Skip to content

Commit 6c713f0

Browse files
author
Jenkins
committed
Compute Library v23.05
1 parent d8bf9b5 commit 6c713f0

File tree

711 files changed

+181341
-200606
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

711 files changed

+181341
-200606
lines changed

‎.bazelrc‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -37,3 +37,4 @@ build --flag_alias=logging=//:logging
3737
build --flag_alias=openmp=//:openmp
3838
build --flag_alias=cppthreads=//:cppthreads
3939
build --flag_alias=enable_bf16_validation=//:enable_bf16_validation
40+
build --flag_alias=enable_sve_validation=//:enable_sve_validation

‎Android.bp‎

Lines changed: 34 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -27,6 +27,7 @@ opencl_srcs = [
2727
"src/core/CL/cl_kernels/common/elementwise_operation.cl",
2828
"src/core/CL/cl_kernels/common/elementwise_operation_quantized.cl",
2929
"src/core/CL/cl_kernels/common/elementwise_unary.cl",
30+
"src/core/CL/cl_kernels/common/elementwise_unary_quantized.cl",
3031
"src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h",
3132
"src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl",
3233
"src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl",
@@ -49,6 +50,8 @@ opencl_srcs = [
4950
"src/core/CL/cl_kernels/common/generate_proposals_quantized.cl",
5051
"src/core/CL/cl_kernels/common/instance_normalization.cl",
5152
"src/core/CL/cl_kernels/common/l2_normalize.cl",
53+
"src/core/CL/cl_kernels/common/mat_mul.cl",
54+
"src/core/CL/cl_kernels/common/mat_mul_quantized.cl",
5255
"src/core/CL/cl_kernels/common/mean_stddev_normalization.cl",
5356
"src/core/CL/cl_kernels/common/memset.cl",
5457
"src/core/CL/cl_kernels/common/minmax_layer.cl",
@@ -175,6 +178,7 @@ cc_library_static {
175178
"build/android-arm64v8a/src/core/CL",
176179
"src/core/common",
177180
"src/core/helpers",
181+
"src/core/NEON/kernels/arm_gemm",
178182
"src/core/NEON/kernels/assembly",
179183
"src/core/NEON/kernels/convolution/common",
180184
"src/core/NEON/kernels/convolution/winograd",
@@ -292,6 +296,7 @@ cc_library_static {
292296
"src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp",
293297
"src/core/NEON/kernels/NERangeKernel.cpp",
294298
"src/core/NEON/kernels/NEReductionOperationKernel.cpp",
299+
"src/core/NEON/kernels/NEReorderKernel.cpp",
295300
"src/core/NEON/kernels/NEReorgLayerKernel.cpp",
296301
"src/core/NEON/kernels/NEReverseKernel.cpp",
297302
"src/core/NEON/kernels/NESelectKernel.cpp",
@@ -301,6 +306,7 @@ cc_library_static {
301306
"src/core/NEON/kernels/NEStridedSliceKernel.cpp",
302307
"src/core/NEON/kernels/NETileKernel.cpp",
303308
"src/core/NEON/kernels/arm_conv/addressing.cpp",
309+
"src/core/NEON/kernels/arm_conv/depthwise/depthwise_common.cpp",
304310
"src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
305311
"src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
306312
"src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
@@ -483,6 +489,9 @@ cc_library_static {
483489
"src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp",
484490
"src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp",
485491
"src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp",
492+
"src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp",
493+
"src/cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp",
494+
"src/cpu/kernels/elementwise_unary/generic/neon/qasymm8_signed.cpp",
486495
"src/cpu/kernels/floor/neon/fp16.cpp",
487496
"src/cpu/kernels/floor/neon/fp32.cpp",
488497
"src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp",
@@ -510,6 +519,7 @@ cc_library_static {
510519
"src/cpu/kernels/l2normlayer/generic/neon/fp16.cpp",
511520
"src/cpu/kernels/l2normlayer/generic/neon/fp32.cpp",
512521
"src/cpu/kernels/l2normlayer/generic/neon/impl.cpp",
522+
"src/cpu/kernels/lut/generic/neon/u8.cpp",
513523
"src/cpu/kernels/maxunpool/generic/neon/fp16.cpp",
514524
"src/cpu/kernels/maxunpool/generic/neon/fp32.cpp",
515525
"src/cpu/kernels/maxunpool/generic/neon/impl.cpp",
@@ -578,6 +588,7 @@ cc_library_static {
578588
"src/cpu/operators/CpuGemmDirectConv2d.cpp",
579589
"src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp",
580590
"src/cpu/operators/CpuGemmLowpOutputStage.cpp",
591+
"src/cpu/operators/CpuMatMul.cpp",
581592
"src/cpu/operators/CpuMaxUnpooling.cpp",
582593
"src/cpu/operators/CpuMul.cpp",
583594
"src/cpu/operators/CpuPermute.cpp",
@@ -630,8 +641,10 @@ cc_library_static {
630641
"src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp",
631642
"src/dynamic_fusion/sketch/gpu/operators/GpuReshape.cpp",
632643
"src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp",
644+
"src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp",
633645
"src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp",
634646
"src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp",
647+
"src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp",
635648
"src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp",
636649
"src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp",
637650
"src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp",
@@ -685,6 +698,8 @@ cc_library_static {
685698
"src/gpu/cl/kernels/ClIm2ColKernel.cpp",
686699
"src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.cpp",
687700
"src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp",
701+
"src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp",
702+
"src/gpu/cl/kernels/ClMatMulNativeKernel.cpp",
688703
"src/gpu/cl/kernels/ClMulKernel.cpp",
689704
"src/gpu/cl/kernels/ClPermuteKernel.cpp",
690705
"src/gpu/cl/kernels/ClPool2dKernel.cpp",
@@ -733,6 +748,7 @@ cc_library_static {
733748
"src/gpu/cl/operators/ClGemmLowpOutputStage.cpp",
734749
"src/gpu/cl/operators/ClIndirectConv2d.cpp",
735750
"src/gpu/cl/operators/ClLogicalNot.cpp",
751+
"src/gpu/cl/operators/ClMatMul.cpp",
736752
"src/gpu/cl/operators/ClMul.cpp",
737753
"src/gpu/cl/operators/ClPRelu.cpp",
738754
"src/gpu/cl/operators/ClPermute.cpp",
@@ -815,6 +831,7 @@ cc_library_static {
815831
