Optimize CPU time of JPEG lossless decoder (#4625)

Signed-off-by: Joaquin Anton <janton@nvidia.com>
NVIDIA · Feb 1, 2023 · 9f0f7e0 · 9f0f7e0
1 parent c36e5a1
commit 9f0f7e0
Show file tree

Hide file tree

Showing 5 changed files with 208 additions and 82 deletions.
diff --git a/dali/imgcodec/decoders/nvjpeg/nvjpeg.cc b/dali/imgcodec/decoders/nvjpeg/nvjpeg.cc
@@ -1,4 +1,4 @@
-// Copyright (c) 2019-2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright (c) 2019-2023, NVIDIA CORPORATION & AFFILIATES. 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.
@@ -144,6 +144,19 @@ NvJpegDecoderInstance::PerThreadResources::~PerThreadResources() {
   }
 }
 
+bool NvJpegDecoderInstance::CanDecode(DecodeContext ctx, ImageSource *in, DecodeParams opts,
+                                      const ROI &roi) {
+  JpegParser jpeg_parser{};
+  if (!jpeg_parser.CanParse(in))
+    return false;
+
+  // This decoder does not support SOF-3 (JPEG lossless) samples
+  auto ext_info = jpeg_parser.GetExtendedInfo(in);
+  std::array<uint8_t, 2> sof3_marker = {0xff, 0xc3};
+  bool is_lossless_jpeg = ext_info.sof_marker == sof3_marker;
+  return !is_lossless_jpeg;
+}
+
 bool NvJpegDecoderInstance::SetParam(const char *name, const any &value) {
   if (strcmp(name, "device_memory_padding") == 0) {
     device_memory_padding_ = any_cast<size_t>(value);

diff --git a/dali/imgcodec/decoders/nvjpeg/nvjpeg.h b/dali/imgcodec/decoders/nvjpeg/nvjpeg.h
@@ -1,4 +1,4 @@
-// Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. 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.
@@ -33,6 +33,9 @@ class DLL_PUBLIC NvJpegDecoderInstance : public BatchParallelDecoderImpl {
  public:
   explicit NvJpegDecoderInstance(int device_id, const std::map<std::string, any> &params);
 
+  using BatchParallelDecoderImpl::CanDecode;
+  bool CanDecode(DecodeContext ctx, ImageSource *in, DecodeParams opts, const ROI &roi) override;
+
   // NvjpegDecoderInstance has to operate on its own thread pool instead of the
   // one passed by the DecodeContext. Overriding thread pool pointer caried in
   // the context argument of this variant of

diff --git a/dali/imgcodec/decoders/nvjpeg_lossless/nvjpeg_lossless.cc b/dali/imgcodec/decoders/nvjpeg_lossless/nvjpeg_lossless.cc
@@ -32,115 +32,190 @@ NvJpegLosslessDecoderInstance::NvJpegLosslessDecoderInstance(
     : BatchedApiDecoderImpl(device_id, params),
       event_(CUDAEvent::Create(device_id)) {
   DeviceGuard dg(device_id_);
+  // TODO(janton): use custom allocators (?)
   CUDA_CALL(nvjpegCreateEx(NVJPEG_BACKEND_LOSSLESS_JPEG, NULL, NULL, 0, &nvjpeg_handle_));
-  CUDA_CALL(nvjpegJpegStreamCreate(nvjpeg_handle_, &jpeg_stream_));
+  per_thread_resources_.push_back(PerThreadResources{nvjpeg_handle_});
   CUDA_CALL(nvjpegJpegStateCreate(nvjpeg_handle_, &state_));
 }
 
 NvJpegLosslessDecoderInstance::~NvJpegLosslessDecoderInstance() {
   DeviceGuard dg(device_id_);
   CUDA_CALL(cudaEventSynchronize(event_));
-  CUDA_CALL(nvjpegJpegStreamDestroy(jpeg_stream_));
+  per_thread_resources_.clear();
   CUDA_CALL(nvjpegJpegStateDestroy(state_));
   CUDA_CALL(nvjpegDestroy(nvjpeg_handle_));
 }
 
