Skip to content

Commit a069632

Browse files
aratajewweb-flow
authored andcommitted
Support for more variants of 2d block reads
This change extends the implementation of `cl_intel_subgroup_2d_block_io` extension with the following builtins: ```c intel_sub_group_2d_block_read_8b_16r32x1c intel_sub_group_2d_block_read_8b_32r32x1c intel_sub_group_2d_block_read_32b_1r16x1c intel_sub_group_2d_block_read_32b_2r16x1c intel_sub_group_2d_block_read_32b_4r16x1c intel_sub_group_2d_block_read_32b_8r16x1c intel_sub_group_2d_block_read_32b_16r16x1c intel_sub_group_2d_block_read_32b_32r16x1c intel_sub_group_2d_block_prefetch_8b_16r32x1c intel_sub_group_2d_block_prefetch_8b_32r32x1c intel_sub_group_2d_block_prefetch_32b_1r16x1c intel_sub_group_2d_block_prefetch_32b_2r16x1c intel_sub_group_2d_block_prefetch_32b_4r16x1c intel_sub_group_2d_block_prefetch_32b_8r16x1c intel_sub_group_2d_block_prefetch_32b_16r16x1c intel_sub_group_2d_block_prefetch_32b_32r16x1c ``` (cherry picked from commit 111c85f)
1 parent fd68158 commit a069632

File tree

5 files changed

+156
-8
lines changed

5 files changed

+156
-8
lines changed

IGC/BiFModule/Implementation/IGCBiF_Intrinsics_Lsc.cl

Lines changed: 22 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -487,15 +487,19 @@ void __builtin_IB_subgroup_block_read_prefetch_u8_m32k32v2(long baseoffset, int
487487
void __builtin_IB_subgroup_block_read_prefetch_u16_m16k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
488488
void __builtin_IB_subgroup_block_read_prefetch_u16_m32k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
489489

490-
ushort __builtin_IB_subgroup_block_read_cacheopts_u8_m1k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
491-
ushort2 __builtin_IB_subgroup_block_read_cacheopts_u8_m2k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
492-
ushort4 __builtin_IB_subgroup_block_read_cacheopts_u8_m4k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
493-
ushort8 __builtin_IB_subgroup_block_read_cacheopts_u8_m8k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
490+
ushort __builtin_IB_subgroup_block_read_cacheopts_u8_m1k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
491+
ushort2 __builtin_IB_subgroup_block_read_cacheopts_u8_m2k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
492+
ushort4 __builtin_IB_subgroup_block_read_cacheopts_u8_m4k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
493+
ushort8 __builtin_IB_subgroup_block_read_cacheopts_u8_m8k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
494+
ushort16 __builtin_IB_subgroup_block_read_cacheopts_u8_m16k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
495+
ushort32 __builtin_IB_subgroup_block_read_cacheopts_u8_m32k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
494496

495497
void __builtin_IB_subgroup_block_read_prefetch_u8_m1k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
496498
void __builtin_IB_subgroup_block_read_prefetch_u8_m2k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
497499
void __builtin_IB_subgroup_block_read_prefetch_u8_m4k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
498500
void __builtin_IB_subgroup_block_read_prefetch_u8_m8k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
501+
void __builtin_IB_subgroup_block_read_prefetch_u8_m16k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
502+
void __builtin_IB_subgroup_block_read_prefetch_u8_m32k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
499503

500504
ushort __builtin_IB_subgroup_block_read_cacheopts_u16_m1k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
501505
ushort2 __builtin_IB_subgroup_block_read_cacheopts_u16_m2k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
@@ -527,6 +531,20 @@ void __builtin_IB_subgroup_block_read_prefetch_u32_m8k8v1(long baseoffset, int w
527531
void __builtin_IB_subgroup_block_read_prefetch_u32_m16k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
528532
void __builtin_IB_subgroup_block_read_prefetch_u32_m32k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
529533

534+
uint __builtin_IB_subgroup_block_read_cacheopts_u32_m1k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
535+
uint2 __builtin_IB_subgroup_block_read_cacheopts_u32_m2k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
536+
uint4 __builtin_IB_subgroup_block_read_cacheopts_u32_m4k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
537+
uint8 __builtin_IB_subgroup_block_read_cacheopts_u32_m8k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
538+
uint16 __builtin_IB_subgroup_block_read_cacheopts_u32_m16k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
539+
uint32 __builtin_IB_subgroup_block_read_cacheopts_u32_m32k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
540+
541+
void __builtin_IB_subgroup_block_read_prefetch_u32_m1k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
542+
void __builtin_IB_subgroup_block_read_prefetch_u32_m2k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
543+
void __builtin_IB_subgroup_block_read_prefetch_u32_m4k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
544+
void __builtin_IB_subgroup_block_read_prefetch_u32_m8k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
545+
void __builtin_IB_subgroup_block_read_prefetch_u32_m16k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
546+
void __builtin_IB_subgroup_block_read_prefetch_u32_m32k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
547+
530548
uint2 __builtin_IB_subgroup_block_read_cacheopts_u32_m1k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
531549
uint2 __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
532550
uint4 __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);

IGC/BiFModule/Languages/OpenCL/IBiF_Sub_Groups.cl

Lines changed: 18 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1013,6 +1013,8 @@ DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_1r32x1c,
10131013
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_2r32x1c, ushort, ushort2, __builtin_IB_subgroup_block_read_cacheopts_u8_m2k32v1)
10141014
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_4r32x1c, ushort, ushort4, __builtin_IB_subgroup_block_read_cacheopts_u8_m4k32v1)
10151015
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_8r32x1c, ushort, ushort8, __builtin_IB_subgroup_block_read_cacheopts_u8_m8k32v1)
1016+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_16r32x1c, ushort, ushort16, __builtin_IB_subgroup_block_read_cacheopts_u8_m16k32v1)
1017+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_32r32x1c, ushort, ushort32, __builtin_IB_subgroup_block_read_cacheopts_u8_m32k32v1)
10161018

