|
- /**
- * \file dnn/test/naive/conv_bias.cpp
- * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
- *
- * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- */
-
- #include "test/naive/fixture.h"
- #include "megdnn/oprs/nn.h"
- #include "test/common/checker.h"
- #include "test/common/workspace_wrapper.h"
-
- using namespace megdnn;
- using namespace test;
-
- namespace {
- class TensorWrapper {
- public:
- TensorWrapper(Handle* handle, TensorLayout layout) : m_handle(handle) {
- m_tensornd.raw_ptr = megdnn_malloc(m_handle, layout.span().dist_byte());
- m_tensornd.layout = layout;
- }
- ~TensorWrapper() { megdnn_free(m_handle, m_tensornd.raw_ptr); }
-
- TensorND tensornd() const { return m_tensornd; }
-
- private:
- Handle* m_handle;
- TensorND m_tensornd;
- };
- } // namespace
-
- TEST_F(NAIVE, CONV_BIAS_QUANTIZED8x8x32) {
- Checker<ConvBias> checker(handle(), /* check_dispatch */false);
- ConvBias::Param param;
- param.format = ConvBias::Param::Format::NCHW;
-
- checker.set_param(param).exect(
- Testcase{
- TensorValue({1, 1, 4, 4}, dtype::QuantizedS8(0.1f),
- {90-128, 136-128, 85-128, 204-128,
- 48-128, 9-128, 226-128, 25-128,
- 118-128, 109-128, 87-128, 132-128,
- 104-128, 163-128, 25-128, 90-128}),
- TensorValue({3, 1, 3, 3}, dtype::QuantizedS8(0.2f),
- {153-124, 170-124, 102-124,
- 103-124, 23-124, 213-124,
- 116-124, 195-124, 191-124,
-
- 44-124, 50-124, 247-124,
- 172-124, 42-124, 32-124,
- 233-124, 163-124, 247-124,
-
- 120-124, 241-124, 209-124,
- 83-124, 201-124, 115-124,
- 32-124, 140-124, 147-124}),
- TensorValue({1, 3, 1, 1}, dtype::QuantizedS32(0.02f),
- {0, 0, 0}),
- TensorValue({1, 3, 2, 2}, dtype::QuantizedS32(0.3f),
- {1234, 0,
- 0, 0,
-
- 0, 0,
- 0, 0,
-
- 0, -234,
- 0, 0}),
- {}},
- Testcase{
- {},
- {},
- {},
- {},
- TensorValue({1, 3, 2, 2}, dtype::QuantizedS32(0.1f * 0.2f),
- {37127, -22475,
- -15694, -1920,
-
- -12813, 4440,
- 18190, -13195,
-
- -9659, 12423,
- -5558, -4969})});
- }
-
- TEST_F(NAIVE, CONV_BIAS_QUANTIZED4x4x32) {
- Checker<ConvBias> checker(handle(), false);
- using Param = ConvBiasForward::Param;
- Param param;
- param.format = Param::Format::NCHW8;
- checker.set_param(param);
- auto GenTensorValueQuint4 = [](const TensorShape& shape,
- dtype::Quantized4Asymm dtype,
- const std::vector<int>& values) {
- TensorND tensor;
- tensor.layout = {shape, dtype};
- tensor.raw_ptr =
- static_cast<dt_byte*>(malloc(tensor.layout.span().dist_byte()));
- uint8_t* ptr = static_cast<uint8_t*>(tensor.raw_ptr);
- megdnn_assert(values.size() == tensor.layout.span().dist_elem());
- for (size_t i = 0; i < tensor.layout.span().dist_elem(); i += 2) {
- int val0 = values[i], val1 = values[i + 1];
