cavis/libnd4j/include/ops/declarable/generic/compression/threshold.cpp

106 lines
4.4 KiB
C++
Raw Normal View History

2021-02-01 13:31:45 +01:00
/* ******************************************************************************
*
compression ops (#436) * Added declarations for decode/encode_bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added implementation for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added helpers for encode/decode bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored encodingBitmap helper. Signed-off-by: shugeo <sgazeos@gmail.com> * threshold encode/decode skeleton * helper skeleton * minor import fix * encoder shape fn & op impl * thresholdEncode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thresholdDecode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Only cosmetical changes. Signed-off-by: shugeo <sgazeos@gmail.com> * placeholder Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added cuda implementation for bitmap decode helper. Signed-off-by: shugeo <sgazeos@gmail.com> * cuda thresholdEstimate Signed-off-by: raver119@gmail.com <raver119@gmail.com> * cuda thresholdDecode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - nano cmakelist update (get rid of Clion section) - fixed forgotten throw in AtomicTests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thesholdEncode cuda impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added tests for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed tests for encode/decode bitmaps. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed crashes with bitmap decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * bitmap encode/decode CPU Signed-off-by: raver119@gmail.com <raver119@gmail.com> * bitmap encode/decode CUDA Signed-off-by: raver119@gmail.com <raver119@gmail.com> * C API removed for threshold/bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeBitmap/DecodeBitmap Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more tests for threshold encoding Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor test tweak Signed-off-by: raver119@gmail.com <raver119@gmail.com> * two special tests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeBitmap CPU fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * parallel_long/parallel_double proper spans fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeThreshold CUDA fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * nano fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * grid tweaks Signed-off-by: raver119@gmail.com <raver119@gmail.com> * RTX adaptation for thresholdEncode Signed-off-by: raver119 <raver119@gmail.com> * don't allow threshold encoding for length < 2 Signed-off-by: raver119@gmail.com <raver119@gmail.com> * get rid of NDArrayCompressor in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor update of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor tweak of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - matmul allows integer data types use - EncodingHandler boundary default value - few tests for integer matmul Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor fix of CUDA bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * re-enable CUDA deallocator Signed-off-by: raver119@gmail.com <raver119@gmail.com> * threshold encoder fix for systems without omp Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - encode_threshold now requires non-negative boundary - minor tweak in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * restore parallelism in decode_bitmap Signed-off-by: raver119@gmail.com <raver119@gmail.com> * fall back to omp for encode_bitmap cpu Signed-off-by: raver119@gmail.com <raver119@gmail.com> * single time casts Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - additional test for encode_threshold - sync buffers to device before calling for shape function Signed-off-by: raver119@gmail.com <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-08 19:59:39 +02:00
*
* This program and the accompanying materials are made available under the
* terms of the Apache License, Version 2.0 which is available at
* https://www.apache.org/licenses/LICENSE-2.0.