"src/runtime/CL/functions/CLLogicalAnd.cpp",
816832
"src/runtime/CL/functions/CLLogicalNot.cpp",
817833
"src/runtime/CL/functions/CLLogicalOr.cpp",
834+
"src/runtime/CL/functions/CLMatMul.cpp",
818835
"src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp",
819836
"src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp",
820837
"src/runtime/CL/functions/CLNormalizationLayer.cpp",
@@ -928,6 +945,7 @@ cc_library_static {
928945
"src/runtime/NEON/functions/NELSTMLayer.cpp",
929946
"src/runtime/NEON/functions/NELSTMLayerQuantized.cpp",
930947
"src/runtime/NEON/functions/NELogical.cpp",
948+
"src/runtime/NEON/functions/NEMatMul.cpp",
931949
"src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp",
932950
"src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp",
933951
"src/runtime/NEON/functions/NENormalizationLayer.cpp",
@@ -946,6 +964,7 @@ cc_library_static {
946964
"src/runtime/NEON/functions/NERange.cpp",
947965
"src/runtime/NEON/functions/NEReduceMean.cpp",
948966
"src/runtime/NEON/functions/NEReductionOperation.cpp",
967+
"src/runtime/NEON/functions/NEReorderLayer.cpp",
949968
"src/runtime/NEON/functions/NEReorgLayer.cpp",
950969
"src/runtime/NEON/functions/NEReshapeLayer.cpp",
951970
"src/runtime/NEON/functions/NEReverse.cpp",
@@ -981,6 +1000,8 @@ cc_library_static {
9811000
"src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigValhall.cpp",
9821001
"src/runtime/heuristics/dwc_native/ClDWCNativeHeuristicsHelpers.cpp",
9831002
"src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp",
1003+
"src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp",
1004+
"src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp",
9841005
"utils/CommonGraphOptions.cpp",
9851006
"utils/GraphUtils.cpp",
9861007
"utils/Utils.cpp",
@@ -1053,6 +1074,16 @@ cc_library_static {
10531074
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
10541075
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
10551076
"src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1077+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1078+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1079+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1080+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1081+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1082+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1083+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1084+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1085+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1086+
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
10561087
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
10571088
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
10581089
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
@@ -1070,17 +1101,14 @@ cc_library_static {
10701101
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp",
10711102
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp",
10721103
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1073-
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
10741104
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
10751105
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
10761106
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
10771107
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1078-
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
10791108
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
10801109
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
10811110
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
10821111
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1083-
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
10841112
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
10851113
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
10861114
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
@@ -1094,10 +1122,8 @@ cc_library_static {
10941122
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
10951123
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
10961124
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1097-
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
10981125
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
10991126
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1100-
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst_strided/generic.cpp",
11011127
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
11021128
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
11031129
"src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
@@ -1261,6 +1287,9 @@ cc_library_static {
12611287
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp",
12621288
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp",
12631289
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp",
1290+
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp",
1291+
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp",
1292+
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp",
12641293
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp",
12651294
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp",
12661295
"src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp",
@@ -1319,9 +1348,6 @@ cc_library_static {
13191348
"src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp",
13201349
"src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp",
13211350
"src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp",
1322-
"src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_fp32_mla_8x1VL/generic.cpp",
1323-
"src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_s8s32_dot_8x1VL/generic.cpp",
1324-
"src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_u8u32_dot_8x1VL/generic.cpp",
13251351
"src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp",
13261352
"src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp",
13271353
"src/core/NEON/kernels/convolution/winograd/input_transforms/sme_fp32_mla_6x6.cpp",

