From caf9d39839b6afd00ab457964055082b94842f62 Mon Sep 17 00:00:00 2001 From: "joanna.wozna.intel" Date: Thu, 18 Feb 2021 10:06:52 +0100 Subject: [PATCH] Add Conv Transpose BF16 (#30877) * Add conv transpose BF16 * Share function GetWeightsTz * Adjust to review and fix op compatibility * Add bias to unique handler name * Remove errors related to paddle enforce * Add conv2d_transpose to bf16 list and kernel refator --- .../framework/ir/graph_pattern_detector.cc | 6 +- .../ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc | 2 +- .../ir/quant_conv2d_dequant_fuse_pass.cc | 2 +- .../ir_passes/tensorrt_subgraph_pass.cc | 2 +- paddle/fluid/operators/conv_transpose_op.cc | 21 +- .../fluid/operators/mkldnn/conv_mkldnn_op.cc | 20 +- .../mkldnn/conv_transpose_mkldnn_op.cc | 520 +++++++++++------- paddle/fluid/platform/mkldnn_helper.h | 13 + paddle/fluid/platform/mkldnn_reuse.h | 11 +- .../test_conv2d_transpose_bf16_mkldnn_op.py | 204 +++++++ .../mkldnn/test_conv2d_transpose_mkldnn_op.py | 7 + .../paddle/fluid/tests/unittests/op_test.py | 10 +- tools/static_mode_white_list.py | 1 + 13 files changed, 580 insertions(+), 239 deletions(-) create mode 100644 python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_bf16_mkldnn_op.py diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 2922f547278..4de75de5ebb 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -2192,9 +2192,9 @@ PDNode *patterns::Bfloat16Placement::operator()( const std::unordered_set &bfloat16_enabled_op_types) { std::unordered_set supported_op_types = std::unordered_set( - {"concat", "conv2d", "elementwise_add", "elementwise_mul", "fc", - "fusion_gru", "gelu", "layer_norm", "matmul", "pool2d", "reshape2", - "softmax", "sum", "transpose2"}); + {"concat", "conv2d", "conv2d_transpose", "elementwise_add", + "elementwise_mul", "fc", "fusion_gru", "gelu", "layer_norm", + "matmul", "pool2d", "reshape2", "softmax", "sum", "transpose2"}); if (!bfloat16_enabled_op_types.empty()) { supported_op_types = bfloat16_enabled_op_types; } diff --git a/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc index 10691ded668..c804eeb9fc3 100644 --- a/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc @@ -160,7 +160,7 @@ REGISTER_PASS(conv_transpose_bias_mkldnn_fuse_pass, REGISTER_PASS_CAPABILITY(conv_transpose_bias_mkldnn_fuse_pass) .AddCombination( paddle::framework::compatible::OpVersionComparatorCombination() - .LE("conv2d_transpose", 1) + .LE("conv2d_transpose", 2) .LE("elementwise_add", 1)); REGISTER_PASS(conv3d_bias_mkldnn_fuse_pass, diff --git a/paddle/fluid/framework/ir/quant_conv2d_dequant_fuse_pass.cc b/paddle/fluid/framework/ir/quant_conv2d_dequant_fuse_pass.cc index 64acac10186..5043fce8885 100644 --- a/paddle/fluid/framework/ir/quant_conv2d_dequant_fuse_pass.cc +++ b/paddle/fluid/framework/ir/quant_conv2d_dequant_fuse_pass.cc @@ -329,7 +329,7 @@ REGISTER_PASS_CAPABILITY(quant_conv2d_dequant_fuse_pass) paddle::framework::compatible::OpVersionComparatorCombination() .LE("conv2d", 1) .EQ("fc", 0) - .LE("conv2d_transpose", 1) + .LE("conv2d_transpose", 2) .EQ("fake_quantize_abs_max", 0) .EQ("fake_quantize_range_abs_max", 0) .EQ("fake_quantize_moving_average_abs_max", 0) diff --git a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc index d0a000fa32a..0ac2c9a9376 100644 --- a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc +++ b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc @@ -390,7 +390,7 @@ REGISTER_PASS_CAPABILITY(tensorrt_subgraph_pass) .LE("elementwise_add", 1) .LE("elementwise_mul", 1) .EQ("prelu", 0) - .LE("conv2d_transpose", 1) + .LE("conv2d_transpose", 2) .LE("leaky_relu", 1) .EQ("fc", 0) .EQ("shuffle_channel", 0) diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc index 018d15e76c9..dc4b416a609 100644 --- a/paddle/fluid/operators/conv_transpose_op.cc +++ b/paddle/fluid/operators/conv_transpose_op.cc @@ -290,6 +290,15 @@ void Conv2DTransposeOpMaker::Make() { AddAttr("use_mkldnn", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); + AddAttr("force_fp32_output", + "(bool, default false) Force BF16 kernel output FP32, only " + "used in MKL-DNN BF16") + .SetDefault(false); + AddAttr( + "mkldnn_data_type", + "(string, default \"float32\"). Data type of mkldnn kernel") + .SetDefault("float32") + .InEnum({"float32", "bfloat16"}); AddAttr("fuse_relu", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); AddAttr("fuse_activation", @@ -671,7 +680,17 @@ REGISTER_OP_VERSION(conv2d_transpose) "output_padding", "In order to add additional size to one side of each dimension " "in the output", - std::vector{})); + std::vector{})) + .AddCheckpoint( + R"ROC( + Upgrade conv2d transpose to add a new attributes [force_fp32_output, mkldnn_data_type]. + )ROC", + paddle::framework::compatible::OpVersionDesc() + .NewAttr("force_fp32_output", + "Force BF16 kernel output FP32, only used in MKL-DNN BF16", + false) + .NewAttr("mkldnn_data_type", "Data type of mkldnn kernel", + "float32")); REGISTER_OP_VERSION(conv3d_transpose) .AddCheckpoint( diff --git a/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc index 67b857aac02..fc11951d743 100644 --- a/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc @@ -33,18 +33,6 @@ using mkldnn::stream; using platform::GetMKLDNNFormat; using platform::to_void_cast; -inline void GetWeightsTz(std::vector& weights_tz, // NOLINT - const int groups) { - if (groups > 1) { - // if (is_conv3d) [o, i, d, h, w]->[g, o/g, i, d, h, w] - // else [o, i, h, w] -> [g, o/g, i, h, w] - weights_tz.push_back(0); - std::rotate(weights_tz.begin(), weights_tz.end() - 1, weights_tz.end()); - weights_tz[0] = groups; - weights_tz[1] = weights_tz[1] / groups; - } -} - inline MKLDNNMemoryFormat GetWeightsFormat(const MKLDNNMemoryFormat format, const int groups, const bool is_conv3d) { @@ -198,7 +186,7 @@ class ConvMKLDNNHandlerT const auto src_tz = paddle::framework::vectorize(input->dims()); auto weights_tz = paddle::framework::vectorize(filter->dims()); - GetWeightsTz(weights_tz, groups); + platform::GetGroupConvWeightsTz(weights_tz, groups); const auto dst_tz = paddle::framework::vectorize(output->dims()); @@ -322,7 +310,7 @@ class ConvMKLDNNHandlerT } else { const K* filter_data = filter->data(); auto weights_tz = framework::vectorize(filter->dims()); - GetWeightsTz(weights_tz, groups); + platform::GetGroupConvWeightsTz(weights_tz, groups); auto user_src_md = platform::MKLDNNMemDesc( weights_tz, platform::MKLDNNGetDataType(), @@ -640,7 +628,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto weights_tz = paddle::framework::vectorize(filter->dims()); int g = std::max(groups, 1); - GetWeightsTz(weights_tz, g); + platform::GetGroupConvWeightsTz(weights_tz, g); auto dst_tz = paddle::framework::vectorize(output->dims()); std::transform(dilations.begin(), dilations.end(), dilations.begin(), @@ -959,7 +947,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto weights_tz = paddle::framework::vectorize(filter->dims()); int g = std::max(groups, 1); - GetWeightsTz(weights_tz, g); + platform::GetGroupConvWeightsTz(weights_tz, g); auto dst_tz = paddle::framework::vectorize(output_grad->dims()); auto src_format = input->format(); diff --git a/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc index f5e62cb44ee..8d43e9f0dca 100644 --- a/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc @@ -25,245 +25,339 @@ namespace operators { using Tensor = framework::Tensor; using framework::DataLayout; -template -class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel { +inline mkldnn::memory::dims GetWeightsTz(const Tensor* filter, + const int groups) { + auto iohw_weights_tz = framework::vectorize(filter->dims()); + auto weights_tz = iohw_weights_tz; + + // IOHW -> OIHW + weights_tz[0] = iohw_weights_tz[1]; + weights_tz[1] = iohw_weights_tz[0]; + int g = std::max(groups, 1); + platform::GetGroupConvWeightsTz(weights_tz, g); + return weights_tz; +} + +template +class ConvTransposeMKLDNNHandlerT + : public platform::MKLDNNHandlerT { public: - void Compute(const paddle::framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true, - paddle::platform::errors::PreconditionNotMet( - "Operator DNNL ConvTranspose must use CPUPlace")); - const bool is_test = ctx.Attr("is_test"); - PADDLE_ENFORCE_EQ(is_test, true, - platform::errors::InvalidArgument( - "ConvTransposeMKLDNN works only for inference. " - "Set is_test = True. but got is_test=False .")); - - auto& dev_ctx = - ctx.template device_context(); - const auto& mkldnn_engine = dev_ctx.GetEngine(); - - auto* input = ctx.Input("Input"); - auto* filter = ctx.Input("Filter"); - auto* bias = ctx.HasInput("Bias") ? ctx.Input("Bias") : nullptr; - auto* output = ctx.Output("Output"); - - PADDLE_ENFORCE_EQ( - input->layout(), DataLayout::kMKLDNN, - platform::errors::InvalidArgument( - "Got wrong layout = %d for Input tensor.", input->layout())); - PADDLE_ENFORCE_NE(input->format(), MKLDNNMemoryFormat::undef, - platform::errors::InvalidArgument( - "Got wrong format for Input tensor.")); - - PADDLE_ENFORCE_EQ( - filter->layout(), DataLayout::kMKLDNN, - platform::errors::InvalidArgument( - "The filter tensor's laytout should be %d, but got %d.", - DataLayout::kMKLDNN, filter->layout())); - PADDLE_ENFORCE_NE(filter->format(), MKLDNNMemoryFormat::undef, - platform::errors::InvalidArgument( - "Got wrong formats for Filter tensor.")); - - PADDLE_ENFORCE_EQ( - input->dims().size(), 