imageDenoising_nlm2_kernel.cuh

/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *  * Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 *  * Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *  * Neither the name of NVIDIA CORPORATION nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

////////////////////////////////////////////////////////////////////////////////
//   0 1 2 3 4 5 6 7
// 0 + . . . . . . .
// 1 . . . . . . . .
// 2 . . . . . . . .
// 3 . . . * . . . .
// 4 . . . . . . . .
// 5 . . . . . . . .
// 6 . . . . . . . .
// 7 . . . . . . . .
//
// * - Base point for every thread, + - pixel around which ColorDistance is
// computed
// The idea behind this method:
// - Every thread in a 8x8 block computes just one ColorDistance
// - It is saved in the weights array that is shared across the threads
// - Threads are synced
// - For every pixel inside the block weights are considered to be constants
////////////////////////////////////////////////////////////////////////////////
#include <cooperative_groups.h>

namespace cg = cooperative_groups;

__global__ void NLM2(TColor *dst, int imageW, int imageH, float Noise,
                     float lerpC, cudaTextureObject_t texImage) {
  // Handle to thread block group
  cg::thread_block cta = cg::this_thread_block();

  // Weights cache
  __shared__ float fWeights[BLOCKDIM_X * BLOCKDIM_Y];

  const int ix = blockDim.x * blockIdx.x + threadIdx.x;
  const int iy = blockDim.y * blockIdx.y + threadIdx.y;
  // Add half of a texel to always address exact texel centers
  const float x = (float)ix + 0.5f;
  const float y = (float)iy + 0.5f;
  const float cx = blockDim.x * blockIdx.x + NLM_WINDOW_RADIUS + 0.5f;
  const float cy = blockDim.x * blockIdx.y + NLM_WINDOW_RADIUS + 0.5f;

  if (ix < imageW && iy < imageH) {
    // Find color distance from current texel to the center of NLM window
    float weight = 0;

    for (float n = -NLM_BLOCK_RADIUS; n <= NLM_BLOCK_RADIUS; n++)
      for (float m = -NLM_BLOCK_RADIUS; m <= NLM_BLOCK_RADIUS; m++)
        weight += vecLen(tex2D<float4>(texImage, cx + m, cy + n),
                         tex2D<float4>(texImage, x + m, y + n));

    // Geometric distance from current texel to the center of NLM window
    float dist =
        (threadIdx.x - NLM_WINDOW_RADIUS) * (threadIdx.x - NLM_WINDOW_RADIUS) +
        (threadIdx.y - NLM_WINDOW_RADIUS) * (threadIdx.y - NLM_WINDOW_RADIUS);

    // Derive final weight from color and geometric distance
    weight = __expf(-(weight * Noise + dist * INV_NLM_WINDOW_AREA));

    // Write the result to shared memory
    fWeights[threadIdx.y * BLOCKDIM_X + threadIdx.x] = weight;
    // Wait until all the weights are ready
    cg::sync(cta);

    // Normalized counter for the NLM weight threshold
    float fCount = 0;
    // Total sum of pixel weights
    float sumWeights = 0;
    // Result accumulator
    float3 clr = {0, 0, 0};

    int idx = 0;

    // Cycle through NLM window, surrounding (x, y) texel
    for (float i = -NLM_WINDOW_RADIUS; i <= NLM_WINDOW_RADIUS + 1; i++)
      for (float j = -NLM_WINDOW_RADIUS; j <= NLM_WINDOW_RADIUS + 1; j++) {
        // Load precomputed weight
        float weightIJ = fWeights[idx++];

        // Accumulate (x + j, y + i) texel color with computed weight
        float4 clrIJ = tex2D<float4>(texImage, x + j, y + i);
        clr.x += clrIJ.x * weightIJ;
        clr.y += clrIJ.y * weightIJ;
        clr.z += clrIJ.z * weightIJ;

        // Sum of weights for color normalization to [0..1] range
        sumWeights += weightIJ;

        // Update weight counter, if NLM weight for current window texel
        // exceeds the weight threshold
        fCount += (weightIJ > NLM_WEIGHT_THRESHOLD) ? INV_NLM_WINDOW_AREA : 0;
      }

    // Normalize result color by sum of weights
    sumWeights = 1.0f / sumWeights;
    clr.x *= sumWeights;
    clr.y *= sumWeights;
    clr.z *= sumWeights;

    // Choose LERP quotient basing on how many texels
    // within the NLM window exceeded the weight threshold
    float lerpQ = (fCount > NLM_LERP_THRESHOLD) ? lerpC : 1.0f - lerpC;

