Skip to content

Commit

Permalink
[SYCL][Bindless][Doc] Add support for cubemaps (intel#12996)
Browse files Browse the repository at this point in the history 
Add cubemap support:
 - Allocation and freeing of cubemapped images
 - Unsampled fetching and writing, and sampled reading
 - Device queries for cubemap support
 - Testing for both unsampled and sampled cubemap examples
 - Update the spec with cubemap support

Remove `const` and `&` qualifiers from spec and implementation for
handle parameters in `write_xxx` functions.

Corresponding UR PR:
oneapi-src/unified-runtime#1433

---------

Co-authored-by: Przemek Malon <przemek.malon@codeplay.com>

Resolved Conflicts in:
- Due to not cherry-picking intel#12840
  - sycl/include/sycl/detail/pi.h
- Due to not cherry-picking intel#13181
  - sycl/include/sycl/device_aspect_macros.hpp
  - sycl/include/sycl/info/aspects.def
  • Loading branch information
@Seanst98 @kbenzie
Seanst98 authored and kbenzie committed Apr 18, 2024
1 parent e9ec31d commit 92e0423
Show file tree
Hide file tree
Showing 20 changed files with 1,168 additions and 40 deletions.
127 changes: 127 additions & 0 deletions libclc/ptx-nvidiacl/libspirv/images/image.cl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -216,6 +216,11 @@ pixelf32 as_pixelf32(int4 v) { return as_float4(v); }
return (to_t##2)((to_t)from.x, (to_t)from.y); \
}

#define _DEFINE_VEC4_TO_SINGLE_CAST(from_t, to_t) \
inline to_t cast_##from_t##4_to_##to_t(from_t##4 from) { \
return (to_t)from[0]; \
}

#define _DEFINE_CAST(from_t, to_t) \
inline to_t cast_##from_t##_to_##to_t(from_t from) { return (to_t)from; }

Expand Down Expand Up @@ -278,6 +283,17 @@ _DEFINE_VEC4_TO_VEC2_CAST(float, half)
_DEFINE_VEC4_TO_VEC2_CAST(int, uint)
_DEFINE_VEC4_TO_VEC2_CAST(short, ushort)

_DEFINE_VEC4_TO_SINGLE_CAST(int, int)
_DEFINE_VEC4_TO_SINGLE_CAST(uint, uint)
_DEFINE_VEC4_TO_SINGLE_CAST(float, float)
_DEFINE_VEC4_TO_SINGLE_CAST(short, short)
_DEFINE_VEC4_TO_SINGLE_CAST(short, char)
_DEFINE_VEC4_TO_SINGLE_CAST(int, short)
_DEFINE_VEC4_TO_SINGLE_CAST(int, char)
_DEFINE_VEC4_TO_SINGLE_CAST(uint, ushort)
_DEFINE_VEC4_TO_SINGLE_CAST(uint, uchar)
_DEFINE_VEC4_TO_SINGLE_CAST(float, half)

_DEFINE_VEC2_CAST(int, float)
_DEFINE_VEC2_CAST(short, char)
_DEFINE_VEC2_CAST(short, uchar)
Expand Down Expand Up @@ -332,6 +348,8 @@ _DEFINE_READ_3D_PIXELF(16, clamp)
#undef _DEFINE_VEC4_CAST
#undef _DEFINE_VEC2_CAST
#undef _DEFINE_CAST
#undef _DEFINE_VEC4_TO_VEC2_CAST
#undef _DEFINE_VEC4_TO_SINGLE_CAST
#undef _DEFINE_READ_1D_PIXELF
#undef _DEFINE_READ_2D_PIXELF
#undef _DEFINE_READ_3D_PIXELF
Expand Down Expand Up @@ -3645,3 +3663,112 @@ _CLC_DEFINE_IMAGE_ARRAY_BINDLESS_BUILTIN_ALL(half, DF16_, f, 16)
#undef _NVVM_FUNC
#undef NVVM_FUNC
#undef MANGLE_FUNC_IMG_HANDLE_HELPER


// <--- CUBEMAP --->
// Cubemap surfaces are handled through the layered images implementation

// Define functions to call intrinsic
float4
__nvvm_tex_cube_v4f32_f32(unsigned long, float, float,
float) __asm("__clc_llvm_nvvm_tex_cube_v4f32_f32");
int4 __nvvm_tex_cube_v4i32_f32(unsigned long, float, float, float) __asm(
"__clc_llvm_nvvm_tex_cube_v4i32_f32");
uint4 __nvvm_tex_cube_v4j32_f32(unsigned long, float, float, float) __asm(
"__clc_llvm_nvvm_tex_cube_v4j32_f32");

