case FmtKind_Hex:
case FmtArgKind_Hex:
{
parsed_str = StringFromU64(arena, arg.value.sint, 16, arg.z);
} break;
case FmtKind_Ptr:
case FmtArgKind_Ptr:
{
parsed_str = StringFromPtr(arena, arg.value.ptr);
} break;
case FmtKind_Float:
case FmtArgKind_Float:
{
parsed_str = StringFromF64(arena, arg.value.f, arg.p);
} break;
case FmtKind_Handle:
case FmtArgKind_Handle:
{
parsed_str = StringFromhandle(arena, arg.value.handle.h64[0], arg.value.handle.h64[1]);
} break;
case FmtKind_Uid:
case FmtArgKind_Uid:
{
parsed_str = StringFromUid(arena, arg.value.uid);
} break;
case FmtKind_End:
case FmtArgKind_End:
{
/* Unexpected end. Not enough FMT args passed to function. */
Assert(0);
@ -636,7 +636,7 @@ String FormatStringV(Arena *arena, String fmt, va_list args)
{
FmtArg last_arg = va_arg(args, FmtArg);
/* End arg not reached. Too many FMT values passed to function. */
Assert(last_arg.kind == FmtKind_End);
Assert(last_arg.kind == FmtArgKind_End);
}
#endif

46
src/base/base_string.h
View File

@ -4,27 +4,27 @@
#define DefaultFmtPrecision 3
#define IntChars ("0123456789abcdef")
Enum(FmtKind)
Enum(FmtArgKind)
{
FmtKind_None,
FmtArgKind_None,
/* Arbitrary magic numbers for argument validation */
FmtKind_Char = 0x0f5281df,
FmtKind_String = 0x0a5ffa9a,
FmtKind_Uint = 0x0746f19b,
FmtKind_Sint = 0x08603694,
FmtKind_Hex = 0x0a3d0792,
FmtKind_Ptr = 0x0c4519e4,
FmtKind_Float = 0x04814143,
FmtKind_Uid = 0x3d1cd407,
FmtKind_Handle = 0x6ead3bec,
FmtArgKind_Char = 0x0f5281df,
FmtArgKind_String = 0x0a5ffa9a,
FmtArgKind_Uint = 0x0746f19b,
FmtArgKind_Sint = 0x08603694,
FmtArgKind_Hex = 0x0a3d0792,
FmtArgKind_Ptr = 0x0c4519e4,
FmtArgKind_Float = 0x04814143,
FmtArgKind_Uid = 0x3d1cd407,
FmtArgKind_Handle = 0x6ead3bec,
FmtKind_End = 0x0ecbc5ae
FmtArgKind_End = 0x0ecbc5ae
};
Struct(FmtArg)
{
FmtKind kind;
FmtArgKind kind;
u32 p; /* Precision */
u32 z; /* Z-fill */
union
@ -100,16 +100,16 @@ String StringFromList(Arena *arena, StringList l, String separator);
#define FMTARG(_kind, ...) ((FmtArg) { .kind = (_kind), .p = DefaultFmtPrecision, __VA_ARGS__ })
#define FmtChar(v, ...) FMTARG(FmtKind_Char, .value.c = (v), __VA_ARGS__)
#define FmtString(v, ...) FMTARG(FmtKind_String, .value.string = (v), __VA_ARGS__)
#define FmtUint(v, ...) FMTARG(FmtKind_Uint, .value.uint = (v), __VA_ARGS__)
#define FmtSint(v, ...) FMTARG(FmtKind_Sint, .value.sint = (v), __VA_ARGS__)
#define FmtHex(v, ...) FMTARG(FmtKind_Hex, .value.uint = (v), __VA_ARGS__)
#define FmtPtr(v, ...) FMTARG(FmtKind_Ptr, .value.ptr = (v), __VA_ARGS__)
#define FmtFloat(v, ...) FMTARG(FmtKind_Float, .value.f = (v), __VA_ARGS__)
#define FmtHandle(v, ...) FMTARG(FmtKind_Handle, .value.handle.h64[0] = (v).idx, .value.handle.h64[1] = (v).gen, __VA_ARGS__)
#define FmtUid(v, ...) FMTARG(FmtKind_Uid, .value.uid = (v), __VA_ARGS__)
#define FmtEnd FMTARG(FmtKind_End) /* Denotes end of VA list */
#define FmtChar(v, ...) FMTARG(FmtArgKind_Char, .value.c = (v), __VA_ARGS__)
#define FmtString(v, ...) FMTARG(FmtArgKind_String, .value.string = (v), __VA_ARGS__)
#define FmtUint(v, ...) FMTARG(FmtArgKind_Uint, .value.uint = (v), __VA_ARGS__)
#define FmtSint(v, ...) FMTARG(FmtArgKind_Sint, .value.sint = (v), __VA_ARGS__)
#define FmtHex(v, ...) FMTARG(FmtArgKind_Hex, .value.uint = (v), __VA_ARGS__)
#define FmtPtr(v, ...) FMTARG(FmtArgKind_Ptr, .value.ptr = (v), __VA_ARGS__)
#define FmtFloat(v, ...) FMTARG(FmtArgKind_Float, .value.f = (v), __VA_ARGS__)
#define FmtHandle(v, ...) FMTARG(FmtArgKind_Handle, .value.handle.h64[0] = (v).idx, .value.handle.h64[1] = (v).gen, __VA_ARGS__)
#define FmtUid(v, ...) FMTARG(FmtArgKind_Uid, .value.uid = (v), __VA_ARGS__)
#define FmtEnd FMTARG(FmtArgKind_End) /* Denotes end of VA list */
#define StringF(arena, lit, ...) FormatString_((arena), Lit(lit), __VA_ARGS__, FmtEnd)
#define FormatString(arena, fmt, ...) FormatString_((arena), (fmt), __VA_ARGS__, FmtEnd)

