tensorflow/lite/delegates/gpu/gl/kernels/mean.cc

/* Copyright 2020 The TensorFlow 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.
==============================================================================*/

#include "tensorflow/lite/delegates/gpu/gl/kernels/mean.h"

#include <algorithm>
#include <cstdint>
#include <cstring>
#include <string>
#include <vector>

#include "absl/memory/memory.h"
#include "absl/status/status.h"
#include "tensorflow/lite/delegates/gpu/common/status.h"
#include "tensorflow/lite/delegates/gpu/common/types.h"
#include "tensorflow/lite/delegates/gpu/common/util.h"

namespace tflite {
namespace gpu {
namespace gl {
namespace {

bool UseSubgroupBasedImpl(const GpuInfo& gpu_info) {
  return gpu_info.IsApiVulkan() &&
         (gpu_info.vulkan_info.api_version_major > 1 ||
          gpu_info.vulkan_info.api_version_minor >= 1) &&
         gpu_info.vulkan_info.subgroup_size >= 32 &&
         gpu_info.vulkan_info.supports_subgroup_arithmetic;
}

// An implementation of Mean for desktop GPUs and some phones with recent
// Vulkan drivers. It is more parallel than the trivial Mean operation, but
// still limited to using a single work group.
void GenerateSubgroupBasedMean(const NodeShader::GenerationContext& ctx,
                               GeneratedCode* generated_code) {
  int height = ctx.input_shapes[0][1];
  int width = ctx.input_shapes[0][2];
  int depth = ctx.input_shapes[0][3];
  std::vector<Variable> parameters = {
      {"input_data_0_h", height},
      {"input_data_0_w", width},
      {"output_data_0_h", 1},
      {"output_data_0_w", 1},
  };

  std::string source = R"(
  // Round columns and rows per invocation up, to ensure that we read the
  // entire input.
  const uint columns_per_invocation =
      ($input_data_0_w$ + (gl_WorkGroupSize.x - 1))/gl_WorkGroupSize.x;
  const uint rows_per_invocation =
      ($input_data_0_h$ + (gl_WorkGroupSize.y - 1))/gl_WorkGroupSize.y;
  const uint first_row = gl_GlobalInvocationID.y*rows_per_invocation;
  const uint first_col = gl_GlobalInvocationID.x*columns_per_invocation;
  const uint last_row_exclusive =
      min(first_row+rows_per_invocation, $input_data_0_h$);
  const uint last_column_exclusive =
      min(first_col+columns_per_invocation, $input_data_0_w$);
  vec4 value = vec4(0);
  for (uint h = first_row; h < last_row_exclusive; ++h) {
    for (uint w = first_col; w < last_column_exclusive; ++w) {
      value += $input_data_0[w, h, gid.z]$;
    }
  }
  highp vec4 subgroup_sum = subgroupAdd(value);
  if(subgroupElect()) {
    subgroup_sums[gl_SubgroupID] = subgroup_sum;
  }

  memoryBarrierShared();
  barrier();
  // Do the final reduction in the first subgroup.
  if(gl_SubgroupID == 0) {
    highp vec4 subtotal = vec4(0);
    if (gl_SubgroupInvocationID < gl_NumSubgroups) {
      subtotal = subgroup_sums[gl_SubgroupInvocationID];
    }
    highp vec4 grand_total = subgroupAdd(subtotal);
    if(subgroupElect()) {
      highp vec4 result = grand_total / $input_data_0_w$ / $input_data_0_h$;
      $output_data_0[0, 0, gid.z] = result$;
    }
  }
  )";

  const uint32_t subgroup_size = ctx.gpu_info->vulkan_info.subgroup_size;
  const uint32_t max_wg_size_x = ctx.gpu_info->GetMaxWorkGroupSizeForX();
  const uint32_t max_wg_size_y = ctx.gpu_info->GetMaxWorkGroupSizeForY();
  // Due to the design of the shader, at most subgroup_size subgroups can be
  // launched. This may limit the maximal workgroup size.
  const uint32_t max_wg_size =
      std::min(static_cast<uint32_t>(ctx.gpu_info->GetMaxWorkGroupTotalSize()),
               subgroup_size * subgroup_size);
  const uint32_t max_number_of_subgroups = max_wg_size / subgroup_size;
  uint32_t wg_size_x = 0;
  uint32_t wg_size_y = 0;
  if (width * height <= max_wg_size && width <= max_wg_size_x &&
      height <= max_wg_size_y) {
    wg_size_x = width;
    wg_size_y = height;
  } else {
    // Approximately square workgroup. Also make sure to limit by driver limit
    // and input size.
    wg_size_x = std::min({static_cast<uint32_t>(std::sqrt(max_wg_size)),
                          max_wg_size_x, static_cast<uint32_t>(width)});
    wg_size_y = std::min({max_wg_size / wg_size_x, max_wg_size_y,
                          static_cast<uint32_t>(height)});
  }

