Add new functions from SPV_INTEL_2d_block_io extensions

Extended the SPV_INTEL_2d_block_io extension to support new functions:
// type d8, block width 16, array length 1
intel_sub_group_2d_block_read_8b_1r16x1c
intel_sub_group_2d_block_read_8b_2r16x1c
intel_sub_group_2d_block_read_8b_4r16x1c
intel_sub_group_2d_block_read_8b_8r16x1c
intel_sub_group_2d_block_read_8b_16r16x1c
// type d8, block width 16, array length 2
intel_sub_group_2d_block_read_8b_1r16x2c
intel_sub_group_2d_block_read_8b_2r16x2c
intel_sub_group_2d_block_read_8b_4r16x2c
intel_sub_group_2d_block_read_8b_8r16x2c
// type d8, block width 16, array length 4
intel_sub_group_2d_block_read_8b_1r16x4c
intel_sub_group_2d_block_read_8b_2r16x4c
intel_sub_group_2d_block_read_8b_4r16x4c
// type d8, block width 64, array length 1
intel_sub_group_2d_block_read_8b_1r64x1c
intel_sub_group_2d_block_read_8b_2r64x1c
intel_sub_group_2d_block_read_8b_4r64x1c
intel_sub_group_2d_block_read_8b_8r64x1c
intel_sub_group_2d_block_read_8b_16r64x1c
intel_sub_group_2d_block_read_8b_32r64x1c
// type d16, block width 32, array length 1
intel_sub_group_2d_block_read_16b_1r32x1c
intel_sub_group_2d_block_read_16b_2r32x1c
intel_sub_group_2d_block_read_16b_4r32x1c
intel_sub_group_2d_block_read_16b_8r32x1c
intel_sub_group_2d_block_read_16b_16r32x1c
intel_sub_group_2d_block_read_16b_32r32x1c

Author: KanclerzPiotr

IGC/BiFModule/Implementation/IGCBiF_Intrinsics_Lsc.cl

@@ -597,17 +597,51 @@ typedef ushort ushort64 __attribute__((ext_vector_type(64)));
 typedef uint uint32 __attribute__((ext_vector_type(32)));
 typedef uint uint64 __attribute__((ext_vector_type(64)));
 
+uchar   __builtin_IB_subgroup_block_read_cacheopts_u8_m1k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar2  __builtin_IB_subgroup_block_read_cacheopts_u8_m2k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar4  __builtin_IB_subgroup_block_read_cacheopts_u8_m4k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar8  __builtin_IB_subgroup_block_read_cacheopts_u8_m8k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar16 __builtin_IB_subgroup_block_read_cacheopts_u8_m16k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 uchar32 __builtin_IB_subgroup_block_read_cacheopts_u8_m32k16v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 
+uchar2  __builtin_IB_subgroup_block_read_cacheopts_u8_m1k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar4  __builtin_IB_subgroup_block_read_cacheopts_u8_m2k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar8  __builtin_IB_subgroup_block_read_cacheopts_u8_m4k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar16 __builtin_IB_subgroup_block_read_cacheopts_u8_m8k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 uchar32 __builtin_IB_subgroup_block_read_cacheopts_u8_m16k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 uchar64 __builtin_IB_subgroup_block_read_cacheopts_u8_m32k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 
+uchar4   __builtin_IB_subgroup_block_read_cacheopts_u8_m1k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar8   __builtin_IB_subgroup_block_read_cacheopts_u8_m2k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar16  __builtin_IB_subgroup_block_read_cacheopts_u8_m4k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar32  __builtin_IB_subgroup_block_read_cacheopts_u8_m8k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar64  __builtin_IB_subgroup_block_read_cacheopts_u8_m16k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uchar128 __builtin_IB_subgroup_block_read_cacheopts_u8_m32k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+
+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);
+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);
+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);
+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);
+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);
+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);
+
+uint   __builtin_IB_subgroup_block_read_cacheopts_u8_m1k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint2  __builtin_IB_subgroup_block_read_cacheopts_u8_m2k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint4  __builtin_IB_subgroup_block_read_cacheopts_u8_m4k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint8  __builtin_IB_subgroup_block_read_cacheopts_u8_m8k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint16 __builtin_IB_subgroup_block_read_cacheopts_u8_m16k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint32 __builtin_IB_subgroup_block_read_cacheopts_u8_m32k64v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+
+
 ushort32  __builtin_IB_subgroup_block_read_cacheopts_u8_m16k32v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 ushort64  __builtin_IB_subgroup_block_read_cacheopts_u8_m32k32v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 
-uchar32  __builtin_IB_subgroup_block_read_cacheopts_u8_m8k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
-uchar64  __builtin_IB_subgroup_block_read_cacheopts_u8_m16k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
-uchar128  __builtin_IB_subgroup_block_read_cacheopts_u8_m32k16v4(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint   __builtin_IB_subgroup_block_read_cacheopts_u16_m1k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint2  __builtin_IB_subgroup_block_read_cacheopts_u16_m2k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint4  __builtin_IB_subgroup_block_read_cacheopts_u16_m4k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint8  __builtin_IB_subgroup_block_read_cacheopts_u16_m8k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint16 __builtin_IB_subgroup_block_read_cacheopts_u16_m16k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
+uint32 __builtin_IB_subgroup_block_read_cacheopts_u16_m32k32v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 
 ushort32  __builtin_IB_subgroup_block_read_cacheopts_u16_m16k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
 ushort64  __builtin_IB_subgroup_block_read_cacheopts_u16_m32k16v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
@@ -622,12 +656,7 @@ void  __builtin_IB_subgroup_block_read_prefetch_u8_m16k8v4(long baseoffset, int
 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);
 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);
 
-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);
-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);
-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);
-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);
-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);
-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);
+
 
 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);
 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);