*
2021-02-01 13:31:45 +01:00
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
compression ops (#436) * Added declarations for decode/encode_bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added implementation for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added helpers for encode/decode bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored encodingBitmap helper. Signed-off-by: shugeo <sgazeos@gmail.com> * threshold encode/decode skeleton * helper skeleton * minor import fix * encoder shape fn & op impl * thresholdEncode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thresholdDecode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Only cosmetical changes. Signed-off-by: shugeo <sgazeos@gmail.com> * placeholder Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added cuda implementation for bitmap decode helper. Signed-off-by: shugeo <sgazeos@gmail.com> * cuda thresholdEstimate Signed-off-by: raver119@gmail.com <raver119@gmail.com> * cuda thresholdDecode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - nano cmakelist update (get rid of Clion section) - fixed forgotten throw in AtomicTests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thesholdEncode cuda impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added tests for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed tests for encode/decode bitmaps. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed crashes with bitmap decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * bitmap encode/decode CPU Signed-off-by: raver119@gmail.com <raver119@gmail.com> * bitmap encode/decode CUDA Signed-off-by: raver119@gmail.com <raver119@gmail.com> * C API removed for threshold/bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeBitmap/DecodeBitmap Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more tests for threshold encoding Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor test tweak Signed-off-by: raver119@gmail.com <raver119@gmail.com> * two special tests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeBitmap CPU fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * parallel_long/parallel_double proper spans fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeThreshold CUDA fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * nano fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * grid tweaks Signed-off-by: raver119@gmail.com <raver119@gmail.com> * RTX adaptation for thresholdEncode Signed-off-by: raver119 <raver119@gmail.com> * don't allow threshold encoding for length < 2 Signed-off-by: raver119@gmail.com <raver119@gmail.com> * get rid of NDArrayCompressor in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor update of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor tweak of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - matmul allows integer data types use - EncodingHandler boundary default value - few tests for integer matmul Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor fix of CUDA bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * re-enable CUDA deallocator Signed-off-by: raver119@gmail.com <raver119@gmail.com> * threshold encoder fix for systems without omp Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - encode_threshold now requires non-negative boundary - minor tweak in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * restore parallelism in decode_bitmap Signed-off-by: raver119@gmail.com <raver119@gmail.com> * fall back to omp for encode_bitmap cpu Signed-off-by: raver119@gmail.com <raver119@gmail.com> * single time casts Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - additional test for encode_threshold - sync buffers to device before calling for shape function Signed-off-by: raver119@gmail.com <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-08 19:59:39 +02:00
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
* License for the specific language governing permissions and limitations
* under the License.
*
* SPDX-License-Identifier: Apache-2.0
******************************************************************************/
//
// @author raver119@gmail.com
//
#include <system/op_boilerplate.h>
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/threshold.h>
namespace sd {
namespace ops {
CUSTOM_OP_IMPL(encode_threshold, 1, 2, true, 1, 0) {
auto x = INPUT_VARIABLE(0);
auto updated = OUTPUT_VARIABLE(0);
auto encoded = OUTPUT_NULLIFIED(1);
float threshold = T_ARG(0);
REQUIRE_TRUE(x->lengthOf() <= DataTypeUtils::max<int>(), 0, "encode_threshold: gradients array must have length <= MAX_INT");
REQUIRE_TRUE(encoded->lengthOf() >= 4, 0, "encode_threshold: array for encoded updates can't have less than 4 elements");
// REQUIRE_TRUE(x->platformBuffer() == updated->platformBuffer(), 0, "encode_threshold: gradients array must be the same at input and output");
// filling header bytes
encoded->p(0, encoded->lengthOf() - 4);
encoded->p(1, (int) x->lengthOf());
encoded->p(2, reinterpret_cast<int *>(&threshold)[0]);
encoded->p(3, 0); // flag for FLEXIBLE_ENCODING
// if there's no updates to process - just skip execution
if (encoded->lengthOf() == 4)
return Status::OK();
helpers::thresholdEncode(*x, *encoded, threshold);
return Status::OK();
}
DECLARE_SHAPE_FN(encode_threshold) {
auto x = INPUT_VARIABLE(0);
// we have limit option here
int boundary = block.numI() > 0 ? I_ARG(0) : DataTypeUtils::max<int>();
float threshold = T_ARG(0);
REQUIRE_TRUE(boundary >= 0, 0, "encode_threshold: boundary must be positive");