4, - platform::errors::InvalidArgument( - "Input must be with 4 dimensions, i.e. NCHW. but got dimension =%d", - input->dims().size())); - PADDLE_ENFORCE_EQ( - filter->dims().size(), 4, - platform::errors::InvalidArgument("Filter must be with 4 dimensions, " - "i.e. OIHW, but got dimension =%d", - filter->dims().size())); + ConvTransposeMKLDNNHandlerT(const framework::ExecutionContext& ctx, + const platform::MKLDNNDeviceContext& dev_ctx, + const mkldnn::engine mkldnn_engine, + platform::Place cpu_place, const Tensor* input, + const Tensor* filter, const Tensor* bias, + Tensor* output, const std::string& unique_name) + : platform::MKLDNNHandlerT( + dev_ctx, mkldnn_engine, cpu_place, + platform::CreateKey(dev_ctx, framework::vectorize(input->dims()), + unique_name)) { + if (!this->isCached()) { + const bool is_test = ctx.Attr("is_test"); + PADDLE_ENFORCE_EQ(is_test, true, + platform::errors::InvalidArgument( + "ConvTransposeMKLDNN works only for inference. " + "The attribute \'is_test\' value should be set to " + "True, but got is_test=False.")); - if (bias) { PADDLE_ENFORCE_EQ( - bias->layout(), DataLayout::kMKLDNN, + input->layout(), DataLayout::kMKLDNN, platform::errors::InvalidArgument( - "The bias tensor's laytout should be %d, but got %d.", - DataLayout::kMKLDNN, bias->layout())); - PADDLE_ENFORCE_NE(bias->format(), MKLDNNMemoryFormat::undef, + "Got wrong layout = %d for Input tensor.", input->layout())); + PADDLE_ENFORCE_NE(input->format(), MKLDNNMemoryFormat::undef, platform::errors::InvalidArgument( - "Got wrong format for Bias tensor.")); + "Got wrong format for Input tensor. The input " + "format is undefined.")); PADDLE_ENFORCE_EQ( - bias->dims().size(), 1, - platform::errors::InvalidArgument("Bias must only have 1 dimension, " - "i.e. X, but got dimension = %d .", - bias->dims().size())); - } - - std::vector strides_temp = ctx.Attr>("strides"); - std::vector strides(begin(strides_temp), end(strides_temp)); - - std::vector paddings_temp = ctx.Attr>("paddings"); - std::vector paddings(begin(paddings_temp), end(paddings_temp)); + filter->layout(), DataLayout::kMKLDNN, + platform::errors::InvalidArgument( + "The filter tensor's laytout should be %d, but got %d.", + DataLayout::kMKLDNN, filter->layout())); + PADDLE_ENFORCE_NE(filter->format(), MKLDNNMemoryFormat::undef, + platform::errors::InvalidArgument( + "Got wrong formats for Filter tensor.")); - std::vector dilations_temp = ctx.Attr>("dilations"); - std::vector dilations(begin(dilations_temp), end(dilations_temp)); + PADDLE_ENFORCE_EQ( + input->dims().size(), 4, + platform::errors::InvalidArgument("Input must be with 4 dimensions, " + "i.e. NCHW. but got dimension =%d", + input->dims().size())); + PADDLE_ENFORCE_EQ( + filter->dims().size(), 4, + platform::errors::InvalidArgument("Filter must be with 4 dimensions, " + "i.e. OIHW, but got dimension =%d", + filter->dims().size())); + + if (bias) { + PADDLE_ENFORCE_EQ( + bias->layout(), DataLayout::kMKLDNN, + platform::errors::InvalidArgument( + "The bias tensor's laytout should be %d, but got %d.", + DataLayout::kMKLDNN, bias->layout())); + PADDLE_ENFORCE_NE(bias->format(), MKLDNNMemoryFormat::undef, + platform::errors::InvalidArgument( + "Got wrong format for Bias tensor.")); + + PADDLE_ENFORCE_EQ(bias->dims().size(), 1, + platform::errors::InvalidArgument( + "Bias must only have 1 dimension, " + "i.e. X, but got dimension = %d .", + bias->dims().size())); + } - int groups = ctx.Attr("groups"); - std::string padding_algorithm = ctx.Attr("padding_algorithm"); + std::vector strides_temp = ctx.Attr>("strides"); + mkldnn::memory::dims strides(begin(strides_temp), end(strides_temp)); - PADDLE_ENFORCE_EQ( - strides.size(), 2, - platform::errors::Unimplemented( - "Now we only support 2d oneDNN convolution transpose op")); + std::vector paddings_temp = ctx.Attr>("paddings"); + mkldnn::memory::dims paddings(begin(paddings_temp), end(paddings_temp)); - auto input_dims = input->dims(); - auto data_dims = framework::slice_ddim(input_dims, 2, input_dims.size()); - auto filter_dims = filter->dims(); - auto filter_data_dims = - framework::slice_ddim(filter_dims, 2, filter_dims.size()); + std::vector dilations_temp = ctx.Attr>("dilations"); + mkldnn::memory::dims dilations(begin(dilations_temp), + end(dilations_temp)); - auto ksize = framework::vectorize(filter_data_dims); + int groups = ctx.Attr("groups"); + std::string padding_algorithm = + ctx.Attr("padding_algorithm"); - UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm, - data_dims, strides, ksize); + PADDLE_ENFORCE_EQ( + strides.size(), 2, + platform::errors::Unimplemented( + "Now we only support 2d oneDNN convolution transpose op")); + + const