#define COORD_INPUT float x, float y, float z
#define COORD_THUNK_PARAMS x, y, z
#define COORD_PARAMS coord.x, coord.y, coord.z

// Macro to generate cubemap fetches to call intrinsics
// float4, int4, uint4 already defined above
#define _CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN( \
elem_t, fetch_elem_t, vec_size, fetch_vec_size, coord_input, coord_params) \
elem_t __nvvm_tex_cube_##vec_size##_f32(unsigned long imageHandle, \
coord_input) { \
fetch_elem_t a = \
__nvvm_tex_cube_##fetch_vec_size##_f32(imageHandle, coord_params); \
return cast_##fetch_elem_t##_to_##elem_t(a); \
}

// Float
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(float, float4, f32, v4f32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(float2, float4, v2f32, v4f32, COORD_INPUT, COORD_THUNK_PARAMS)
// Int
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(int, int4, i32, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(int2, int4, v2i32, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
// Uint
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(uint, uint4, j32, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(uint2, uint4, v2j32, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
// Short
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(short, int4, i16, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(short2, int4, v2i16, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(short4, int4, v4i16, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
// UShort
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(ushort, uint4, t16, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(ushort2, uint4, v2t16, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(ushort4, uint4, v4t16, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
// Char
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(char, int4, i8, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(char2, int4, v2i8, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(char4, int4, v4i8, v4i32, COORD_INPUT, COORD_THUNK_PARAMS)
// UChar
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(uchar, uint4, h8, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(uchar2, uint4, v2h8, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(uchar4, uint4, v4h8, v4j32, COORD_INPUT, COORD_THUNK_PARAMS)
// Half
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(half, float4, f16, v4f32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(half2, float4, v2f16, v4f32, COORD_INPUT, COORD_THUNK_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN(half4, float4, v4f16, v4f32, COORD_INPUT, COORD_THUNK_PARAMS)

// Macro to generate the mangled names for cubemap fetches
#define _CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(elem_t, elem_t_mangled, \
vec_size, coord_mangled, \
coord_input, coord_params) \
_CLC_DEF elem_t MANGLE_FUNC_IMG_HANDLE( \
26, __spirv_ImageSampleCubemap, I, \
elem_t_mangled##coord_mangled##ET0_T_T1_)(ulong imageHandle, \
coord_input) { \
return __nvvm_tex_cube_##vec_size##_f32(imageHandle, coord_params); \
}

// Float
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(float, f, f32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(float2, Dv2_f, v2f32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(float4, Dv4_f, v4f32, Dv3_f, float3 coord, COORD_PARAMS)
// Int
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(int, i, i32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(int2, Dv2_i, v2i32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(int4, Dv4_i, v4i32, Dv3_f, float3 coord, COORD_PARAMS)
// Uint
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uint, j, j32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uint2, Dv2_j, v2j32, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uint4, Dv4_j, v4j32, Dv3_f, float3 coord, COORD_PARAMS)
// Short
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(short, s, i16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(short2, Dv2_s, v2i16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(short4, Dv4_s, v4i16, Dv3_f, float3 coord, COORD_PARAMS)
// UShort
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(ushort, t, t16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(ushort2, Dv2_t, v2t16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(ushort4, Dv4_t, v4t16, Dv3_f, float3 coord, COORD_PARAMS)
// Char
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(char, a, i8, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(char2, Dv2_a, v2i8, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(char4, Dv4_a, v4i8, Dv3_f, float3 coord, COORD_PARAMS)
// UChar
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uchar, h, h8, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uchar2, Dv2_h, v2h8, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(uchar4, Dv4_h, v4h8, Dv3_f, float3 coord, COORD_PARAMS)
// Half
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(half, DF16_, f16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(half2, Dv2_DF16_, v2f16, Dv3_f, float3 coord, COORD_PARAMS)
_CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN(half4, Dv4_DF16_, v4f16, Dv3_f, float3 coord, COORD_PARAMS)


#undef _CLC_DEFINE_CUBEMAP_BINDLESS_THUNK_READS_BUILTIN
#undef COORD_INPUT
#undef COORD_THUNK_PARAMS
#undef COORD_PARAMS
#undef _CLC_DEFINE_CUBEMAP_BINDLESS_READS_BUILTIN
25 changes: 25 additions & 0 deletions libclc/ptx-nvidiacl/libspirv/images/image_helpers.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -625,3 +625,28 @@ entry:
%1 = tail call <4 x i32> @__clc_struct32_to_vector({i32,i32,i32,i32} %0)
ret <4 x i32> %1
}

