diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-28 13:18:46 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-28 17:25:14 +0000 |
commit | 8e2f64f214fa3ce5834db966222fa3b804e236a2 (patch) | |
tree | a5259b9930a6791ce2d7c01403c4376b12f2080a | |
parent | a86433a64ab25c2ea8e274bd2f357a9709636f5b (diff) | |
download | ComputeLibrary-8e2f64f214fa3ce5834db966222fa3b804e236a2.tar.gz |
Create custom flags for enabling fp16 support
Resolves: COMPMID-4655
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I1c1398b39da6da089b288d032074b270d87510c4
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6010
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
-rw-r--r-- | SConstruct | 12 | ||||
-rw-r--r-- | scripts/arm_compute_library_nn_driver.go | 1 | ||||
-rw-r--r-- | src/core/CPP/Validate.h | 14 |
3 files changed, 21 insertions, 6 deletions
diff --git a/SConstruct b/SConstruct index 8dbb68952c..60d220ee2f 100644 --- a/SConstruct +++ b/SConstruct @@ -222,10 +222,13 @@ elif 'v8' in env['arch']: else: env.Append(CXXFLAGS = ['-march=armv8-a']) - if 'v8.6-a' in env['arch'] or env['fat_binary']: + if 'v8.6-a' in env['arch']: env.Append(CPPDEFINES = ['ARM_COMPUTE_ENABLE_I8MM', 'ARM_COMPUTE_ENABLE_BF16']) if "disable_mmla_fp" not in env['custom_options']: env.Append(CPPDEFINES = ['ARM_COMPUTE_ENABLE_SVEF32MM']) + if 'v8.' in env['arch']: + env.Append(CPPDEFINES = ['ARM_COMPUTE_ENABLE_FP16']) + elif 'x86' in env['arch']: if env['estate'] == '32': env.Append(CCFLAGS = ['-m32']) @@ -309,9 +312,10 @@ if env['fat_binary']: if env['arch'] != 'armv8.2-a': print("Currently fat binary is only supported with armv8.2-a") Exit(1) - env.Append(CXXFLAGS = ['-DENABLE_SVE', '-DARM_COMPUTE_ENABLE_SVE', - '-DARM_COMPUTE_ENABLE_BF16', '-DARM_COMPUTE_ENABLE_I8MM', '-DARM_COMPUTE_ENABLE_SVEF32MM']) - env.Append(CXXFLAGS = ['-DENABLE_NEON', '-DARM_COMPUTE_ENABLE_NEON']) + env.Append(CXXFLAGS = ['-DENABLE_NEON', '-DARM_COMPUTE_ENABLE_NEON', + '-DENABLE_SVE', '-DARM_COMPUTE_ENABLE_SVE', + '-DARM_COMPUTE_ENABLE_FP16', '-DARM_COMPUTE_ENABLE_BF16', + '-DARM_COMPUTE_ENABLE_I8MM', '-DARM_COMPUTE_ENABLE_SVEF32MM']) if env['data_type_support']: if any(i in env['data_type_support'] for i in ['all', 'fp16']): diff --git a/scripts/arm_compute_library_nn_driver.go b/scripts/arm_compute_library_nn_driver.go index 2461838d4f..689aa3b8f2 100644 --- a/scripts/arm_compute_library_nn_driver.go +++ b/scripts/arm_compute_library_nn_driver.go @@ -25,6 +25,7 @@ func globalFlags(ctx android.BaseContext) []string { theArch := a.ArchType.String() if theArch == "armv8-2a" { cppflags = append(cppflags, "-march=armv8.2-a+fp16") + cppflags = append(cppflags, "-DARM_COMPUTE_ENABLE_FP16") } } } diff --git a/src/core/CPP/Validate.h b/src/core/CPP/Validate.h index 4a97e4789e..df192b5131 100644 --- a/src/core/CPP/Validate.h +++ b/src/core/CPP/Validate.h @@ -41,8 +41,13 @@ namespace arm_compute inline Status error_on_unsupported_cpu_fp16(const char *function, const char *file, const int line, const ITensorInfo *tensor_info) { + bool fp16_kernels_enabled = false; +#if defined(ARM_COMPUTE_ENABLE_FP16) && defined(ENABLE_FP16_KERNELS) + fp16_kernels_enabled = true; +#endif /* defined(ARM_COMPUTE_ENABLE_FP16) && defined(ENABLE_FP16_KERNELS) */ + ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor_info == nullptr, function, file, line); - ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG((tensor_info->data_type() == DataType::F16) && !CPUInfo::get().has_fp16(), + ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG((tensor_info->data_type() == DataType::F16) && (!CPUInfo::get().has_fp16() || !fp16_kernels_enabled), function, file, line, "This CPU architecture does not support F16 data type, you need v8.2 or above"); return Status{}; } @@ -59,8 +64,13 @@ inline Status error_on_unsupported_cpu_fp16(const char *function, const char *fi inline Status error_on_unsupported_cpu_bf16(const char *function, const char *file, const int line, const ITensorInfo *tensor_info) { + bool bf16_kernels_enabled = false; +#if defined(ARM_COMPUTE_ENABLE_BF16) + bf16_kernels_enabled = true; +#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */ + ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor_info == nullptr, function, file, line); - ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG(tensor_info->data_type() == DataType::BFLOAT16 && !CPUInfo::get().has_bf16(), + ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG((tensor_info->data_type() == DataType::BFLOAT16) && (!CPUInfo::get().has_bf16() || !bf16_kernels_enabled), function, file, line, "This CPU architecture does not support BFloat16 data type, you need v8.6 or above"); return Status{}; } |