+NvJpegLosslessDecoderInstance::PerThreadResources::PerThreadResources(nvjpegHandle_t handle) {
+  CUDA_CALL(nvjpegJpegStreamCreate(handle, &jpeg_stream));
+}
+
+NvJpegLosslessDecoderInstance::PerThreadResources::PerThreadResources(PerThreadResources&& other)
+  : jpeg_stream(other.jpeg_stream) {
+  other.jpeg_stream = nullptr;
+}
+
+NvJpegLosslessDecoderInstance::PerThreadResources::~PerThreadResources() {
+  if (jpeg_stream) {
+    CUDA_CALL(nvjpegJpegStreamDestroy(jpeg_stream));
+  }
+}
+
 bool NvJpegLosslessDecoderInstance::CanDecode(DecodeContext ctx, ImageSource *in, DecodeParams opts,
                                               const ROI &roi) {
   if (opts.format != DALI_ANY_DATA && opts.format != DALI_GRAY) {
     return false;
   }
 
-  try {
-    CUDA_CALL(nvjpegJpegStreamParseHeader(nvjpeg_handle_, in->RawData<unsigned char>(), in->Size(),
-                                          jpeg_stream_));
-
-    int is_supported = 0;  // 0 means yes
-    CUDA_CALL(nvjpegDecodeBatchedSupported(nvjpeg_handle_, jpeg_stream_, &is_supported));
-    return is_supported == 0;
-  } catch (...) {
+  JpegParser jpeg_parser{};
+  if (!jpeg_parser.CanParse(in))
     return false;
+
+  // This decoder only supports SOF-3 (JPEG lossless) samples
+  auto ext_info = jpeg_parser.GetExtendedInfo(in);
+  std::array<uint8_t, 2> sof3_marker = {0xff, 0xc3};
+  bool is_lossless_jpeg = ext_info.sof_marker == sof3_marker;
+  return is_lossless_jpeg;
+}
+
+void NvJpegLosslessDecoderInstance::Parse(DecodeResultsPromise &promise,
+                                          DecodeContext ctx,
+                                          cspan<ImageSource *> in,
+                                          DecodeParams opts,
+                                          cspan<ROI> rois) {
+  int nsamples = in.size();
+  assert(rois.empty() || rois.size() == nsamples);
+  assert(ctx.tp != nullptr);
+  int nthreads = ctx.tp->NumThreads();
+  if (nthreads > static_cast<int>(per_thread_resources_.size())) {
+    per_thread_resources_.reserve(nthreads);
+    for (int i = per_thread_resources_.size(); i < nthreads; i++)
+      per_thread_resources_.emplace_back(nvjpeg_handle_);
+  }
+  sample_meta_.clear();
+  sample_meta_.resize(nsamples);
+
+  // temporary solution: Just to check when the parsing has finished
+  DecodeResultsPromise parse_promise(nsamples);
+  for (int i = 0; i < nsamples; i++) {
+    int tid = 0;
+    ctx.tp->AddWork(
+        [&, i](int tid) {
+          auto &jpeg_stream = per_thread_resources_[tid].jpeg_stream;
+          auto *sample = in[i];
+          auto &meta = sample_meta_[i];
+          try {
+            CUDA_CALL(nvjpegJpegStreamParseHeader(nvjpeg_handle_, sample->RawData<unsigned char>(),
+                                                  sample->Size(), jpeg_stream));
+            int is_supported = 0;  // 0 means yes
+            CUDA_CALL(nvjpegDecodeBatchedSupported(nvjpeg_handle_, jpeg_stream, &is_supported));
+            meta.can_decode = (is_supported == 0);
+            if (!meta.can_decode) {
+              promise.set(i, {false, nullptr});
