leonarski_fandClaude Opus 5.5
9c5141c7c0
The CPU-only image loop is DRAM-bound on 16M frames: decode, preprocess, the adaptive finder's plain ring pass and FlagRings each streamed the whole frame through memory. - JFJochDecompressHperfBlocks hands each decoded bitshuffle block to a callback; with no output buffer the block is unshuffled into a reused block-sized scratch (JFJochDecompressBlocks). - MXAnalysisWithoutFPGA::PreprocessCPU preprocesses each block into the int32 buffer (ImagePreprocessorCPU::AnalyzeBlock) and, when the fused CPU finder runs, puts it through the plain ring pass + fused azint (AdaptiveSpotFinderCPU::AccumulateRingsBlock) while it is in cache. Detect() then starts from those sums. The per-worker decompression buffer is no longer allocated for bitshuffled data. - FlagRings becomes FlagRow, called by DetectAt's first pass for row y+NBX just before that row enters the vertical sums; first_pass_needed is marked from each row's candidates at the same point. Exact: blocks arrive in pixel order, so the float azint sums see the same pixels in the same order; the per-pixel expressions are unchanged; everything else is integer. p.hkl, p.mtz, p_P1.mtz and p_unmerged.mtz byte-identical to rc173 on myob, cytc, lyso, sparse (CPU-only build), GPU myob identical (the GPU path does not take this route). CPU-only, 32 workers, under the gpulock on a shared (loaded) machine, base -> fused, two rounds (second in reversed order): myob 155.9 -> 97.7 s, 139.9 -> 88.1 s (loop 46.1 -> 26.8 s/pass; user 3690 -> 2221 s) cytc 220.8 -> 156.6 s, 216.6 -> 155.1 s (user 5160 -> 4128 s) lyso 134.5 -> 123.3 s, 69.5 -> 64.8 s peak RSS myob 15.2 -> 10.8 GB, cytc 12.6 -> 10.4 GB, lyso 7.0 -> 6.7 GB Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01D1G8gJVAy6gp1K5Dz3NE5C
285 lines
13 KiB
C++
285 lines
13 KiB
C++
// SPDX-FileCopyrightText: 2024 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
|
|
// SPDX-License-Identifier: GPL-3.0-only
|
|
|
|
#include <algorithm>
|
|
#include <bit>
|
|
#include <bitset>
|
|
#include <cmath>
|
|
|
|
#include "ImageSpotFinderCPU.h"
|
|
#include "StrongPixelSet.h"
|
|
|
|
ImageSpotFinderCPU::ImageSpotFinderCPU(int32_t in_width, int32_t in_height)
|
|
: ImageSpotFinder(in_width, in_height), first_pass_buffer(OutputSize(), 0) {}
|
|
|
|
void ImageSpotFinderCPU::Detect(const ImagePreprocessorBuffer &image,
|
|
const SpotFindingSettings &settings) {
|
|
// Two passes, as ImageSpotFinderGPU::Detect does. The second recomputes every local background
|
|
// with the pixels the first found strong taken out of it, and keeps those pixels strong. It
|
|
// matters because a spot wide enough to reach into its own background window inflates the mean
|
|
// and variance it is then tested against, so its outer pixels fail the SNR test on a single
|
|
// pass. The GPU has always done this; running one pass here made the two finders return
|
|
// different spot lists for the same frame.
|
|
DetectPass(image, settings, nullptr, first_pass_buffer);
|
|
DetectPass(image, settings, first_pass_buffer.data(), output_buffer);
|
|
}
|
|
|
|
namespace {
|
|
|
|
// The local-box SNR test of one pixel. sum/sum2/valid are its window's, centre pixel included.
