Skip to content

Commit fd68158

Browse files
aratajewweb-flow
aratajew
authored and
web-flow
committed
Fix return type for 32-bit 2d block reads
Returned types for 32-bit data 2d block reads were too wide. Example: the entire block of data read by `intel_sub_group_2d_block_read_32b_2r8x2c` per subgroup is: `4 bytes * 2 * 8 * 2 = 128 bytes` The function is designed to be compiled in SIMD16, therefore one workitem receives `128 bytes / 16 = 8 bytes (uint2)`. (cherry picked from commit 2f7b0e0)
1 parent 593ab67 commit fd68158

File tree

3 files changed

+32
-32
lines changed

3 files changed

+32
-32
lines changed

‎IGC/BiFModule/Implementation/IGCBiF_Intrinsics_Lsc.cl

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -514,11 +514,11 @@ void __builtin_IB_subgroup_block_read_prefetch_u16_m32k16v1(long baseoffset, int
514514
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);
515515

516516
uint __builtin_IB_subgroup_block_read_cacheopts_u32_m1k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
517-
uint2 __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
518-
uint4 __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
519-
uint8 __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
520-
uint16 __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
521-
uint32 __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
517+
uint __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
518+
uint2 __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
519+
uint4 __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
520+
uint8 __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
521+
uint16 __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
522522

523523
void __builtin_IB_subgroup_block_read_prefetch_u32_m1k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
524524
void __builtin_IB_subgroup_block_read_prefetch_u32_m2k8v1(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
@@ -528,11 +528,11 @@ void __builtin_IB_subgroup_block_read_prefetch_u32_m16k8v1(long baseoffset, int
528528
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);
529529

530530
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);
531-
uint4 __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);
532-
uint8 __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);
533-
uint16 __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
534-
uint32 __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
535-
uint64 __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
531+
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);
532+
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);
533+
uint8 __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
534+
uint16 __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
535+
uint32 __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
536536

537537
void __builtin_IB_subgroup_block_read_prefetch_u32_m1k8v2(long baseoffset, int width_minus_one, int height_minus_one, int pitch_minus_one, int2 coord, enum LSC_LDCC cacheOpt);
538538
void __builtin_IB_subgroup_block_read_prefetch_u32_m2k8v2(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: 12 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -1028,18 +1028,18 @@ DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_16b_16r16x1c,
10281028
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_16b_32r16x1c, ushort, ushort32, __builtin_IB_subgroup_block_read_cacheopts_u16_m32k16v1)
10291029

10301030
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_1r8x1c, uint, uint, __builtin_IB_subgroup_block_read_cacheopts_u32_m1k8v1)
1031-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_2r8x1c, uint, uint2, __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v1)
1032-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_4r8x1c, uint, uint4, __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v1)
1033-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r8x1c, uint, uint8, __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v1)
1034-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_16r8x1c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v1)
1035-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_32r8x1c, uint, uint32, __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v1)
1036-
1037-
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)
1038-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_2r8x2c, uint, uint4, __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v2)
1039-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_4r8x2c, uint, uint8, __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v2)
1040-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r8x2c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v2)
1041-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_16r8x2c, uint, uint32, __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v2)
1042-
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_32r8x2c, uint, uint64, __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v2)
1031+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_2r8x1c, uint, uint, __builtin_IB_subgroup_block_read_cacheopts_u32_m2k8v1)
1032+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_4r8x1c, uint, uint2, __builtin_IB_subgroup_block_read_cacheopts_u32_m4k8v1)
1033+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r8x1c, uint, uint4, __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v1)
1034+
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)
1035+
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)
1036+
1037+
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)
1038+
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)
1039+
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)
1040+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_8r8x2c, uint, uint8, __builtin_IB_subgroup_block_read_cacheopts_u32_m8k8v2)
1041+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_16r8x2c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_u32_m16k8v2)
1042+
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_32b_32r8x2c, uint, uint32, __builtin_IB_subgroup_block_read_cacheopts_u32_m32k8v2)
10431043