+              parse_promise.set(i, {false, nullptr});
+              return;
+            }
+            meta.needs_processing = opts.dtype != DALI_UINT16;
+            meta.needs_processing |= !rois.empty() && rois[i].use_roi();
+            if (opts.use_orientation) {
+              auto &ori = meta.orientation = JpegParser().GetInfo(sample).orientation;
+              meta.needs_processing |= (ori.rotate || ori.flip_x || ori.flip_y);
+            }
+
+            unsigned int precision;
+            CUDA_CALL(nvjpegJpegStreamGetSamplePrecision(jpeg_stream, &precision));
+            meta.dyn_range_multiplier = 1.0f;
+            if (NeedDynamicRangeScaling(precision, DALI_UINT16)) {
+              meta.dyn_range_multiplier = DynamicRangeMultiplier(precision, DALI_UINT16);
+              meta.needs_processing = true;
+            }
+            parse_promise.set(i, {true, nullptr});
+          } catch (...) {
+            meta.can_decode = false;
+            promise.set(i, {false, std::current_exception()});
+            parse_promise.set(i, {false, nullptr});
+          }
+        }, in[i]->Size());
   }
+  ctx.tp->RunAll(false);
+  parse_promise.get_future().wait_all();
+
+  batch_sz_ = 0;
+  for (auto &meta : sample_meta_) {
+    if (meta.can_decode)
+      meta.idx_in_batch = batch_sz_++;
+  }
+}
+
+void NvJpegLosslessDecoderInstance::RunDecode(kernels::DynamicScratchpad& s,
+                                              DecodeContext ctx,
+                                              span<SampleView<GPUBackend>> out,
+                                              cspan<ImageSource *> in,
+                                              DecodeParams opts,
+                                              cspan<ROI> rois) {
+  if (batch_sz_ <= 0)
+    return;
+  int nsamples = in.size();
+  encoded_.clear();
+  encoded_.resize(batch_sz_);
+  encoded_len_.clear();
+  encoded_len_.resize(batch_sz_);
+  decoded_.clear();
+  decoded_.resize(batch_sz_);
+  for (int i = 0; i < nsamples; i++) {
+    auto &meta = sample_meta_[i];
+    if (!meta.can_decode)
+      continue;
+    auto *sample = in[i];
+    auto &out_sample = out[i];
+    auto roi = rois.empty() ? ROI{} : rois[i];
+    assert(sample->Kind() == InputKind::HostMemory);
+    encoded_[meta.idx_in_batch] = sample->RawData<unsigned char>();
+    encoded_len_[meta.idx_in_batch] = sample->Size();
+    auto &o = decoded_[meta.idx_in_batch];
+    auto sh = out_sample.shape();
+    o.pitch[0] = sh[1] * sh[2] * sizeof(uint16_t);
+    if (meta.needs_processing) {
+      int64_t nbytes = volume(sh) * sizeof(uint16_t);
+      o.channel[0] = s.Allocate<mm::memory_kind::device, uint8_t>(nbytes);
+    } else {
+      o.channel[0] = static_cast<uint8_t *>(out_sample.raw_mutable_data());
+    }
+  }
+
+  CUDA_CALL(nvjpegDecodeBatchedInitialize(nvjpeg_handle_, state_, batch_sz_, 1,
+                                          NVJPEG_OUTPUT_UNCHANGEDI_U16));
+  CUDA_CALL(nvjpegDecodeBatched(nvjpeg_handle_, state_, encoded_.data(), encoded_len_.data(),
+                                decoded_.data(), ctx.stream));
 }
 