REQUIRE_TRUE(x->lengthOf() <= DataTypeUtils::max<int>(), 0, "encode_threshold: gradients array must have length <= MAX_INT");
// we must calculate number of elements that >= threshold
auto elements = sd::math::nd4j_min<int>(helpers::thresholdEstimate(*x, threshold), boundary);
if (elements < 2)
elements = 0;
// result array must have 4 additional int elements for header
return SHAPELIST(x->shapeInfo(), sd::ConstantShapeHelper::getInstance().vectorShapeInfo(elements + 4, DataType::INT32));
compression ops (#436) * Added declarations for decode/encode_bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added implementation for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Added helpers for encode/decode bitmap ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored encodingBitmap helper. Signed-off-by: shugeo <sgazeos@gmail.com> * threshold encode/decode skeleton * helper skeleton * minor import fix * encoder shape fn & op impl * thresholdEncode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thresholdDecode cpu impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Only cosmetical changes. Signed-off-by: shugeo <sgazeos@gmail.com> * placeholder Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added cuda implementation for bitmap decode helper. Signed-off-by: shugeo <sgazeos@gmail.com> * cuda thresholdEstimate Signed-off-by: raver119@gmail.com <raver119@gmail.com> * cuda thresholdDecode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - nano cmakelist update (get rid of Clion section) - fixed forgotten throw in AtomicTests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * thesholdEncode cuda impl Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Added tests for bitmap encoding/decoding ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed tests for encode/decode bitmaps. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed crashes with bitmap decode/encode helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * bitmap encode/decode CPU Signed-off-by: raver119@gmail.com <raver119@gmail.com> * bitmap encode/decode CUDA Signed-off-by: raver119@gmail.com <raver119@gmail.com> * C API removed for threshold/bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeBitmap/DecodeBitmap Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * EncodeThreshold/DecodeThreshold Java side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more tests for threshold encoding Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor test tweak Signed-off-by: raver119@gmail.com <raver119@gmail.com> * two special tests Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeBitmap CPU fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * parallel_long/parallel_double proper spans fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * encodeThreshold CUDA fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * nano fix Signed-off-by: raver119@gmail.com <raver119@gmail.com> * grid tweaks Signed-off-by: raver119@gmail.com <raver119@gmail.com> * RTX adaptation for thresholdEncode Signed-off-by: raver119 <raver119@gmail.com> * don't allow threshold encoding for length < 2 Signed-off-by: raver119@gmail.com <raver119@gmail.com> * get rid of NDArrayCompressor in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor update of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more minor tweak of EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - matmul allows integer data types use - EncodingHandler boundary default value - few tests for integer matmul Signed-off-by: raver119@gmail.com <raver119@gmail.com> * minor fix of CUDA bitmap encode Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * boundary changed to integer everywhere Signed-off-by: raver119@gmail.com <raver119@gmail.com> * re-enable CUDA deallocator Signed-off-by: raver119@gmail.com <raver119@gmail.com> * threshold encoder fix for systems without omp Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - encode_threshold now requires non-negative boundary - minor tweak in EncodingHandler Signed-off-by: raver119@gmail.com <raver119@gmail.com> * restore parallelism in decode_bitmap Signed-off-by: raver119@gmail.com <raver119@gmail.com> * fall back to omp for encode_bitmap cpu Signed-off-by: raver119@gmail.com <raver119@gmail.com> * single time casts Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - additional test for encode_threshold - sync buffers to device before calling for shape function Signed-off-by: raver119@gmail.com <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-08 19:59:39 +02:00
}
DECLARE_TYPES(encode_threshold) {
getOpDescriptor()
->setAllowedInputTypes(0, {ALL_FLOATS})
->setAllowedOutputTypes(0, {ALL_FLOATS})
->setAllowedOutputTypes(1, DataType::INT32);
}
CUSTOM_OP_IMPL(decode_threshold, 2, 1, true, 0, 0) {
auto weights = INPUT_VARIABLE(0);
auto encoded = INPUT_VARIABLE(1);
auto updates = OUTPUT_VARIABLE(0);
REQUIRE_TRUE(encoded->lengthOf() >= 4, 0, "decode_threshold: encoded array can't have length < 4");
REQUIRE_TRUE(updates->lengthOf() == encoded->e<int>(1), 0, "decode_threshold: updates array must have length equal to [%i]", encoded->e<int>(1));
REQUIRE_TRUE(encoded->e<int>(3) == 0, 0, "decode_threshold: encoded array doesn't look like threshold-encoded");
helpers::thresholdDecode(*encoded, *updates);
return Status::OK();
}
DECLARE_SHAPE_FN(decode_threshold) {
auto weights = inputShape->at(0);
return SHAPELIST(weights);
}
DECLARE_TYPES(decode_threshold) {
getOpDescriptor()
->setAllowedInputTypes(0, {ALL_FLOATS})
->setAllowedInputTypes(1, DataType::INT32)
->setAllowedOutputTypes(0,{ALL_FLOATS});
}
}
}