    // Write final result to global memory
    float4 clr00 = tex2D<float4>(texImage, x, y);
    clr.x = lerpf(clr.x, clr00.x, lerpQ);
    clr.y = lerpf(clr.y, clr00.y, lerpQ);
    clr.z = lerpf(clr.z, clr00.z, lerpQ);
    dst[imageW * iy + ix] = make_color(clr.x, clr.y, clr.z, 0);
  }
}

extern "C" void cuda_NLM2(TColor *d_dst, int imageW, int imageH, float Noise,
                          float LerpC, cudaTextureObject_t texImage) {
  dim3 threads(BLOCKDIM_X, BLOCKDIM_Y);
  dim3 grid(iDivUp(imageW, BLOCKDIM_X), iDivUp(imageH, BLOCKDIM_Y));

  NLM2<<<grid, threads>>>(d_dst, imageW, imageH, Noise, LerpC, texImage);
}

////////////////////////////////////////////////////////////////////////////////
// Stripped NLM2 kernel, only highlighting areas with different LERP directions
////////////////////////////////////////////////////////////////////////////////
__global__ void NLM2diag(TColor *dst, int imageW, int imageH, float Noise,
                         float LerpC, cudaTextureObject_t texImage) {
  // Handle to thread block group
  cg::thread_block cta = cg::this_thread_block();

  // Weights cache
  __shared__ float fWeights[BLOCKDIM_X * BLOCKDIM_Y];

  const int ix = blockDim.x * blockIdx.x + threadIdx.x;
  const int iy = blockDim.y * blockIdx.y + threadIdx.y;
  // Add half of a texel to always address exact texel centers
  const float x = (float)ix + 0.5f;
  const float y = (float)iy + 0.5f;
  const float cx = blockDim.x * blockIdx.x + NLM_WINDOW_RADIUS + 0.5f;
  const float cy = blockDim.x * blockIdx.y + NLM_WINDOW_RADIUS + 0.5f;

  if (ix < imageW && iy < imageH) {
    // Find color distance from current texel to the center of NLM window
    float weight = 0;

    for (float n = -NLM_BLOCK_RADIUS; n <= NLM_BLOCK_RADIUS + 1; n++)
      for (float m = -NLM_BLOCK_RADIUS; m <= NLM_BLOCK_RADIUS + 1; m++)
        weight += vecLen(tex2D<float4>(texImage, cx + m, cy + n),
                         tex2D<float4>(texImage, x + m, y + n));

    // Geometric distance from current texel to the center of NLM window
    float dist =
        (threadIdx.x - NLM_WINDOW_RADIUS) * (threadIdx.x - NLM_WINDOW_RADIUS) +
        (threadIdx.y - NLM_WINDOW_RADIUS) * (threadIdx.y - NLM_WINDOW_RADIUS);

    // Derive final weight from color and geometric distance
    weight = __expf(-(weight * Noise + dist * INV_NLM_WINDOW_AREA));

    // Write the result to shared memory
    fWeights[threadIdx.y * BLOCKDIM_X + threadIdx.x] = weight;
    // Wait until all the weights are ready
    cg::sync(cta);

    // Normalized counter for the NLM weight threshold
    float fCount = 0;
    int idx = 0;

    // Cycle through NLM window, surrounding (x, y) texel
    for (float n = -NLM_WINDOW_RADIUS; n <= NLM_WINDOW_RADIUS + 1; n++)
      for (float m = -NLM_WINDOW_RADIUS; m <= NLM_WINDOW_RADIUS + 1; m++) {
        // Load precomputed weight
        float weightIJ = fWeights[idx++];

        // Update weight counter, if NLM weight for current window texel
        // exceeds the weight threshold
        fCount += (weightIJ > NLM_WEIGHT_THRESHOLD) ? INV_NLM_WINDOW_AREA : 0;
      }

    // Choose LERP quotient basing on how many texels
    // within the NLM window exceeded the weight threshold
    float lerpQ = (fCount > NLM_LERP_THRESHOLD) ? 1.0f : 0.0f;

    // Write final result to global memory
    dst[imageW * iy + ix] = make_color(lerpQ, 0, (1.0f - lerpQ), 0);
  };
}

extern "C" void cuda_NLM2diag(TColor *d_dst, int imageW, int imageH,
                              float Noise, float LerpC,
                              cudaTextureObject_t texImage) {
  dim3 threads(BLOCKDIM_X, BLOCKDIM_Y);
  dim3 grid(iDivUp(imageW, BLOCKDIM_X), iDivUp(imageH, BLOCKDIM_Y));

  NLM2diag<<<grid, threads>>>(d_dst, imageW, imageH, Noise, LerpC, texImage);
}