‎BUILD.bazel‎

Lines changed: 27 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -72,6 +72,12 @@ bool_flag(
7272
visibility = ["//visibility:public"],
7373
)
7474

75+
bool_flag(
76+
name = "enable_sve_validation",
77+
build_setting_default = False,
78+
visibility = ["//visibility:public"],
79+
)
80+
7581
#---------------------------------------------------------------------
7682
# Flag variables
7783
config_setting(
@@ -116,6 +122,13 @@ config_setting(
116122
},
117123
)
118124

125+
config_setting(
126+
name = "sve_validation_flag",
127+
flag_values = {
128+
":enable_sve_validation": "true",
129+
},
130+
)
131+
119132

120133
#---------------------------------------------------------------------
121134
# Common defines used for all targets
@@ -138,8 +151,6 @@ cc_library(
138151
"DARM_COMPUTE_GRAPH_ENABLED",
139152
"ARM_COMPUTE_ENABLE_SVEF32MM",
140153
"ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS",
141-
"ENABLE_SVE",
142-
"ARM_COMPUTE_ENABLE_SVE",
143154
"_GLIBCXX_USE_NANOSLEEP"
144155
] + select({
145156
"//:debug_flag": [
@@ -181,7 +192,7 @@ cc_library(
181192
name = "arm_compute_graph",
182193
srcs = ["//src:arm_compute_graph_srcs"],
183194
copts = [
184-
"-march=armv8.2-a+fp16", # What arch is it we should go for here?
195+
"-march=armv8.2-a+fp16",
185196
] + select({
186197
"//:debug_flag": [
187198
"-O0",
@@ -200,7 +211,7 @@ cc_library(
200211
}),
201212
visibility = ["//visibility:public"],
202213
deps = [
203-
"arm_compute",
214+
"arm_compute_core",
204215
"//:common_defines",
205216
"//arm_compute:graph_headers",
206217
],
@@ -214,7 +225,7 @@ cc_library(
214225
name = "arm_compute_sve2",
215226
srcs = ["//src:arm_compute_sve2_srcs"],
216227
copts = [
217-
"-march=armv8.6-a+sve2+fp16+dotprod", # What arch is it we should go for here?
228+
"-march=armv8.6-a+sve2+fp16+dotprod",
218229
] + select({
219230
"//:debug_flag": [
220231
"-O0",
@@ -243,8 +254,10 @@ cc_library(
243254
"//conditions:default": [],
244255
}),
245256
local_defines = [
257+
"ENABLE_SVE",
258+
"ARM_COMPUTE_ENABLE_SVE",
246259
"ARM_COMPUTE_ENABLE_SVE2",
247-
"ARM_COMPUTE_ENABLE_BF16"
260+
"ARM_COMPUTE_ENABLE_BF16",
248261
],
249262
deps = [
250263
"//:common_defines",
@@ -263,7 +276,7 @@ cc_library(
263276
name = "arm_compute_sve",
264277
srcs = ["//src:arm_compute_sve_srcs"],
265278
copts = [
266-
"-march=armv8.2-a+sve+fp16+dotprod", # What arch is it we should go for here?
279+
"-march=armv8.2-a+sve+fp16+dotprod",
267280
] + select({
268281
"//:debug_flag": [
269282
"-O0",
@@ -292,6 +305,8 @@ cc_library(
292305
"//conditions:default": [],
293306
}),
294307
local_defines = [
308+
"ENABLE_SVE",
309+
"ARM_COMPUTE_ENABLE_SVE",
295310
"ARM_COMPUTE_ENABLE_BF16",
296311
],
297312
deps = [
@@ -308,8 +323,8 @@ cc_library(
308323
# Core and Runtime library
309324

310325
cc_library(
311-
name = "arm_compute",
312-
srcs = ["//src:arm_compute_srcs"],
326+
name = "arm_compute_core",
327+
srcs = ["//src:arm_compute_core_srcs"],
313328
hdrs = glob([
314329
"core/NEON/kernels/**/*.h",
315330
"core/NEON/kernels/**/*.hpp",
@@ -318,7 +333,7 @@ cc_library(
318333
"//:create_version_file",
319334
],
320335
copts = [
321-
"-march=armv8.2-a+fp16", # What arch is it we should go for here?
336+
"-march=armv8.2-a+fp16",
322337
] + select({
323338
"//:debug_flag": [
324339
"-O0",
@@ -348,6 +363,8 @@ cc_library(
348363
"//conditions:default": [],
349364
}),
350365
local_defines = [
366+
"ENABLE_SVE",
367+
"ARM_COMPUTE_ENABLE_SVE",
351368
"ARM_COMPUTE_ENABLE_BF16",
352369
],
353370
visibility = ["//visibility:public"],

0 commit comments

Comments
 (0)