2
src/gpu/gpu.lay
View File

@ -5,7 +5,9 @@
//- Api
@IncludeC gpu_core.h
@IncludeC gpu_shader_extras.h
@IncludeC gpu_extras.h
@IncludeGpu gpu_shader_extras.h
//- Impl
@IncludeC gpu_extras.c

12
src/gpu/gpu_shader_extras.h Normal file
View File

@ -0,0 +1,12 @@
////////////////////////////////////////////////////////////
//~ Static handles (common resources available to all shaders)
#if IsLanguageC
#define GPU_StaticHandle(type, v) ((type) { (v) })
#elif IsLanguageHlsl
#define GPU_StaticHandle(type, v) (type(v))
#endif
#define GPU_ShaderPrintBufferHandle RWByteAddressBufferHandle (1)
#define GPU_BasicPointSamplerHandle SamplerStateHandle (2)
#define GPU_BasicNoiseTextureHandle Texture2DHandle (3)

222
src/gpu/gpu_shader_extras.hlsl
View File

@ -1,14 +1,220 @@
////////////////////////////////////////////////////////////
//~ Shader printf types
////////////////////////////////////////////////////////////
//~ Shader printf
// #define Test(fmt_cstr, ...) do { \
// } while (0)
/* This technique comes from MJP's article:
* https://therealmjp.github.io/posts/hlsl-printf/
*/
void Test(u32 count)
template<typename T> u32 U32FromChar(in T c)
{
if(c == ' ')
return 32;
if(c == '!')
return 33;
if(c == '\"' || c == '\"')
return 34;
if(c == '#')
return 35;
if(c == '$')
return 36;
if(c == '%')
return 37;
if(c == '&')
return 38;
if(c == '\'')
return 39;
if(c == '(')
return 40;
if(c == ')')
return 41;
if(c == '*')
return 42;
if(c == '+')
return 43;
if(c == ',')
return 44;
if(c == '-')
return 45;
if(c == '.')
return 46;
if(c == '/')
return 47;
if(c == '0')
return 48;
if(c == '1')
return 49;
if(c == '2')
return 50;
if(c == '3')
return 51;
if(c == '4')
return 52;
if(c == '5')
return 53;
if(c == '6')
return 54;
if(c == '7')
return 55;
if(c == '8')
return 56;
if(c == '9')
return 57;
if(c == ':')
return 58;
if(c == ';')
return 59;
if(c == '<')
return 60;
if(c == '=')
return 61;
if(c == '>')
return 62;
if(c == '?')
return 63;
if(c == '@')
return 64;
if(c == 'A')
return 65;
if(c == 'B')
return 66;
if(c == 'C')
return 67;
if(c == 'D')
return 68;
if(c == 'E')
return 69;
if(c == 'F')
return 70;
if(c == 'G')
return 71;
if(c == 'H')
return 72;
if(c == 'I')
return 73;
if(c == 'J')
return 74;
if(c == 'K')
return 75;
if(c == 'L')
return 76;
if(c == 'M')
return 77;
if(c == 'N')
return 78;
if(c == 'O')
return 79;
if(c == 'P')
return 80;
if(c == 'Q')
return 81;
if(c == 'R')
return 82;
if(c == 'S')
return 83;
if(c == 'T')
return 84;
if(c == 'U')
return 85;
if(c == 'V')
return 86;
if(c == 'W')
return 87;
if(c == 'X')
return 88;
if(c == 'Y')
return 89;
if(c == 'Z')
return 90;
if(c == '[')
return 91;
if(c == '\\')
return 92;
if(c == ']')
return 93;
if(c == '^')
return 94;
if(c == '_')
return 95;
if(c == '`')
return 96;
if(c == 'a')
return 97;
if(c == 'b')
return 98;
if(c == 'c')
return 99;
if(c == 'd')
return 100;
if(c == 'e')
return 101;
if(c == 'f')
return 102;
if(c == 'g')
return 103;
if(c == 'h')
return 104;
if(c == 'i')
return 105;
if(c == 'j')
return 106;
if(c == 'k')
return 107;
if(c == 'l')
return 108;
if(c == 'm')
return 109;
if(c == 'n')
return 110;
if(c == 'o')
return 111;
if(c == 'p')
return 112;
if(c == 'q')
return 113;
if(c == 'r')
return 114;
if(c == 's')
return 115;
if(c == 't')
return 116;
if(c == 'u')
return 117;
if(c == 'v')
return 118;
if(c == 'w')
return 119;
if(c == 'x')
return 120;
if(c == 'y')
return 121;
if(c == 'z')
return 122;
if(c == '{')
return 123;
if(c == '|')
return 124;
if(c == '}')
return 125;
if(c == '~')
return 126;
return 0;
}
#if GPU_DEBUG
#define DebugPrint(fmt_cstr) do { \
u32 __strlen = 0; \
for (;;) { if (U32FromChar(fmt_cstr[__strlen]) == 0) { break; } ++__strlen; } \
RWByteAddressBuffer __print_buff = RWByteAddressBufferFromHandle(GPU_ShaderPrintBufferHandle); \
u32 __pos; \
__print_buff.InterlockedAdd(0, __strlen, __pos); \
if (__pos < countof(__print_buff)) \
{ \
for (u32 char_idx = 0; char_idx < __strlen; ++char_idx) \
{ \
__print_buff.Store(__pos + char_idx, U32FromChar(fmt_cstr[char_idx])); \
} \
} \
} while (0)
#else
#define DebugPrint(...)
#endif