auto& input_dims = input->dims(); + const auto data_dims = + framework::slice_ddim(input_dims, 2, input_dims.size()); + const auto& filter_dims = filter->dims(); + const auto filter_data_dims = + framework::slice_ddim(filter_dims, 2, filter_dims.size()); + + const auto ksize = framework::vectorize(filter_data_dims); + + UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm, + data_dims, strides, ksize); + + std::transform(dilations.begin(), dilations.end(), dilations.begin(), + [](int64_t i) { return i - 1; }); + + const auto src_tz = framework::vectorize(input->dims()); + const auto weights_tz = GetWeightsTz(filter, groups); + const auto dst_tz = framework::vectorize(output->dims()); + const auto mkldnn_paddings = platform::ToMkldnnPadding(paddings); + + /* create memory descriptor for convolution without specified format + * ('any') which lets a primitive (convolution in this case) choose + * the memory format preferred for best performance + */ + const auto chosen_memory_format = MKLDNNMemoryFormat::any; + const std::string fuse_activation = + ctx.Attr("fuse_activation"); + const float fuse_alpha = ctx.Attr("fuse_alpha"); + const float fuse_beta = ctx.Attr("fuse_beta"); + + auto data_type = mkldnn::memory::data_type::f32; + if (ctx.Attr("mkldnn_data_type") == "bfloat16" || + std::is_same::value) + data_type = mkldnn::memory::data_type::bf16; + + const auto src_md = + platform::MKLDNNMemDesc(src_tz, data_type, chosen_memory_format); + const auto weights_md = + platform::MKLDNNMemDesc(weights_tz, data_type, chosen_memory_format); + const auto dst_md = platform::MKLDNNMemDesc( + dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + + const mkldnn::primitive_attr conv_trans_attr = + CreatePostOps(fuse_activation, fuse_alpha, fuse_beta); + auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference + : mkldnn::prop_kind::forward_training; + if (bias) { + std::vector bias_tz = framework::vectorize(bias->dims()); + const auto bias_md = + platform::MKLDNNMemDesc(bias_tz, data_type, MKLDNNMemoryFormat::x); + this->AcquireForwardPrimitiveDescriptor( + conv_trans_attr, fwd_prop_kind, + dnnl::algorithm::deconvolution_direct, src_md, weights_md, bias_md, + dst_md, strides, dilations, mkldnn_paddings[0], mkldnn_paddings[1]); + } else { + this->AcquireForwardPrimitiveDescriptor( + conv_trans_attr, fwd_prop_kind, + dnnl::algorithm::deconvolution_direct, src_md, weights_md, dst_md, + strides, dilations, mkldnn_paddings[0], mkldnn_paddings[1]); + } + } + } - std::transform(dilations.begin(), dilations.end(), dilations.begin(), - [](int64_t i) { return i - 1; }); + mkldnn::primitive_attr CreatePostOps(const std::string& fuse_activation, + const float& fuse_alpha, + const float& fuse_beta) { + mkldnn::primitive_attr conv_attr; + mkldnn::post_ops post_operations; + + // Fusion with ReLU layer is executed through the PostOps feature. Create a + // PostOps object and configure it to execute an eltwise relu operation. + if (fuse_activation == "relu" || fuse_activation == "leaky_relu") { + constexpr float scale = 1.0f; + post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu, + fuse_alpha, fuse_beta); + } else if (fuse_activation == "relu6") { + constexpr float scale = 1.0f; + post_operations.append_eltwise(scale, + mkldnn::algorithm::eltwise_bounded_relu, + fuse_alpha, fuse_beta); + } else if (fuse_activation == "swish") { + constexpr float scale = 1.0f; + post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_swish, + fuse_alpha, fuse_beta); + } + conv_attr.set_post_ops(post_operations); + return conv_attr; + } + std::shared_ptr AcquireSrcMemoryWithReorder( + const framework::Tensor* input) { const T* input_data = input->data(); - const T* filter_data = filter->data(); - - auto src_tz = paddle::framework::vectorize(input->dims()); - auto iohw_weights_tz = - paddle::framework::vectorize(filter->dims()); - auto weights_tz = iohw_weights_tz; - - // IOHW -> OIHW - weights_tz[0] = iohw_weights_tz[1]; - weights_tz[1] = iohw_weights_tz[0]; - - // Custom Reorder from IOHW to OIHW - auto iohw2oihw_reorder = - [&iohw_weights_tz](const T* filter_data) -> std::shared_ptr { - int o = iohw_weights_tz[1]; - int c = iohw_weights_tz[0]; - int h = iohw_weights_tz[2]; - int w = iohw_weights_tz[3]; - std::shared_ptr reordered_filter_data(new T[o * c * h * w](), - std::default_delete()); - for (int i = 0; i < c; ++i) { - for (int j = 0; j < o; ++j) { - int in_offset = j * h * w + i * o * h * w; - int out_offset = j * c * h * w + i * h * w; - std::memcpy(&(reordered_filter_data.get())[out_offset], - &filter_data[in_offset], h * w * sizeof(T)); - } + const std::string user_key_suffix{"@src_mem_p_user"}; + auto user_src_mem_p = this->AcquireMemory(user_key_suffix); + if (!user_src_mem_p) { + auto user_src_md = platform::MKLDNNMemDesc( + framework::vectorize(input->dims()), platform::MKLDNNGetDataType(), + input->format()); + return this->AcquireMemoryWithReorder( + user_src_md, this->fwd_pd_->src_desc(), + platform::to_void_cast(input_data), "@src_mem_p"); + } else { + const std::string target_key_suffix{"@src_mem_p_target"}; + const auto target_src_mem_p = this->AcquireMemory(target_key_suffix); + user_src_mem_p->set_data_handle(platform::to_void_cast(input_data)); + if (user_src_mem_p != target_src_mem_p) { + this->AcquireReorder(user_src_mem_p, target_src_mem_p, "@src_mem_p"); } + return target_src_mem_p; + } + } + + std::shared_ptr AcquireWeightsMemoryWithReorder( + const framework::Tensor* filter, const int& groups, const bool& is_test) { + // This is workaround to make execution faster, delete + // if statement after including md inside Tensor + auto weights_mem_p = this->AcquireMemory("@weights_mem_p_target"); + if (is_test && weights_mem_p) { + return weights_mem_p; + } else { + const K* filter_data = filter->data(); + auto weights_tz = GetWeightsTz(filter, groups); + int g = std::max(groups, 1); + + auto user_src_md = platform::MKLDNNMemDesc( + weights_tz, platform::MKLDNNGetDataType(), + (g == 1) ? filter->format() : MKLDNNMemoryFormat::goihw); + + auto iohw_weights_tz = framework::vectorize(filter->dims()); + // Custom Reorder from IOHW to OIHW + auto iohw2oihw_reorder = + [&iohw_weights_tz](const K* filter_data) -> std::shared_ptr { + int o = iohw_weights_tz[1]; + int c = iohw_weights_tz[0]; + int h = iohw_weights_tz[2]; + int w = iohw_weights_tz[3]; + std::shared_ptr reordered_filter_data(new K[o * c * h * w](), + std::default_delete()); + for (int i = 0; i < c; ++i) { + for (int j = 0; j < o; ++j) { + int in_offset = j * h * w + i * o * h * w; + int out_offset = j * c * h * w + i * h * w; + std::memcpy(&(reordered_filter_data.get())[out_offset], + &filter_data[in_offset], h * w * sizeof(K)); + } + } - return reordered_filter_data; - }; - - int g = std::max(groups, 1); - if (g > 1) { - int o = weights_tz[0]; - int i = weights_tz[1]; - int h = weights_tz[2]; - int w = weights_tz[3]; - weights_tz.resize(5); - weights_tz[0] = g; - weights_tz[1] = o / g; - weights_tz[2] = i; - weights_tz[3] = h; - weights_tz[4] = w; + return reordered_filter_data; + }; + + return this->template AcquireMemoryWithReorder( + user_src_md, this->fwd_pd_->weights_desc(), + platform::to_void_cast(filter_data), "@weights_mem_p", is_test, + iohw2oihw_reorder); } - auto dst_tz = paddle::framework::vectorize(output->dims()); - - // Get unique name for storing MKLDNN primitives - const std::string key = - platform::CreateKey(dev_ctx, src_tz, ctx.OutputName("Output")); - - std::vector pipeline; - - auto user_src_md = platform::MKLDNNMemDesc( - {src_tz}, platform::MKLDNNGetDataType(), input->format()); - auto user_weights_md = platform::MKLDNNMemDesc( - {weights_tz}, platform::MKLDNNGetDataType(), - (g == 1) ? MKLDNNMemoryFormat::oihw : MKLDNNMemoryFormat::goihw); - - /* create memory descriptor for convolution without specified format - * ('any') which lets a primitive (convolution in this case) choose - * the memory format preferred for best performance - */ - auto chosen_memory_format = MKLDNNMemoryFormat::any; - std::string fuse_activation = ctx.Attr("fuse_activation"); - float fuse_alpha = ctx.Attr("fuse_alpha"); - float fuse_beta = ctx.Attr("fuse_beta"); - - auto src_md = platform::MKLDNNMemDesc( - src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); - auto weights_md = platform::MKLDNNMemDesc( - weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); - std::vector bias_tz; - auto dst_md = platform::MKLDNNMemDesc( - dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); - - platform::ConvTransposeMKLDNNHandler handler(dev_ctx, mkldnn_engine, key); - // create a deconv(conv transpose) primitive descriptor and save it for - // usage in backward - std::shared_ptr - conv_transpose_pd; - auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference - : mkldnn::prop_kind::forward_training; - if (bias) { - bias_tz = paddle::framework::vectorize(bias->dims()); - auto bias_md = platform::MKLDNNMemDesc( - bias_tz, platform::MKLDNNGetDataType(), MKLDNNMemoryFormat::x); - conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor( - src_md, weights_md, bias_md, dst_md, strides, dilations, paddings, - mkldnn_engine, fuse_activation, fuse_alpha, fuse_beta, false, - fwd_prop_kind); + } + + std::shared_ptr AcquireBiasMemoryWithReorder( + const framework::Tensor* bias, const bool& is_test) { + auto bias_mem_p = this->AcquireMemory("@bias_mem_p_target"); + if (is_test && bias_mem_p) { + return bias_mem_p; } else { - conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor( - src_md, weights_md, boost::none, dst_md, strides, dilations, paddings, - mkldnn_engine, fuse_activation, fuse_alpha, fuse_beta, false, - fwd_prop_kind); + const K* bias_data = bias->data(); + auto user_bias_md = platform::MKLDNNMemDesc( + framework::vectorize(bias->dims()), platform::MKLDNNGetDataType(), + MKLDNNMemoryFormat::x); + return this->AcquireMemoryWithReorder( + user_bias_md, this->fwd_pd_->bias_desc(), + platform::to_void_cast(bias_data), "@bias_mem_p", is_test); } + } +}; - // create mkldnn memory from input tensors (data/weights) - auto user_src_memory_p = handler.AcquireSrcMemory( - user_src_md, platform::to_void_cast(input_data)); - auto user_weights_memory_p = handler.AcquireWeightsMemory( - user_weights_md, platform::to_void_cast(filter_data), - is_test ? iohw2oihw_reorder : platform::user_function()); +template +class ConvTransposeMKLDNNOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true, + platform::errors::PreconditionNotMet( + "Operator DNNL ConvTranspose must use CPUPlace")); + const bool is_bfloat16 = + ctx.Attr("mkldnn_data_type") == "bfloat16"; + const bool force_fp32_output = ctx.Attr("force_fp32_output"); + if (is_bfloat16) { + if (force_fp32_output) + Execute(ctx); + else + Execute(ctx); + } else { + Execute(ctx); + } + } - // create reorder primitive if the input format is not the preferred one - auto src_memory_p = - handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); - auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( - user_weights_memory_p, pipeline, is_test); + template + void Execute(const framework::ExecutionContext& ctx) const { + auto& dev_ctx = + ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); - auto output_data = - output->mutable_data(ctx.GetPlace(), handler.GetDstMemorySize()); - auto dst_memory_p = handler.AcquireDstMemoryFromPrimitive( - platform::to_void_cast(output_data)); + const bool is_test = ctx.Attr("is_test"); - auto conv_p = handler.AcquireConvolution(); + const auto* input = ctx.Input("Input"); + const auto* filter = ctx.Input("Filter"); + const auto* bias = + ctx.HasInput("Bias") ? ctx.Input("Bias") : nullptr; + auto* output = ctx.Output("Output"); + const std::string unique_name = ctx.InputName("Input") + + ctx.InputName("Filter") + + (bias ? ctx.InputName("Bias") : ""); + ConvTransposeMKLDNNHandlerT handler( + ctx, dev_ctx, mkldnn_engine, ctx.GetPlace(), input, filter, bias, + output, unique_name); + auto src_memory_p = handler.AcquireSrcMemoryWithReorder(input); + auto weights_memory_p = handler.AcquireWeightsMemoryWithReorder( + filter, ctx.Attr("groups"), is_test); + + std::shared_ptr dst_memory_p = + handler.template AcquireDstMemory(output); + auto conv_p = handler.AcquireForwardPrimitive(); + + std::unordered_map args = { + {MKLDNN_ARG_SRC, *src_memory_p}, + {MKLDNN_ARG_WEIGHTS, *weights_memory_p}, + {MKLDNN_ARG_DST, *dst_memory_p}}; - auto& astream = platform::MKLDNNDeviceContext::tls().get_stream(); if (bias) { - const T* bias_data = bias->data(); - auto user_bias_md = platform::MKLDNNMemDesc( - {bias_tz}, platform::MKLDNNGetDataType(), MKLDNNMemoryFormat::x); - auto user_bias_memory_p = handler.AcquireBiasMemory( - user_bias_md, platform::to_void_cast(bias_data)); - - auto bias_memory_p = - handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline); - - conv_p->execute(astream, {{MKLDNN_ARG_SRC, *src_memory_p}, - {MKLDNN_ARG_WEIGHTS, *weights_memory_p}, - {MKLDNN_ARG_BIAS, *bias_memory_p}, - {MKLDNN_ARG_DST, *dst_memory_p}}); - } else { - conv_p->execute(astream, {{MKLDNN_ARG_SRC, *src_memory_p}, - {MKLDNN_ARG_WEIGHTS, *weights_memory_p}, - {MKLDNN_ARG_DST, *dst_memory_p}}); + auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(bias, is_test); + args.insert({MKLDNN_ARG_BIAS, *bias_memory_p}); } + auto& astream = platform::MKLDNNDeviceContext::tls().get_stream(); + conv_p->execute(astream, args); astream.wait(); - output->set_layout(DataLayout::kMKLDNN); output->set_format(platform::GetMKLDNNFormat(*dst_memory_p)); } @@ -274,5 +368,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel { namespace ops = paddle::operators; -REGISTER_OP_KERNEL(conv2d_transpose, MKLDNN, ::paddle::platform::CPUPlace, - ops::ConvTransposeMKLDNNOpKernel); +REGISTER_OP_KERNEL( + conv2d_transpose, MKLDNN, ::paddle::platform::CPUPlace, + ops::ConvTransposeMKLDNNOpKernel, + ops::ConvTransposeMKLDNNOpKernel); diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index 79c536508da..20e6dfe1c39 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -492,6 +492,19 @@ inline