+
 FutureDecodeResults NvJpegLosslessDecoderInstance::ScheduleDecode(DecodeContext ctx,
                                                                   span<SampleView<GPUBackend>> out,
                                                                   cspan<ImageSource *> in,
                                                                   DecodeParams opts,
                                                                   cspan<ROI> rois) {
   int nsamples = in.size();
   assert(out.size() == nsamples);
-  assert(rois.empty() || rois.size() == nsamples);
-  assert(ctx.tp != nullptr);
 
   DecodeResultsPromise promise(nsamples);
   auto set_promise = [&](DecodeResult result) {
     for (int i = 0; i < nsamples; i++)
       promise.set(i, result);
   };
 
+  if (opts.format != DALI_ANY_DATA && opts.format != DALI_GRAY)
+    set_promise({false, std::make_exception_ptr(
+                            std::invalid_argument("Only ANY_DATA and GRAY are supported."))});
+
+  // scratchpad should not go out of scope until we launch the postprocessing
   kernels::DynamicScratchpad s({}, ctx.stream);
+  Parse(promise, ctx, in, opts, rois);
   try {
-    if (opts.format != DALI_ANY_DATA && opts.format != DALI_GRAY)
-      throw std::invalid_argument("Only ANY_DATA and GRAY are supported.");
-
-    sample_meta_.clear();
-    sample_meta_.resize(nsamples);
-    encoded_.clear();
-    encoded_.resize(nsamples);
-    encoded_len_.clear();
-    encoded_len_.resize(nsamples);
-    decoded_.clear();
-    decoded_.resize(nsamples);
-
-    for (int i = 0; i < nsamples; i++) {
-      auto *sample = in[i];
-      auto &out_sample = out[i];
-      assert(sample->Kind() == InputKind::HostMemory);
-      auto *data_ptr = sample->RawData<unsigned char>();
-      auto data_size = sample->Size();
-      encoded_[i] = data_ptr;
-      encoded_len_[i] = data_size;
-      sample_meta_[i].needs_processing = opts.dtype != DALI_UINT16;
-      if (!rois.empty() && rois[i].use_roi()) {
-        sample_meta_[i].needs_processing = true;
-      }
-      if (opts.use_orientation) {
-        auto &ori = sample_meta_[i].orientation = JpegParser().GetInfo(in[i]).orientation;
-        if (ori.rotate || ori.flip_x || ori.flip_y)
-          sample_meta_[i].needs_processing = true;
-      }
-
-      CUDA_CALL(nvjpegJpegStreamParseHeader(nvjpeg_handle_, sample->RawData<unsigned char>(),
-                                            sample->Size(), jpeg_stream_));
-      unsigned int precision;
-      CUDA_CALL(nvjpegJpegStreamGetSamplePrecision(jpeg_stream_, &precision));
-      sample_meta_[i].dyn_range_multiplier = 1.0f;
-      if (NeedDynamicRangeScaling(precision, DALI_UINT16)) {
-        sample_meta_[i].dyn_range_multiplier = DynamicRangeMultiplier(precision, DALI_UINT16);
-        sample_meta_[i].needs_processing = true;
-      }
-
-      auto &o = decoded_[i];
-      auto sh = out_sample.shape();
-      o.pitch[0] = sh[1] * sh[2] * sizeof(uint16_t);
-      if (sample_meta_[i].needs_processing) {
-        int64_t nbytes = volume(sh) * sizeof(uint16_t);
-        o.channel[0] = s.Allocate<mm::memory_kind::device, uint8_t>(nbytes);
-      } else {
-        o.channel[0] = static_cast<uint8_t *>(out_sample.raw_mutable_data());
-      }
-    }
-
-    CUDA_CALL(nvjpegDecodeBatchedInitialize(nvjpeg_handle_, state_, nsamples, 1,
-                                            NVJPEG_OUTPUT_UNCHANGEDI_U16));
-    CUDA_CALL(nvjpegDecodeBatched(nvjpeg_handle_, state_, encoded_.data(), encoded_len_.data(),
-                                  decoded_.data(), ctx.stream));
-  } catch (...) {
+    RunDecode(s, ctx, out, in, opts, rois);
+  } catch(...) {
     set_promise({false, std::current_exception()});
-    return promise.get_future();
   }
-
   Postprocess(promise, ctx, out, opts, rois);
   CUDA_CALL(cudaEventRecord(event_, ctx.stream));
   return promise.get_future();
@@ -149,27 +224,34 @@ FutureDecodeResults NvJpegLosslessDecoderInstance::ScheduleDecode(DecodeContext
 void NvJpegLosslessDecoderInstance::Postprocess(DecodeResultsPromise &promise, DecodeContext ctx,
                                                 span<SampleView<GPUBackend>> out, DecodeParams opts,
                                                 cspan<ROI> rois) {
+  if (batch_sz_ <= 0)
+    return;
   int nsamples = out.size();
-  for (int i = 0; i < nsamples; i++) {
-    if (!sample_meta_[i].needs_processing) {
+  for (int i = 0, j = 0; i < nsamples; i++) {
+    const auto &meta = sample_meta_[i];
+    if (!meta.can_decode)
+      continue;  // we didn't try to decode this sample
+
+    const auto &decoded = decoded_[j++];  // decoded only has samples where can_decode == true
+    if (!meta.needs_processing) {
       promise.set(i, {true, nullptr});
       continue;
     }
     auto sh = out[i].shape();
-    SampleView<GPUBackend> decoded_view(decoded_[i].channel[0], sh, DALI_UINT16);
+    auto roi = rois.empty() ? ROI{} : rois[i];
+    SampleView<GPUBackend> decoded_view(decoded.channel[0], sh, DALI_UINT16);
     DALIImageType decoded_format = sh[2] == 1 ? DALI_GRAY : DALI_ANY_DATA;
     try {
       Convert(out[i], "HWC", opts.format, decoded_view, "HWC", decoded_format,
-              ctx.stream, rois.empty() ? ROI{} : rois[i], sample_meta_[i].orientation,
-              sample_meta_[i].dyn_range_multiplier);
+              ctx.stream, roi, meta.orientation, meta.dyn_range_multiplier);
       promise.set(i, {true, nullptr});
     } catch (...) {
       promise.set(i, {false, std::current_exception()});
     }
   }
 }
 