- ptr[i / 2] = val0 | (val1 << 4);
- }
- return tensor;
- };
- checker.set_param(param).exect(
- Testcase{
- GenTensorValueQuint4(
- {1, 1, 4, 4, 8},
- dtype::Quantized4Asymm(0.1f, uint8_t(8)),
- {0, 6, 14, 5, 11, 2, 9, 9, 2, 1, 2, 11, 5,
- 0, 4, 8, 12, 15, 7, 7, 11, 0, 4, 1, 14, 9,
- 2, 0, 1, 11, 7, 13, 6, 11, 14, 4, 14, 6, 4,
- 3, 4, 2, 8, 15, 10, 6, 7, 0, 11, 13, 3, 9,
- 5, 13, 0, 5, 4, 5, 10, 5, 5, 0, 3, 13, 5,
- 4, 14, 10, 8, 3, 15, 1, 13, 5, 8, 9, 13, 10,
- 15, 13, 9, 0, 1, 11, 15, 4, 12, 11, 4, 5, 2,
- 9, 10, 9, 3, 1, 15, 10, 0, 1, 4, 6, 11, 2,
- 4, 9, 14, 6, 12, 0, 10, 13, 9, 7, 14, 14, 3,
- 14, 14, 7, 2, 4, 1, 9, 4, 7, 15, 10}),
- GenTensorValueQuint4(
- {8, 1, 3, 3, 8},
- dtype::Quantized4Asymm(0.2f, uint8_t(7)),
- {6, 8, 3, 6, 1, 9, 7, 8, 10, 0, 4, 11, 0,
- 1, 9, 8, 3, 3, 0, 9, 3, 2, 2, 2, 10, 5,
- 8, 7, 12, 10, 1, 11, 3, 1, 9, 8, 2, 15, 5,
- 0, 14, 3, 8, 15, 14, 7, 15, 4, 3, 3, 11, 9,
- 8, 4, 7, 14, 4, 6, 10, 7, 5, 5, 2, 0, 5,
- 0, 1, 10, 13, 1, 7, 12, 9, 11, 12, 7, 3, 15,
- 1, 10, 7, 8, 9, 1, 6, 8, 7, 0, 4, 12, 12,
- 11, 4, 0, 14, 1, 6, 15, 15, 4, 1, 2, 10, 9,
- 6, 0, 13, 2, 5, 8, 11, 1, 1, 2, 4, 13, 3,
- 3, 12, 11, 6, 5, 8, 11, 13, 12, 0, 13, 9, 13,
- 12, 1, 7, 10, 6, 12, 8, 13, 11, 1, 3, 5, 0,
- 10, 4, 8, 15, 13, 9, 7, 2, 14, 9, 9, 10, 7,
- 13, 0, 9, 4, 7, 10, 15, 4, 10, 10, 9, 13, 8,
- 7, 10, 9, 13, 12, 14, 8, 3, 6, 4, 8, 5, 5,
- 6, 3, 6, 6, 10, 4, 3, 0, 12, 8, 7, 3, 14,
- 7, 3, 2, 3, 7, 7, 3, 0, 8, 11, 3, 14, 1,
- 13, 10, 5, 7, 9, 15, 8, 9, 1, 3, 11, 13, 13,
- 6, 0, 6, 0, 10, 0, 1, 4, 3, 11, 3, 7, 1,
- 7, 10, 7, 2, 13, 15, 12, 0, 2, 0, 6, 15, 9,
- 13, 2, 10, 2, 1, 13, 13, 7, 7, 2, 10, 1, 12,
- 9, 5, 2, 8, 11, 13, 12, 5, 3, 1, 9, 14, 12,
- 6, 12, 12, 3, 7, 0, 8, 1, 9, 12, 2, 10, 11,
- 5, 11, 10, 10, 13, 9, 3, 1, 4, 9, 6, 2, 15,
- 8, 12, 5, 14, 0, 8, 1, 3, 2, 14, 1, 6, 4,
- 4, 10, 9, 5, 15, 8, 2, 4, 3, 11, 6, 12, 6,
- 3, 14, 5, 11, 5, 9, 15, 8, 3, 5, 3, 11, 9,
- 5, 7, 14, 9, 0, 5, 11, 9, 14, 13, 2, 1, 10,
- 6, 6, 6, 15, 0, 7, 9, 12, 6, 6, 5, 0, 14,
- 15, 9, 10, 10, 13, 7, 12, 5, 13, 2, 7, 14, 7,
- 14, 13, 0, 12, 10, 7, 4, 12, 1, 8, 7, 8, 0,
- 11, 12, 12, 4, 7, 9, 15, 1, 15, 11, 7, 6, 9,
- 0, 10, 6, 7, 5, 11, 14, 13, 14, 6, 3, 0, 3,
- 6, 10, 3, 5, 0, 7, 6, 14, 15, 8, 4, 13, 11,
- 3, 1, 5, 6, 2, 14, 1, 15, 4, 4, 4, 8, 7,
- 13, 0, 8, 14, 10, 8, 14, 7, 11, 0, 2, 15, 13,
- 15, 0, 7, 8, 15, 6, 6, 4, 2, 4, 10, 13, 10,
- 6, 1, 10, 14, 13, 6, 9, 6, 8, 11, 10, 13, 2,
- 6, 10, 0, 1, 6, 15, 7, 6, 6, 13, 9, 2, 9,
- 0, 2, 15, 15, 14, 0, 2, 13, 15, 15, 0, 7, 10,
- 10, 13, 15, 6, 13, 8, 5, 4, 12, 9, 4, 14, 8,
- 6, 13, 15, 2, 8, 10, 11, 6, 11, 0, 15, 0, 1,
- 5, 1, 14, 13, 7, 2, 6, 3, 9, 7, 6, 15, 12,
