diff options
author | Sheri Zhang <sheri.zhang@arm.com> | 2020-03-09 14:29:52 +0000 |
---|---|---|
committer | Sheri Zhang <sheri.zhang@arm.com> | 2020-03-25 15:58:42 +0000 |
commit | 1b14c75c0d591c4abe4d2d41b7e4e165fbf58382 (patch) | |
tree | 41e671befde3f61247d0728d16907ff281d6294d /src | |
parent | 2e5fd637205770ec5e11096e6e19b8efc67d544e (diff) | |
download | ComputeLibrary-1b14c75c0d591c4abe4d2d41b7e4e165fbf58382.tar.gz |
COMPMID-2968: Add support for QASYMM8_SIGNED in CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I37e6e76dbd5546c0eaedfacd01ea905c37148e8a
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2861
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src')
-rw-r--r-- | src/core/CL/cl_kernels/gemmlowp.cl | 24 | ||||
-rw-r--r-- | src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.cpp (renamed from src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp) | 60 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp | 28 |
3 files changed, 69 insertions, 43 deletions
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl index 3fba781ede..7f2828689a 100644 --- a/src/core/CL/cl_kernels/gemmlowp.cl +++ b/src/core/CL/cl_kernels/gemmlowp.cl @@ -2317,9 +2317,9 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16(TENSOR3D_DE #endif // defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT) #if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET) -/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8 +/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED * - * This kernel takes a final int32 accumulator value (the output of matrix multiplication), and processes it to obtain the final QASYMM8 value. + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value. * The following computations will be performed by the kernel: * * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier @@ -2327,11 +2327,14 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16(TENSOR3D_DE * -# Requantize * -# Add offset to each result * -# Clamp the value between the specified min and max bounds - * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. + * -# Clamp the resulting int32 values: + * - to the [0..255] range and cast to QASYMM8. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. * * @attention The offset and scalar scale factor must be passed at compile time using -DRESULT_OFFSET, -DREAL_MULTIPLIER * * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time + * @note The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. * These values can be used to implement "rectified linear unit" activation functions * @@ -2388,19 +2391,20 @@ __kernel void gemmlowp_output_stage_quantize_down_float(TENSOR3D_DECLARATION(src #endif // defined(ADD_BIAS) // Convert to float - float16 input_values_f = convert_float4(input_values); - input_values_f = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET); + float4 input_values_f = convert_float4(input_values); + input_values_f = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET); - uchar4 res = convert_uchar4_sat(input_values_f); + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4) + res = CONVERT_SAT(input_values_f, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4)); #if defined(MIN_BOUND) - res = max(res, (uchar4)MIN_BOUND); + res = max(res, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4))MIN_BOUND); #endif // defined(MIN_BOUND) #if defined(MAX_BOUND) - res = min(res, (uchar4)MAX_BOUND); + res = min(res, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4))MAX_BOUND); #endif // defined(MAX_BOUND) // Store the result - vstore4(res, 0, dst_addr); + vstore4(res, 0, (__global OUTPUT_DATA_TYPE *)dst_addr); } -#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET) +#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
\ No newline at end of file diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.cpp index 7097dc9248..5a554f3111 100644 --- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.cpp @@ -21,9 +21,10 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h" +#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h" #include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" @@ -32,7 +33,7 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" - +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "support/StringSupport.h" namespace arm_compute @@ -40,10 +41,13 @@ namespace arm_compute namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, - int min, int max) + const GEMMLowpOutputStageInfo *info) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); - ARM_COMPUTE_RETURN_ERROR_ON(min > max); + ARM_COMPUTE_RETURN_ERROR_ON((info->output_data_type != DataType::QASYMM8) && (info->output_data_type != DataType::QASYMM8_SIGNED)); + ARM_COMPUTE_RETURN_ERROR_ON(info->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(info->output_data_type))); + ARM_COMPUTE_RETURN_ERROR_ON(info->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(info->output_data_type)) + || info->gemmlowp_min_bound > info->gemmlowp_max_bound); // Check biases if exist if(bias != nullptr) @@ -55,15 +59,18 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con if(output->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() != info->output_data_type, "Mismatching output data type"); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); } return Status{}; } -std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output) +std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, DataType output_data_type) { + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output, input->clone()->set_data_type(output_data_type)); + constexpr unsigned int num_elems_processed_per_iteration = 4; // Output auto inizialitation if not yet initialized @@ -77,14 +84,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen bool window_changed = update_window_and_padding(win, input_access); - if(output->total_size() != 0) - { - Window win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); - AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win_out, output_result_access); - - output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); - } + AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration); + window_changed = window_changed || update_window_and_padding(win, output_result_access); + output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); if(bias != nullptr) { @@ -98,39 +100,39 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen } // namespace class Coordinates; -CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel() +CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel() : _input(nullptr), _bias(nullptr), _output(nullptr) { } -Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max) +Status CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, + const GEMMLowpOutputStageInfo *info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), - (bias != nullptr) ? bias->clone().get() : nullptr, - output->clone().get()) - .first); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, info)); return Status{}; } -void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, - float multiplier, int offset, - int min, int max) +void CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, + const GEMMLowpOutputStageInfo *info) { // Perform validate step ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), info)); _input = input; _bias = bias; _output = output; + auto min = info->gemmlowp_min_bound; + auto max = info->gemmlowp_max_bound; + // Set the arguments to pass at compile time CLBuildOptions build_opts; - build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(multiplier)); - build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(offset)); + build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(info->gemmlowp_real_multiplier)); + build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(info->gemmlowp_offset)); + build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type())); build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min)); build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max)); build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); @@ -139,12 +141,12 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::configure(const ICLTe _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down_float", build_opts.options())); // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info()); + auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), info->output_data_type); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure_internal(win_config.second); } -void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::run(const Window &window, cl::CommandQueue &queue) +void CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::run(const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp index e86f303ff4..fbd1820098 100644 --- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp +++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp @@ -24,11 +24,11 @@ #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" #include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h" #include "support/MemorySupport.h" namespace arm_compute @@ -90,15 +90,24 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::configure(const ICLTensor * float multiplier, int offset, int min, int max) { - auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel>(); - k->configure(input, bias, output, multiplier, offset, min, max); + GEMMLowpOutputStageInfo info = GEMMLowpOutputStageInfo(); + info.gemmlowp_offset = offset; + info.gemmlowp_real_multiplier = multiplier; + info.gemmlowp_min_bound = min; + info.gemmlowp_max_bound = max; + + auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel>(); + k->configure(input, bias, output, &info); _kernel = std::move(k); } Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max) { - return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(input, bias, output, min, max); + GEMMLowpOutputStageInfo info = GEMMLowpOutputStageInfo(); + info.gemmlowp_min_bound = min; + info.gemmlowp_max_bound = max; + return CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(input, bias, output, &info); } void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, @@ -165,6 +174,13 @@ void CLGEMMLowpOutputStage::configure(const ICLTensor *input, const ICLTensor *b } break; } + case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT: + { + auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel>(); + k->configure(input, bias, output, &info); + _kernel = std::move(k); + break; + } default: ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type."); } @@ -202,6 +218,10 @@ Status CLGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorIn return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type."); } } + case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT: + { + return CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(input, bias, output, &info); + } default: return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type."); } |