RegCount max registers calculation

Differential Revision: D59494644

Author: estebanpadilla

backends/vulkan/tools/gpuinfo/include/stats.h

@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+#include <array>
+#include <cstdint>
+
+template <typename T>
+class AvgStats {
+  T sum_ = 0;
+  uint64_t n_ = 0;
+
+ public:
+  typedef T value_t;
+
+  void push(T value) {
+    sum_ += value;
+    n_ += 1;
+  }
+  inline bool has_value() const {
+    return n_ != 0;
+  }
+  operator T() const {
+    return sum_ / n_;
+  }
+};
+
+template <typename T, size_t NTap>
+class NTapAvgStats {
+  std::array<double, NTap> hist_;
+  size_t cur_idx_;
+  bool ready_;
+
+ public:
+  typedef T value_t;
+
+  void push(T value) {
+    hist_[cur_idx_++] = value;
+    if (cur_idx_ >= NTap) {
+      cur_idx_ = 0;
+      ready_ = true;
+    }
+  }
+  inline bool has_value() const {
+    return ready_;
+  }
+  operator T() const {
+    double out = 0.0;
+    for (double x : hist_) {
+      out += x;
+    }
+    out /= NTap;
+    return out;
+  }
+};

backends/vulkan/tools/gpuinfo/include/utils.h

@@ -8,6 +8,8 @@
 
 #include <executorch/backends/vulkan/runtime/api/api.h>
 
+#include "stats.h"
+
 using namespace vkcompute;
 using namespace api;
 
@@ -47,3 +49,48 @@ void ensure_min_niter(
     niter = uint32_t(niter * min_time_us / t);
   }
 }
+
+template <uint32_t NTap>
+struct DtJumpFinder {
+ private:
+  NTapAvgStats<double, NTap> time_avg_;
+  AvgStats<double> dtime_avg_;
+  double compensation_;
+  double threshold_;
+
+ public:
+  // Compensation is a tiny additive to give on delta time so that the algorithm
+  // works smoothly when a sequence of identical timing is ingested, which is
+  // pretty common in our tests. Threshold is simply how many times the new
+  // delta has to be to be recognized as a deviation.
+  DtJumpFinder(double compensation = 0.01, double threshold = 10)
+      : time_avg_(),
+        dtime_avg_(),
+        compensation_(compensation),
+        threshold_(threshold) {}
+
+  // Returns true if the delta time regarding to the last data point seems
+  // normal; returns false if it seems the new data point is too much away from
+  // the historical records.
+  bool push(double time) {
+    if (time_avg_.has_value()) {
+      double dtime = std::abs(time - time_avg_) + (compensation_ * time_avg_);
+      if (dtime_avg_.has_value()) {
+        double ddtime = std::abs(dtime - dtime_avg_);
+        if (ddtime > threshold_ * dtime_avg_) {
+          return true;
+        }
+      }
+      dtime_avg_.push(dtime);
+    }
+    time_avg_.push(time);
+    return false;
+  }
+
+  double dtime_avg() const {
+    return dtime_avg_;
+  }
+  double compensate_time() const {
+    return compensation_ * time_avg_;
+  }
+};

backends/vulkan/tools/gpuinfo/src/app.cpp

@@ -14,6 +14,11 @@
 
 void reg_count() {
   const uint32_t NREG_MIN = 1;
+  const uint32_t NREG_MAX = 512;
+  const uint32_t NREG_STEP = 1;
+
+  const double COMPENSATE = 0.01;
+  const double THRESHOLD = 3;
 
   uint32_t NITER;
 
@@ -43,6 +48,26 @@ void reg_count() {
   std::cout << "Calculating NITER..." << std::endl;
   ensure_min_niter(1000, NITER, [&]() { return bench(1, 1, NREG_MIN); });
   std::cout << "NITER: " << NITER << std::endl;
+
+  uint32_t nreg_max;
+
+  DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
+  uint32_t nreg = NREG_MIN;
+  for (; nreg <= NREG_MAX; nreg += NREG_STEP) {
+    double time = bench(1, 1, nreg);
+    std::cout << "Testing nreg=\t" << nreg << "\tTime=\t" << time << std::endl;
+    if (dj.push(time)) {
+      nreg -= NREG_STEP;
+      nreg_max = nreg;
+      break;
+    }
+  }
+  if (nreg >= NREG_MAX) {
+    std::cout << "Unable to conclude a maximal register count" << std::endl;
+    nreg_max = NREG_STEP;
+  } else {
+    std::cout << nreg_max << " available at most" << std::endl;
+  }
 }
 
 int main(int argc, const char** argv) {