|
|
bool StrongInWindow(int64_t pxl_val, int64_t sum, int64_t sum2, int64_t valid,
|
|
const SpotFindingSettings &settings, float strong2) {
|
|
const int64_t sum_local = sum - pxl_val;
|
|
const int64_t sum2_local = sum2 - pxl_val * pxl_val;
|
|
const int64_t valid_local = valid - 1;
|
|
|
|
const int64_t var = valid_local * sum2_local - (sum_local * sum_local);
|
|
const int64_t in_minus_mean = pxl_val * valid_local - sum_local;
|
|
|
|
return (pxl_val == INT32_MAX) // saturated pixel, or strong in the previous pass, is accepted always
|
|
|| ((pxl_val != INT32_MIN && // pixel is not bad pixel
|
|
valid_local > ImageSpotFinder::MIN_VALID_PIXELS && // too many bad pixels around will give poor statistics
|
|
(pxl_val > settings.photon_count_threshold) && // pixel is above count threshold
|
|
(in_minus_mean > 0) && // pixel value is larger than mean
|
|
(in_minus_mean * in_minus_mean > static_cast<int64_t>(std::ceil(var * strong2)))));
|
|
// pixel is above SNR threshold
|
|
}
|
|
|
|
} // namespace
|
|
|
|
void ImageSpotFinderCPU::DetectAt(const ImagePreprocessorBuffer &image, const SpotFindingSettings &settings,
|
|
const std::vector<uint32_t> &candidates,
|
|
const std::function<void(int32_t)> &fill_row) {
|
|
candidate_windows.clear();
|
|
// Filled in by DetectPass as the candidates of each row become known.
|
|
first_pass_needed.assign(static_cast<size_t>(height) * ((width + 31) / 32), 0);
|
|
DetectPass(image, settings, nullptr, first_pass_buffer, candidates.data(), fill_row);
|
|
|
|
std::fill(output_buffer.begin(), output_buffer.end(), 0);
|
|
const float strong2 = settings.signal_to_noise_threshold * settings.signal_to_noise_threshold;
|
|
const auto first_pass = [&](int32_t pxl) { return (first_pass_buffer[pxl / 32] >> (pxl % 32)) & 1U; };
|
|
|
|
for (const auto &c : candidate_windows) {
|
|
const int32_t line = c.pxl / width;
|
|
const int32_t col = c.pxl % width;
|
|
int64_t sum = c.sum, sum2 = c.sum2, valid = c.valid;
|
|
// Take out of the window what the second pass does not count: the first pass's strong pixels.
|
|
for (int32_t y = std::max(line - NBX, 0); y <= std::min(line + NBX, height - 1); y++) {
|
|
const int32_t first = y * width + std::max(col - NBX, 0);
|
|
const int32_t last = y * width + std::min(col + NBX, width - 1);
|
|
for (int32_t w = first / 32; w <= last / 32; w++) {
|
|
uint32_t bits = first_pass_buffer[w];
|
|
if (w == first / 32)
|
|
bits &= UINT32_MAX << (first % 32);
|
|
if (w == last / 32 && last % 32 != 31)
|
|
bits &= (1U << (last % 32 + 1)) - 1;
|
|
while (bits) {
|
|
const int32_t q = w * 32 + std::countr_zero(bits);
|
|
bits &= bits - 1;
|
|
const int64_t v = image[q];
|
|
if (v != INT32_MAX && v != INT32_MIN) {
|
|
sum -= v;
|
|
sum2 -= v * v;
|
|
valid -= 1;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
const int64_t pxl_val = first_pass(c.pxl) ? INT32_MAX : image[c.pxl];
|
|
if (StrongInWindow(pxl_val, sum, sum2, valid, settings, strong2))
|
|
output_buffer[c.pxl / 32] |= 1U << (c.pxl % 32);
|
|
}
|
|
}
|
|
|
|
void ImageSpotFinderCPU::DetectPass(const ImagePreprocessorBuffer &image,
|
|
const SpotFindingSettings &settings,
|
|
const uint32_t *prev_strong,
|
|
std::vector<uint32_t> &out_buffer,
|
|
const uint32_t *candidates,
|
|
const std::function<void(int32_t)> &fill_row) {
|
|
for (int i = 0; i < OutputSize(); i++)
|
|
out_buffer[i] = 0;
|
|
|
|
// The first pass's bits are read only inside the window of a candidate (and at the candidate), so
|
|
// with candidates it tests only the pixels of the row/column blocks such a window reaches; the bits
|
|
// it leaves unset there are never read. Those blocks are marked from each row's candidates as the
|
|
// row enters the vertical sums below, which is before any row within NBX of it is tested - so the
|
|
// candidates themselves can be filled in a row at a time (fill_row), while the row is in cache.