- 14, 2, 10, 12, 8, 14, 5, 12, 13, 15, 10, 9, 7,
- 7, 13, 6, 11, 13, 9, 4, 8, 9, 2, 11, 13, 8,
- 1, 0, 14, 6}),
- TensorValue({1, 1, 1, 1, 8},
- dtype::QuantizedS32(0.1f * 0.2f),
- {0, 0, 0, 0, 0, 0, 0, 0}),
- TensorValue(
- {1, 1, 2, 2, 8}, dtype::QuantizedS32(0.3f),
- {0, 0, 0, 0, 0, 0, 0, 0,
- 0, 0, 0, 0, 0, 0, 0, 0,
- 0, 0, 0, 0, 0, 0, -87, 0,
- 0, 0, 0, 0, 0, 0, 0, 0}),
- {}},
- Testcase{
- {},
- {},
- {},
- {},
- TensorValue(
- {1, 1, 2, 2, 8}, dtype::QuantizedS32(0.1f * 0.2f),
- {275, -232, 55, -123, 81, -55, -324, 64,
- -104, -391, 242, -2, -162, -150, -232, -160,
- -192, -72, -52, -154, 198, -48, -1073, -105,
- 103, -218, -22, 446, -81, 90, -152, -126}),
- });
- }
-
- TEST_F(NAIVE, CONV_BIAS_QUANTIZED8x8x32_NCHW32) {
- Checker<ConvBias> checker(handle(), /* check_dispatch */ false);
- ConvBias::Param param;
- param.format = ConvBias::Param::Format::NCHW32;
-
- size_t N = 2, IC = 32, IH = 4, IW = 4, OC = 32, PH = 1, PW = 1, SH = 1,
- SW = 1, FH = 3, FW = 3;
- auto&& conv_opr = handle()->create_operator<ConvBias>();
- conv_opr->param().format = ConvBias::Param::Format::NCHW4;
- conv_opr->param().pad_h = param.pad_h = PH;
- conv_opr->param().pad_w = param.pad_w = PW;
- conv_opr->param().stride_h = param.stride_h = SH;
- conv_opr->param().stride_w = param.stride_w = SW;
- size_t OH = infer_conv_shape(IH, FH, SH, PH);
- size_t OW = infer_conv_shape(IW, FW, SW, PW);
- auto i8_min = std::numeric_limits<int8_t>().min();
- auto i8_max = std::numeric_limits<int8_t>().max();
- UniformIntRNG int_rng{i8_min, i8_max};
- TensorLayout src_layout_4{{N, IC / 4, IH, IW, 4}, dtype::QuantizedS8(0.1f)};
- TensorWrapper src_ts_4{handle(), src_layout_4};
- int_rng.gen(src_ts_4.tensornd());
-
- TensorLayout filter_layout_4{{OC, IC / 4, FH, FW, 4},
- dtype::QuantizedS8(0.2f)};
- TensorWrapper filter_ts_4{handle(), filter_layout_4};
- int_rng.gen(filter_ts_4.tensornd());
-
- TensorLayout bias_layout_4{{1, OC / 4, 1, 1, 4},
- dtype::QuantizedS32(0.02f)};
- TensorWrapper bias_ts_4{handle(), bias_layout_4};
- int_rng.gen(bias_ts_4.tensornd());
-
- TensorLayout dst_layout_4{{N, OC / 4, OH, OW, 4}, dtype::QuantizedS8(0.2f)};
- TensorWrapper dst_ts_4{handle(), dst_layout_4};
- TensorLayout z_layout_4{dst_layout_4, dtype::QuantizedS8(0.3f)};
- TensorWrapper z_ts_4{handle(), z_layout_4};
- int_rng.gen(z_ts_4.tensornd());
-
- size_t ws_size = conv_opr->get_workspace_in_bytes(
- src_layout_4, filter_layout_4, bias_layout_4, z_layout_4,
- dst_layout_4);
- WorkspaceWrapper ws{handle(), ws_size};
- conv_opr->exec(src_ts_4.tensornd(), filter_ts_4.tensornd(),
- bias_ts_4.tensornd(), z_ts_4.tensornd(), dst_ts_4.tensornd(),