1
src/gpu/gpu_static_extras.h Normal file
View File

@ -0,0 +1 @@
/* This file contains shader resource handles

5
src/meta/meta.c
View File

@ -408,7 +408,8 @@ void BuildEntryPoint(WaveLaneCtx *lane)
//- Dxc
{
PushStringToList(perm, &cp.flags_dxc, Lit("-Od"));
// PushStringToList(perm, &cp.flags_dxc, Lit("-Od"));
PushStringToList(perm, &cp.flags_dxc, Lit("-O3"));
PushStringToList(perm, &cp.flags_dxc, Lit("-Zi -Qembed_debug"));
}
}
@ -955,7 +956,7 @@ void BuildEntryPoint(WaveLaneCtx *lane)
for (u32 gpu_obj_idx = 0; gpu_obj_idx < build.gpu_objs.count; ++gpu_obj_idx)
{
GpuObj *gpu_obj = &build.gpu_objs.array[gpu_obj_idx];
if (!disp_obj || TrimWhitespace(disp_obj->output).len == 0 || disp_obj->return_code != 0)
if (!disp_obj || TrimWhitespace(disp_obj->output).len == 0 || gpu_obj->return_code != 0)
{
disp_obj = gpu_obj;
}

4
src/pp/pp_vis/pp_vis_shaders.hlsl
View File

@ -46,7 +46,7 @@ VertexShader(V_DQuadVS, V_DQuadPSInput)
Vec2 target_pos = 0;
V_DQuadPSInput result;
result.SV_Position = Vec4(NdcFromPos(target_pos, Count2D(target)).xy, 0, 1);
result.SV_Position = Vec4(NdcFromPos(target_pos, countof(target)).xy, 0, 1);
result.quad_idx = SV_InstanceID;
return result;
}
@ -96,7 +96,7 @@ VertexShader(V_DVertVS, V_DVertPSInput)
Vec2 target_pos = vert.pos;
V_DVertPSInput result;
result.SV_Position = Vec4(NdcFromPos(target_pos, Count2D(target)).xy, 0, 1);
result.SV_Position = Vec4(NdcFromPos(target_pos, countof(target)).xy, 0, 1);
result.color_lin = vert.color_lin;
return result;
}

10
src/ui/ui_shaders.hlsl
View File

@ -192,15 +192,7 @@ PixelShader(UI_BlitPS, UI_BlitPSOutput, UI_BlitPSInput input)
// LogDebugF(
// "uv: (%F, %F), result: (%F, %F, %F, %F)",
// FmtFloat(uv.x),
// FmtFloat(uv.y),
// FmtFloat(result.x),
// FmtFloat(result.y),
// FmtFloat(result.z),
// FmtFloat(result.w),
// );
DebugPrint("Hello there");