|
|
const int32_t nblocks = (width + 31) / 32;
|
|
const auto new_row = [&](int32_t r) {
|
|
if (!candidates)
|
|
return;
|
|
if (fill_row)
|
|
fill_row(r);
|
|
const int32_t first = r * width;
|
|
const int32_t last = first + width - 1;
|
|
for (int32_t w = first / 32; w <= last / 32; w++) {
|
|
uint32_t bits = candidates[w];
|
|
if (w == first / 32)
|
|
bits &= UINT32_MAX << (first % 32);
|
|
if (w == last / 32 && last % 32 != 31)
|
|
bits &= (1U << (last % 32 + 1)) - 1;
|
|
for (; bits; bits &= bits - 1) {
|
|
const int32_t col = w * 32 + std::countr_zero(bits) - first;
|
|
for (int32_t y = std::max(r - NBX, 0); y <= std::min(r + NBX, height - 1); y++)
|
|
for (int32_t b = std::max(col - NBX, 0) / 32; b <= std::min(col + NBX, width - 1) / 32; b++)
|
|
first_pass_needed[static_cast<size_t>(y) * nblocks + b] = 1;
|
|
}
|
|
}
|
|
};
|
|
|
|
// A pixel found strong by the previous pass reads as INT32_MAX, which the accumulation below
|
|
// already skips and the acceptance test below already takes as strong - the same substitution
|
|
// the GPU kernel makes when it reads prev_out.
|
|
auto value_at = [&](int32_t pxl) -> int32_t {
|
|
if (prev_strong && (prev_strong[pxl / 32] & (1U << (pxl % 32))))
|
|
return INT32_MAX;
|
|
return image[pxl];
|
|
};
|
|
|
|
std::bitset<32> out = 0;
|
|
|
|
if (settings.signal_to_noise_threshold <= 0.0) {
|
|
if (settings.photon_count_threshold > 0) {
|
|
for (int pxl = 0; pxl < height * width; pxl++) {
|
|
int32_t bit = pxl % 32;
|
|
int32_t pxl_val = value_at(pxl);
|
|
if (pxl_val == INT32_MAX || (pxl_val > settings.photon_count_threshold && pxl_val != INT32_MIN))
|
|
out.set(bit);
|
|
|
|
if (bit == 31) {
|
|
out_buffer[pxl / 32] = out.to_ulong();
|
|
out.reset();
|
|
}
|
|
}
|
|
}
|
|
} else {
|
|
float strong2 = settings.signal_to_noise_threshold * settings.signal_to_noise_threshold;
|
|
|
|
// Sum and sum of squares of (2*NBY+1) vertical elements
|
|
// These are updated after each line is finished
|
|
// 64-bit integer guarantees calculations are made without rounding errors
|
|
std::vector<int64_t> sum_vert(width, 0);
|
|
std::vector<int64_t> sum2_vert(width, 0);
|
|
std::vector<uint16_t> valid_vert(width, 0);
|
|
|
|
for (int line = 0; line < NBX; line++) {
|
|
new_row(line);
|
|
for (int col = 0; col < width; col++) {
|
|
auto pxl = line * width + col;
|
|
|
|
int64_t tmp = value_at(pxl);
|
|
if (tmp != INT32_MAX && tmp != INT32_MIN) {
|
|
sum_vert[col] += tmp;
|
|
sum2_vert[col] += tmp * tmp;
|
|
valid_vert[col] += 1;
|
|
}
|
|
}
|
|
}
|
|
|
|
for (int line = 0; line < height; line++) {
|
|
if (line < height - NBX)
|
|
new_row(line + NBX);
|
|
for (int col = 0; col < width; col++) {
|
|
if (line < height - NBX) {
|
|
auto pxl = (line + NBX) * width + col;
|
|
|
|
int64_t tmp = value_at(pxl);
|
|
if (tmp != INT32_MAX && tmp != INT32_MIN) {
|
|
sum_vert[col] += tmp;
|
|
sum2_vert[col] += tmp * tmp;
|
|
valid_vert[col] += 1;
|
|
}
|
|
}
|
|
|
|
if (line >= NBX + 1) {
|
|
auto pxl = (line - (NBX + 1)) * width + col;
|
|
int64_t tmp = value_at(pxl);
|
|
if (tmp != INT32_MAX && tmp != INT32_MIN) {
|
|
sum_vert[col] -= tmp;
|
|
sum2_vert[col] -= tmp * tmp;
|
|
valid_vert[col] -= 1;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (candidates) {
|
|
// Only the runs of 32-column blocks whose result is read (first_pass_needed; every
|
|
// candidate lies in one). Each run starts its window from the vertical sums over the
|
|
// columns it covers there - integers, so the same sums the running window holds at
|
|
// that column - and slides it as below; the bits outside the runs stay 0, as below.