- ws.workspace());
-
- TensorLayout src_layout_32{{N, IC / 32, IH, IW, 32},
- dtype::QuantizedS8(0.1f)};
- TensorWrapper src_ts_32{handle(), src_layout_32};
-
- TensorLayout filter_layout_32{{OC, IC / 32, FH, FW, 32},
- dtype::QuantizedS8(0.2f)};
- TensorWrapper filter_ts_32{handle(), filter_layout_32};
-
- TensorLayout bias_layout_32{{1, OC / 32, 1, 1, 32},
- dtype::QuantizedS32(0.02f)};
- TensorWrapper bias_ts_32{handle(), bias_layout_32};
-
- TensorLayout dst_layout_32{{N, OC / 32, OH, OW, 32},
- dtype::QuantizedS8(0.2f)};
- TensorWrapper dst_ts_32{handle(), dst_layout_32};
-
- TensorLayout z_layout_32{dst_layout_32, dtype::QuantizedS8(0.3f)};
- TensorWrapper z_ts_32{handle(), z_layout_32};
-
- auto from_nchw4_to_nchw32 = [](const TensorND in, const TensorND out) {
- size_t n = out.layout[0], c = out.layout[1], h = out.layout[2],
- w = out.layout[3];
- if (in.layout.dtype.enumv() == DTypeEnum::QuantizedS8) {
- int8_t* in_ptr = in.compatible_ptr<int8_t>();
- int8_t* out_ptr = out.compatible_ptr<int8_t>();
- for (size_t b = 0; b < n; b++) {
- for (size_t ch_out = 0; ch_out < c; ch_out++) {
- for (size_t h_ = 0; h_ < h; h_++) {
- for (size_t w_ = 0; w_ < w; w_++) {
- for (size_t ch_in = 0; ch_in < 32; ch_in++) {
- size_t ch = ch_out * 32 + ch_in;
- size_t ch_out_ = ch / 4;
- size_t ch_in_ = ch % 4;
- *out_ptr = in_ptr[b * c * h * w * 32 +
- ch_out_ * h * w * 4 +
- h_ * w * 4 + w_ * 4 + ch_in_];
- out_ptr++;
- }
- }
- }
- }
- }
- }
- if (in.layout.dtype.enumv() == DTypeEnum::QuantizedS32) {
- int32_t* in_ptr = in.compatible_ptr<int32_t>();
- int32_t* out_ptr = out.compatible_ptr<int32_t>();
- for (size_t b = 0; b < n; b++) {
- for (size_t ch_out = 0; ch_out < c; ch_out++) {
- for (size_t h_ = 0; h_ < h; h_++) {
- for (size_t w_ = 0; w_ < w; w_++) {
- for (size_t ch_in = 0; ch_in < 32; ch_in++) {
- size_t ch = ch_out * 32 + ch_in;
- size_t ch_out_ = ch / 4;
- size_t ch_in_ = ch % 4;
- *out_ptr = in_ptr[b * c * h * w * 32 +
- ch_out_ * h * w * 4 +
- h_ * w * 4 + w_ * 4 + ch_in_];
- out_ptr++;
- }
- }
- }
- }
- }
- }
- };
-
- from_nchw4_to_nchw32(src_ts_4.tensornd(), src_ts_32.tensornd());
- from_nchw4_to_nchw32(filter_ts_4.tensornd(), filter_ts_32.tensornd());
- from_nchw4_to_nchw32(bias_ts_4.tensornd(), bias_ts_32.tensornd());
- from_nchw4_to_nchw32(dst_ts_4.tensornd(), dst_ts_32.tensornd());
- from_nchw4_to_nchw32(z_ts_4.tensornd(), z_ts_32.tensornd());
-
- checker.set_param(param).exect(
- TensorNDArray{src_ts_32.tensornd(),
- filter_ts_32.tensornd(),
- bias_ts_32.tensornd(),
- z_ts_32.tensornd(), {}},
- TensorNDArray{{}, {}, {}, {}, dst_ts_32.tensornd()});
- }
-
- // vim: syntax=cpp.doxygen
|