std::vector> ToMkldnnPadding( } } +// The function adjusts the vector of weight dimensions for group convolutions +inline void GetGroupConvWeightsTz(std::vector& weights_tz, // NOLINT + const int groups) { + if (groups > 1) { + // if (is_conv3d) [o, i, d, h, w]->[g, o/g, i, d, h, w] + // else [o, i, h, w] -> [g, o/g, i, h, w] + weights_tz.push_back(0); + std::rotate(weights_tz.begin(), weights_tz.end() - 1, weights_tz.end()); + weights_tz[0] = groups; + weights_tz[1] = weights_tz[1] / groups; + } +} + inline bool HasOpINT8DataType(const paddle::framework::OpDesc* op) { return (op->GetAttrIfExists("mkldnn_data_type") == "int8" || op->GetAttrIfExists("use_quantizer")); diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h index 2cff67670f6..3e02a8672c3 100644 --- a/paddle/fluid/platform/mkldnn_reuse.h +++ b/paddle/fluid/platform/mkldnn_reuse.h @@ -250,10 +250,12 @@ class MKLDNNHandlerT { astream.wait(); } + template std::shared_ptr AcquireMemoryWithReorder( const mkldnn::memory::desc& user_md, const mkldnn::memory::desc& target_md, void* ptr, - const std::string& suffix, bool is_persistent = false) { + const std::string& suffix, bool is_persistent = false, + std::function(const F*)> custom_reorder_func = {}) { const auto target_key = key_ + suffix + "_target"; const auto key_reorder_p = key_ + suffix + "reorder_p"; const auto user_key = key_ + suffix + "_user"; @@ -262,6 +264,12 @@ class MKLDNNHandlerT { std::static_pointer_cast(dev_ctx_.GetBlob(target_key)); if (target_memory_p == nullptr) { + if (custom_reorder_func) { + auto reordered_data = + custom_reorder_func(reinterpret_cast(ptr)); + dev_ctx_.SetBlob(key_reorder_p + "-custom_reorder", reordered_data); + ptr = reinterpret_cast(reordered_data.get()); + } auto user_memory_p = std::make_shared(user_md, engine_, ptr); if (user_md != target_md) { @@ -1288,6 +1296,5 @@ static void SetDstMemoryQuantized( dst_memory.reset( new mkldnn::memory(*dst_md, engine, to_void_cast(output_data))); } - } // namespace platform } // namespace paddle diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_bf16_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_bf16_mkldnn_op.py new file mode 100644 index 00000000000..c6b7c175d90 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_bf16_mkldnn_op.py @@ -0,0 +1,204 @@ +# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# 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. + +from __future__ import print_function + +import unittest +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16 + +from paddle.fluid.tests.unittests.test_conv2d_transpose_op import conv2dtranspose_forward_naive +from paddle import enable_static + + +def conv2d_bias_naive(out, bias): + _, out_c, _, _ = out.shape + + for l in range(out_c): + out[:, l, :, :] = out[:, l, :, :] + bias[l] + return out + + +@unittest.skipIf(not core.supports_bfloat16(), + "place does not support BF16 evaluation") +class TestConv2DTransposeBF16MKLDNNOp(OpTest): + def test_check_output(self): + self.check_output_with_place(core.CPUPlace()) + + def test_check_grad(self): + pass + + def test_check_grad_no_input(self): + pass + + def test_check_grad_no_filter(self): + pass + + def init_op_type(self): + self.data_format = "NCHW" + self.op_type = 'conv2d_transpose' + self._cpu_only = True + + def init_test_case(self): + self.pad = [0, 0] + self.fuse_bias = False + self.use_mkldnn = True + self.is_test = True + self.bias_size = None + self.fuse_activation = "" + self.fuse_alpha = 0.0 + self.fuse_beta = 0.0 + self.stride = [1, 1] + self.dilations = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + self.groups = 1 + self.output_size = None + self.output_padding = [] + self.data_format = "NCHW" + self.pad = [0, 0] + self.padding_algorithm = "EXPLICIT" + self.force_fp32_output = False + + def setUp(self): + self.input_type = np.uint16 + self.dtype = np.uint16 + self.mkldnn_data_type = "bfloat16" + self.init_op_type() + self.init_test_case() + + input = np.random.random(self.input_size).astype(np.float32) + filter = np.random.random(self.filter_size).astype(np.float32) + + self.attrs = { + 'strides': self.stride, + 'paddings': self.pad, + 'padding_algorithm': self.padding_algorithm, + 'groups': self.groups, + 'dilations': self.dilations, + 'is_test': self.is_test, + 'use_mkldnn': self.use_mkldnn, + 'mkldnn_data_type': self.mkldnn_data_type, + 'force_fp32_output': self.force_fp32_output, + 'data_format': self.data_format, + 'fuse_activation': self.fuse_activation, + 'fuse_alpha': self.fuse_alpha, + 'fuse_beta': self.fuse_beta + } + if self.output_size is not None: + self.attrs['output_size'] = self.output_size + + if len(self.output_padding) > 0: + self.attrs['output_padding'] = self.output_padding + + output = conv2dtranspose_forward_naive(input, filter, + self.attrs).astype(np.float32) + + if self.input_type is not np.float32: + input = convert_float_to_uint16(input) + + self.inputs = { + 'Input': input.view(self.input_type), + 'Filter': OpTest.np_dtype_to_fluid_dtype(filter) + } + + if self.fuse_bias and self.bias_size is not None: + bias = np.random.random(self.bias_size).astype(np.float32) + output = conv2d_bias_naive(output, bias) + output = output.astype(np.float32) + self.attrs['fuse_bias'] = self.fuse_bias + self.inputs['Bias'] = OpTest.np_dtype_to_fluid_dtype(bias) + + if self.fuse_activation == "relu": + output = np.maximum(output, 0).astype(np.float32) + output = output.astype(np.float32) + + if not self.force_fp32_output: + output = convert_float_to_uint16(output, self.attrs['data_format']) + + self.outputs['Output'] = output + + +class TestMKLDNNFuseBias(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNFuseBias, self).init_test_case() + self.pad = [1, 1] + self.fuse_bias = True + self.bias_size = [6] + + +class TestMKLDNNWithPad(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNWithPad, self).init_test_case() + self.pad = [1, 1] + self.input_size = [2, 3, 10, 10] + + +class TestMKLDNNWithStride(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNWithStride, self).init_test_case() + self.pad = [1, 1] + self.stride = [2, 2] + self.input_size = [2, 3, 6, 6] # NCHW + + +class TestMKLDNNWithAsymPad(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNWithAsymPad, self).init_test_case() + self.pad = [0, 0, 1, 2] + self.padding_algorithm = "EXPLICIT" + + +class TestMKLDNNWithSamePad(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNWithSamePad, self).init_test_case() + self.pad = [0, 0] + self.padding_algorithm = "SAME" + + +class TestMKLDNNWithValidPad(TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestMKLDNNWithValidPad, self).init_test_case() + self.pad = [1, 1] + self.padding_algorithm = "VALID" + + +class TestMKLDNNWithValidPad_NHWC(TestMKLDNNWithValidPad): + def init_test_case(self): + super(TestMKLDNNWithValidPad_NHWC, self).init_test_case() + self.data_format = 'NHWC' + N, C, H, W = self.input_size + self.input_size = [N, H, W, C] + + +class TestConv2DTransposeMKLDNNWithDilationsExplicitPad( + TestConv2DTransposeBF16MKLDNNOp): + def init_test_case(self): + super(TestConv2DTransposeMKLDNNWithDilationsExplicitPad, + self).init_test_case() + self.stride = [2, 1] + self.dilations = [1, 2] + self.groups = 1 + self.input_size = [4, 3, 8, 7] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 4, 3] + self.pad = [1, 3, 2, 1] + self.padding_algorithm = "EXPLICIT" + + +if __name__ == '__main__': + enable_static() + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py index 7da274917a5..f31ddf921f8 100644 --- a/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py @@ -82,6 +82,8 @@ class TestConv2DTransposeMKLDNNOp(TestConv2DTransposeOp): self.attrs['fuse_activation'] = self.fuse_activation self.attrs['fuse_alpha'] = self.fuse_alpha self.attrs['fuse_beta'] = self.fuse_beta + self.attrs['mkldnn_data_type'] = 'float32' + self.attrs['force_fp32_output'] = False self.outputs['Output'] = output @@ -150,3 +152,8 @@ class TestConv2DTransposeMKLDNNWithDilationsExplicitPad( self.filter_size = [f_c, 6, 4, 3] self.pad = [1, 3, 2, 1] self.padding_algorithm = "EXPLICIT" + + +if __name__ == '__main__': + enable_static() + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index e3e84a73301..8bb0779bc04 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -221,12 +221,18 @@ def copy_bits_from_float_to_uint16(f): return struct.unpack('> 16 -def convert_float_to_uint16(float_list): +def convert_float_to_uint16(float_list, data_format="NCHW"): + if data_format == "NHWC": + float_list = np.transpose(float_list, [0, 3, 1, 2]) + new_output = [] for x in np.nditer(float_list): new_output.append(np.uint16(copy_bits_from_float_to_uint16(x))) + new_output = np.reshape(new_output, float_list.shape).view(np.uint16) - return np.reshape(new_output, float_list.shape).view(np.uint16) + if data_format == "NHWC": + new_output = np.transpose(new_output, [0, 2, 3, 1]) + return new_output class OpTest(unittest.TestCase): diff --git a/tools/static_mode_white_list.py b/tools/static_mode_white_list.py index 0c36d0cda3f..872fd857381 100644 --- a/tools/static_mode_white_list.py +++ b/tools/static_mode_white_list.py @@ -590,6 +590,7 @@ STATIC_MODE_TESTING_LIST = [ 'test_conv2d_int8_mkldnn_op', 'test_conv2d_mkldnn_op', 'test_conv2d_transpose_mkldnn_op', + 'test_conv2d_transpose_bf16_mkldnn_op', 'test_conv3d_mkldnn_op', 'test_dequantize_mkldnn_op', 'test_elementwise_add_mkldnn_op', -- GitLab