|
|
const uint8_t *needed = first_pass_needed.data() + static_cast<size_t>(line) * nblocks;
|
|
for (int32_t b = 0; b < nblocks;) {
|
|
if (!needed[b]) { b++; continue; }
|
|
const int32_t c0 = b * 32;
|
|
while (b < nblocks && needed[b]) b++;
|
|
const int32_t c1 = std::min(b * 32, width);
|
|
int64_t sum = 0, sum2 = 0, valid = 0;
|
|
for (int32_t c = std::max(c0 - NBX, 0); c <= std::min(c0 + NBX, width - 1); c++) {
|
|
sum += sum_vert[c];
|
|
sum2 += sum2_vert[c];
|
|
valid += valid_vert[c];
|
|
}
|
|
for (int32_t col = c0; col < c1; col++) {
|
|
if (col > c0) {
|
|
if (col < width - NBX) {
|
|
sum += sum_vert[col + NBX];
|
|
sum2 += sum2_vert[col + NBX];
|
|
valid += valid_vert[col + NBX];
|
|
}
|
|
if (col >= NBX + 1) {
|
|
sum -= sum_vert[col - NBX - 1];
|
|
sum2 -= sum2_vert[col - NBX - 1];
|
|
valid -= valid_vert[col - NBX - 1];
|
|
}
|
|
}
|
|
const int32_t pxl = line * width + col;
|
|
if (candidates[pxl / 32] & (1U << (pxl % 32)))
|
|
candidate_windows.push_back({pxl, sum, sum2, valid});
|
|
if (StrongInWindow(value_at(pxl), sum, sum2, valid, settings, strong2))
|
|
out_buffer[pxl / 32] |= 1U << (pxl % 32);
|
|
}
|
|
}
|
|
continue;
|
|
}
|
|
|
|
int64_t sum = 0;
|
|
int64_t sum2 = 0;
|
|
int64_t valid = 0;
|
|
|
|
for (int col = 0; col < NBX; col++) {
|
|
sum += sum_vert[col];
|
|
sum2 += sum2_vert[col];
|
|
valid += valid_vert[col];
|
|
}
|
|
|
|
for (int col = 0; col < width; col++) {
|
|
if (col < width - NBX) {
|
|
sum += sum_vert[col + NBX];
|
|
sum2 += sum2_vert[col + NBX];
|
|
valid += valid_vert[col + NBX];
|
|
}
|
|
|
|
if (col >= NBX + 1) {
|
|
sum -= sum_vert[col - NBX - 1];
|
|
sum2 -= sum2_vert[col - NBX - 1];
|
|
valid -= valid_vert[col - NBX - 1];
|
|
}
|
|
const int32_t pxl = line * width + col;
|
|
const int32_t bit = pxl % 32;
|
|
|
|
if (StrongInWindow(value_at(pxl), sum, sum2, valid, settings, strong2))
|
|
out.set(bit);
|
|
|
|
if (bit == 31) {
|
|
out_buffer[pxl / 32] = out.to_ulong();
|
|
out.reset() ;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
if (height * width % 32 != 0)
|
|
out_buffer[OutputSize() - 1] |= out.to_ulong(); // zeroed above; the candidates path sets its bits directly
|
|
}
|