10441044
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_transform_16b_32r16x1c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_transform_u16_k32v1)
10451045
DEFN_INTEL_SUB_GROUP_2D_BLOCK_READ(intel_sub_group_2d_block_read_transform_16b_16r16x2c, uint, uint16, __builtin_IB_subgroup_block_read_cacheopts_transform_u16_k16v2)

‎IGC/ocloc_tests/Builtins/cl_intel_subgroup_2d_block_io/PVC/block_reads.cl

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -165,31 +165,31 @@ SPDX-License-Identifier: MIT
165165
// CHECK-VISAASM-32B-1R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
166166

167167
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
168-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_2r8x1c -DDST_ARRAY_EL_TYPE=uint2 -DDST_ARRAY_EL_NUM=1" \
168+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_2r8x1c -DDST_ARRAY_EL_TYPE=uint -DDST_ARRAY_EL_NUM=1" \
169169
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-2R-8X1C
170170

171171
// CHECK-VISAASM-32B-2R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
172172

173173
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
174-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_4r8x1c -DDST_ARRAY_EL_TYPE=uint4 -DDST_ARRAY_EL_NUM=1" \
174+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_4r8x1c -DDST_ARRAY_EL_TYPE=uint2 -DDST_ARRAY_EL_NUM=1" \
175175
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-4R-8X1C
176176

177177
// CHECK-VISAASM-32B-4R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
178178

179179
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
180-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_8r8x1c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=1" \
180+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_8r8x1c -DDST_ARRAY_EL_TYPE=uint4 -DDST_ARRAY_EL_NUM=1" \
181181
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-8R-8X1C
182182

183183
// CHECK-VISAASM-32B-8R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
184184

185185
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
186-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_16r8x1c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=2" \
186+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_16r8x1c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=1" \
187187
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-16R-8X1C
188188

189189
// CHECK-VISAASM-32B-16R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
190190

191191
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
192-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_32r8x1c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=4" \
192+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_32r8x1c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=2" \
193193
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-32R-8X1C
194194

195195
// CHECK-VISAASM-32B-32R-8X1C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.8x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
@@ -201,31 +201,31 @@ SPDX-License-Identifier: MIT
201201
// CHECK-VISAASM-32B-1R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
202202

203203
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
204-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_2r8x2c -DDST_ARRAY_EL_TYPE=uint4 -DDST_ARRAY_EL_NUM=1" \
204+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_2r8x2c -DDST_ARRAY_EL_TYPE=uint2 -DDST_ARRAY_EL_NUM=1" \
205205
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-2R-8X2C
206206

207207
// CHECK-VISAASM-32B-2R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
208208

209209
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
210-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_4r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=1" \
210+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_4r8x2c -DDST_ARRAY_EL_TYPE=uint4 -DDST_ARRAY_EL_NUM=1" \
211211
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-4R-8X2C
212212

213213
// CHECK-VISAASM-32B-4R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
214214

215215
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
216-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_8r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=2" \
216+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_8r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=1" \
217217
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-8R-8X2C
218218

219219
// CHECK-VISAASM-32B-8R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
220220

221221
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
222-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_16r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=4" \
222+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_16r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=2" \
223223
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-16R-8X2C
224224

225225
// CHECK-VISAASM-32B-16R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
226226

227227
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
228-
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_32r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=8" \
228+
// RUN: -DINPUT_TYPE=uint -DOUTPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_read_32b_32r8x2c -DDST_ARRAY_EL_TYPE=uint8 -DDST_ARRAY_EL_NUM=4" \
229229
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-32R-8X2C
230230

231231
// CHECK-VISAASM-32B-32R-8X2C: lsc_load_block2d.ugm (M1, 1) V{{[0-9]+}}:d32.2x8x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}

0 commit comments

Comments
 (0)