; <--- CUBEMAP --->
declare {float,float,float,float} @llvm.nvvm.tex.unified.cube.v4f32.f32(i64, float, float, float)
define <4 x float> @__clc_llvm_nvvm_tex_cube_v4f32_f32(i64 %img, float %x, float %y, float %z) nounwind alwaysinline {
entry:
%0 = tail call {float,float,float,float} @llvm.nvvm.tex.unified.cube.v4f32.f32(i64 %img, float %x, float %y, float %z);
%1 = tail call <4 x float> @__clc_structf32_to_vector({float,float,float,float} %0)
ret <4 x float> %1
}

declare {i32,i32,i32,i32} @llvm.nvvm.tex.unified.cube.v4s32.f32(i64, float, float, float)
define <4 x i32> @__clc_llvm_nvvm_tex_cube_v4i32_f32(i64 %img, float %x, float %y, float %z) nounwind alwaysinline {
entry:
%0 = tail call {i32,i32,i32,i32} @llvm.nvvm.tex.unified.cube.v4s32.f32(i64 %img, float %x, float %y, float %z);
%1 = tail call <4 x i32> @__clc_struct32_to_vector({i32,i32,i32,i32} %0)
ret <4 x i32> %1
}

declare {i32,i32,i32,i32} @llvm.nvvm.tex.unified.cube.v4u32.f32(i64, float, float, float)
define <4 x i32> @__clc_llvm_nvvm_tex_cube_v4j32_f32(i64 %img, float %x, float %y, float %z) nounwind alwaysinline {
entry:
%0 = tail call {i32,i32,i32,i32} @llvm.nvvm.tex.unified.cube.v4u32.f32(i64 %img, float %x, float %y, float %z);
%1 = tail call <4 x i32> @__clc_struct32_to_vector({i32,i32,i32,i32} %0)
ret <4 x i32> %1
}
5 changes: 4 additions & 1 deletion llvm/include/llvm/SYCLLowerIR/DeviceConfigFile.td
View file
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,8 @@ def AspectExt_oneapi_interop_semaphore_export : Aspect<"ext_oneapi_interop_semap
def AspectExt_oneapi_mipmap : Aspect<"ext_oneapi_mipmap">;
def AspectExt_oneapi_mipmap_anisotropy : Aspect<"ext_oneapi_mipmap_anisotropy">;
def AspectExt_oneapi_mipmap_level_reference : Aspect<"ext_oneapi_mipmap_level_reference">;
def AspectExt_oneapi_cubemap : Aspect<"ext_oneapi_cubemap">;
def AspectExt_oneapi_cubemap_seamless_filtering : Aspect<"ext_oneapi_cubemap_seamless_filtering">;
def AspectExt_intel_esimd : Aspect<"ext_intel_esimd">;
def AspectExt_oneapi_ballot_group : Aspect<"ext_oneapi_ballot_group">;
def AspectExt_oneapi_fixed_size_group : Aspect<"ext_oneapi_fixed_size_group">;
Expand Down Expand Up @@ -119,7 +121,8 @@ def : TargetInfo<"__TestAspectList",
AspectExt_oneapi_bindless_images_shared_usm, AspectExt_oneapi_bindless_images_1d_usm, AspectExt_oneapi_bindless_images_2d_usm,
AspectExt_oneapi_interop_memory_import, AspectExt_oneapi_interop_memory_export,
AspectExt_oneapi_interop_semaphore_import, AspectExt_oneapi_interop_semaphore_export,
AspectExt_oneapi_mipmap, AspectExt_oneapi_mipmap_anisotropy, AspectExt_oneapi_mipmap_level_reference, AspectExt_intel_esimd,
AspectExt_oneapi_mipmap, AspectExt_oneapi_mipmap_anisotropy, AspectExt_oneapi_mipmap_level_reference, AspectExt_oneapi_cubemap,
AspectExt_oneapi_cubemap_seamless_filtering, AspectExt_intel_esimd,
AspectExt_oneapi_ballot_group, AspectExt_oneapi_fixed_size_group, AspectExt_oneapi_opportunistic_group,
AspectExt_oneapi_tangle_group, AspectExt_intel_matrix, AspectExt_oneapi_is_composite, AspectExt_oneapi_is_component,
AspectExt_oneapi_graph, AspectExt_intel_fpga_task_sequence],
Expand Down
Loading

0 comments on commit 92e0423

Please sign in to comment.