-REGISTER_DECODER("JPEG", NvJpegLosslessDecoderFactory, CUDADecoderPriority - 1);
+REGISTER_DECODER("JPEG", NvJpegLosslessDecoderFactory, CUDADecoderPriority + 1);
 
 }  // namespace imgcodec
 }  // namespace dali
diff --git a/dali/imgcodec/decoders/nvjpeg_lossless/nvjpeg_lossless.h b/dali/imgcodec/decoders/nvjpeg_lossless/nvjpeg_lossless.h
@@ -24,6 +24,7 @@
 #include "dali/core/cuda_stream_pool.h"
 #include "dali/imgcodec/decoders/decoder_batched_api_impl.h"
 #include "dali/imgcodec/decoders/nvjpeg/nvjpeg_memory.h"
+#include "dali/kernels/dynamic_scratchpad.h"
 #include "dali/pipeline/data/buffer.h"
 
 namespace dali {
@@ -42,20 +43,37 @@ class DLL_PUBLIC NvJpegLosslessDecoderInstance : public BatchedApiDecoderImpl {
                                      cspan<ROI> rois = {}) override;
 
  private:
+  // Parses encoded streams and populates SampleMeta, and batch_sz_
+  void Parse(DecodeResultsPromise &promise, DecodeContext ctx, cspan<ImageSource *> in,
+             DecodeParams opts, cspan<ROI> rois);
+
+  // Invokes nvJPEG decoding (sample_meta_ and batch_sz_ to be populated)
+  void RunDecode(kernels::DynamicScratchpad &s, DecodeContext ctx, span<SampleView<GPUBackend>> out,
+                 cspan<ImageSource *> in, DecodeParams opts, cspan<ROI> rois = {});
+
   void Postprocess(DecodeResultsPromise &promise, DecodeContext ctx,
                    span<SampleView<GPUBackend>> out, DecodeParams opts, cspan<ROI> rois);
 
   nvjpegHandle_t nvjpeg_handle_;
-  nvjpegJpegStream_t jpeg_stream_;
+
+  struct PerThreadResources {
+    explicit PerThreadResources(nvjpegHandle_t handle);
+    PerThreadResources(PerThreadResources&& other);
+    ~PerThreadResources();
+    nvjpegJpegStream_t jpeg_stream;
+  };
+  std::vector<PerThreadResources> per_thread_resources_;
   CUDAEvent event_;
   nvjpegJpegState_t state_;
 
+  int batch_sz_ = 0;  // number of samples to be decoded by nvJPEG
   struct SampleMeta {
+    bool can_decode;
+    int idx_in_batch;  // only relevant if can_decode == true
     bool needs_processing;
     Orientation orientation;
     float dyn_range_multiplier;
   };
-
   std::vector<SampleMeta> sample_meta_;
   std::vector<const unsigned char*> encoded_;
   std::vector<size_t> encoded_len_;