  std::vector<Variable> shared_variables = {
      {"subgroup_sums", std::vector<float4>(max_number_of_subgroups)},
  };

  *generated_code = {
      /*parameters=*/std::move(parameters),
      /*objects=*/{},
      /*shared_variables=*/{std::move(shared_variables)},
      // Make sure we get one dispatch of size wg_size_x*wg_size_y*1 per layer.
      /*workload=*/
      uint3(wg_size_x, wg_size_y, uint32_t(DivideRoundUp(depth, 4))),
      /*workgroup=*/uint3(wg_size_x, wg_size_y, 1u),
      /*source_code=*/std::move(source),
      /*input=*/IOStructure::ONLY_DEFINITIONS,
      /*output=*/IOStructure::ONLY_DEFINITIONS,
  };
}

void GenerateTrivialMean(const NodeShader::GenerationContext& ctx,
                         GeneratedCode* generated_code) {
  std::vector<Variable> parameters = {
      {"input_data_0_h", static_cast<int>(ctx.input_shapes[0][1])},
      {"input_data_0_w", static_cast<int>(ctx.input_shapes[0][2])}};

  std::string source = R"(
    // Shaders may be compiled with a precision hint mediump, which means that
    // GLSL compiler may drop the size of float data type from 32 to 16 bits.
    // If "sum" and "size" variables are 16bit floats, their values range
    // become not enough for providing a good results accuracy. That is why
    // their precision is forced to be 32bit by using highp qualifier.

    highp vec4 sum = vec4(0.0);
    highp float size = float($input_data_0_w$ * $input_data_0_h$);
    for (int w = 0; w < $input_data_0_w$; w++) {
      for (int h = 0; h < $input_data_0_h$; h++) {
        sum += $input_data_0[w, h, gid.z]$;
      }
    }
    value_0 = sum / size;
  )";
  *generated_code = {
      /*parameters=*/std::move(parameters),
      /*objects=*/{},
      /*shared_variables=*/{},
      /*workload=*/uint3(),
      /*workgroup=*/uint3(1, 1, 4),
      /*source_code=*/std::move(source),
      /*input=*/IOStructure::ONLY_DEFINITIONS,
      /*output=*/IOStructure::AUTO,
  };
}

class Mean : public NodeShader {
 public:
  absl::Status GenerateCode(const GenerationContext& ctx,
                            GeneratedCode* generated_code) const final {
    const auto& attr = absl::any_cast<const MeanAttributes&>(ctx.op_attr);
    if (attr.dims != std::set<Axis>({Axis::HEIGHT, Axis::WIDTH})) {
      return absl::InvalidArgumentError(
          "Mean calculation is supported only for height and width.");
    }

    if (!(ctx.input_shapes.size() == 1 && ctx.output_shapes.size() == 1 &&
          ctx.output_shapes[0][1] == 1 && ctx.output_shapes[0][2] == 1 &&
          ctx.output_shapes[0][3] == ctx.input_shapes[0][3])) {
      return absl::InvalidArgumentError(
          "Mean calculation is supported for one input and one 1x1 output with "
          "the same channel count.");
    }

    if (UseSubgroupBasedImpl(*ctx.gpu_info)) {
      GenerateSubgroupBasedMean(ctx, generated_code);
    } else {
      GenerateTrivialMean(ctx, generated_code);
    }
    return absl::OkStatus();
  }
};

}  // namespace

std::unique_ptr<NodeShader> NewMeanNodeShader() {
  return absl::make_unique<Mean>();
}

}  // namespace gl
}  // namespace gpu
}  // namespace tflite