10171019
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_16r32x2c, ushort, ushort32, __builtin_IB_subgroup_block_read_cacheopts_u8_m16k32v2)
10181020
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_8b_32r32x2c, ushort, ushort64, __builtin_IB_subgroup_block_read_cacheopts_u8_m32k32v2)
@@ -1034,6 +1036,13 @@ DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r8x1c,
10341036
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_16r8x1c, uint, uint8, __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v1)
10351037
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_32r8x1c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v1)
10361038

1039+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_1r16x1c, uint, uint, __builtin_IB_subgroup_block_read_cacheopts_u32_m1k16v1)
1040+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_2r16x1c, uint, uint2, __builtin_IB_subgroup_block_read_cacheopts_u32_m2k16v1)
1041+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_4r16x1c, uint, uint4, __builtin_IB_subgroup_block_read_cacheopts_u32_m4k16v1)
1042+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r16x1c, uint, uint8, __builtin_IB_subgroup_block_read_cacheopts_u32_m8k16v1)
1043+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_16r16x1c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_u32_m16k16v1)
1044+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_32r16x1c, uint, uint32, __builtin_IB_subgroup_block_read_cacheopts_u32_m32k16v1)
1045+
10371046
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_1r8x2c, uint, uint2, __builtin_IB_subgroup_block_read_cacheopts_u32_m1k8v2)
10381047
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_2r8x2c, uint, uint2, __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v2)
10391048
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_4r8x2c, uint, uint4, __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v2)
@@ -1078,6 +1087,8 @@ DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_1r32
10781087
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_2r32x1c, __builtin_IB_subgroup_block_read_prefetch_u8_m2k32v1)
10791088
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_4r32x1c, __builtin_IB_subgroup_block_read_prefetch_u8_m4k32v1)
10801089
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_8r32x1c, __builtin_IB_subgroup_block_read_prefetch_u8_m8k32v1)
1090+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_16r32x1c, __builtin_IB_subgroup_block_read_prefetch_u8_m16k32v1)
1091+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_32r32x1c, __builtin_IB_subgroup_block_read_prefetch_u8_m32k32v1)
10811092

10821093
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_16r32x2c, __builtin_IB_subgroup_block_read_prefetch_u8_m16k32v2)
10831094
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_8b_32r32x2c, __builtin_IB_subgroup_block_read_prefetch_u8_m32k32v2)
@@ -1097,6 +1108,13 @@ DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_4r8
10971108
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_8r8x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m8k8v1)
10981109
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_32r8x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m32k8v1)
10991110

1111+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_1r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m1k16v1)
1112+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_2r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m2k16v1)
1113+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_4r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m4k16v1)
1114+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_8r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m8k16v1)
1115+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_16r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m16k16v1)
1116+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_32r16x1c, __builtin_IB_subgroup_block_read_prefetch_u32_m32k16v1)
1117+
11001118
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_1r8x2c, __builtin_IB_subgroup_block_read_prefetch_u32_m1k8v2)
11011119
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_2r8x2c, __builtin_IB_subgroup_block_read_prefetch_u32_m2k8v2)
11021120
DEFN_INTEL_SUB_GROUP_2D_BLOCK_PREFETCH(intel_sub_group_2d_block_prefetch_32b_4r8x2c, __builtin_IB_subgroup_block_read_prefetch_u32_m4k8v2)

IGC/BiFModule/Languages/OpenCL/PreRelease/opencl_cth_pre_release.h

Lines changed: 20 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -2643,17 +2643,21 @@ void intel_subgroup_block_write_cacheopts_u16_m8k16v1(__global void* base_addres
26432643
// `^` - are implemented in opencl_cth_released.h
26442644

26452645
////////////////////////////////////////////////////////////////
2646-
// 2D Block Load / Prefetch, 8-bit data, Rows in [1, 2, 4, 8, 16*, 32*], Columns in [32]:
2646+
// 2D Block Load / Prefetch, 8-bit data, Rows in [1, 2, 4, 8, 16, 32], Columns in [32]:
26472647

26482648
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_1r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
26492649
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_2r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
26502650
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_4r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
26512651
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_8r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
2652+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_16r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
2653+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_8b_32r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private ushort* destination);
26522654

26532655
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_1r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
26542656
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_2r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
26552657
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_4r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
26562658
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_8r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2659+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_16r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2660+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_8b_32r32x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
26572661

26582662
////////////////////////////////////////////////////////////////
26592663

@@ -2723,9 +2727,21 @@ void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_32r8x1c
27232727

27242728

27252729
////////////////////////////////////////////////////////////////
2726-
// 2D Block Load, 32-bit data, Rows in [1*, 2*, 4*, 8*, 16*, 32*], Columns in [16]:
2727-
2728-
// TBD
2730+
// 2D Block Load, 32-bit data, Rows in [1, 2, 4, 8, 16, 32], Columns in [16]:
2731+
2732+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_1r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2733+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_2r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2734+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_4r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2735+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_8r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2736+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_16r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2737+
void __attribute__((overloadable)) intel_sub_group_2d_block_read_32b_32r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord, __private uint* destination);
2738+
2739+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_1r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2740+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_2r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2741+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_4r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2742+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_8r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2743+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_16r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
2744+
void __attribute__((overloadable)) intel_sub_group_2d_block_prefetch_32b_32r16x1c(__global void* base_address, int width, int height, int pitch, int2 coord);
27292745

27302746
////////////////////////////////////////////////////////////////
27312747

0 commit comments

Comments
 (0)