RegCount concurrency calculation

Differential Revision: D59497314

Author: Esteban Padilla Cerdio

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

@@ -20,6 +20,10 @@ void reg_count() {
   const double COMPENSATE = 0.01;
   const double THRESHOLD = 3;
 
+  const uint32_t NGRP_MIN = 1;
+  const uint32_t NGRP_MAX = 64;
+  const uint32_t NGRP_STEP = 1;
+
   uint32_t NITER;
 
   auto bench = [&](uint32_t nthread, uint32_t ngrp, uint32_t nreg) {
@@ -68,6 +72,41 @@ void reg_count() {
   } else {
     std::cout << nreg_max << " available at most" << std::endl;
   }
+
+  auto find_ngrp_by_nreg = [&](uint32_t nreg) {
+    DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
+    for (auto ngrp = NGRP_MIN; ngrp <= NGRP_MAX; ngrp += NGRP_STEP) {
+      auto time = bench(1, ngrp, nreg);
+      std::cout << "Testing occupation (nreg=" << nreg << "); ngrp=" << ngrp
+                << ", time=" << time << " us" << std::endl;
+
+      if (dj.push(time)) {
+        ngrp -= NGRP_STEP;
+        std::cout << "Using " << nreg << " registers can have " << ngrp
+                  << " concurrent single-thread workgroups" << std::endl;
+        return ngrp;
+      }
+    }
+    std::cout
+        << "Unable to conclude a maximum number of concurrent single-thread workgroups when "
+        << nreg << " registers are occupied" << std::endl;
+    return (uint32_t)1;
+  };
+
+  uint32_t ngrp_full, ngrp_half;
+  ngrp_full = find_ngrp_by_nreg(nreg_max);
+  ngrp_half = find_ngrp_by_nreg(nreg_max / 2);
+
+  std::string reg_ty;
+
+  if (ngrp_full * 1.5 < ngrp_half) {
+    std::cout << "All physical threads in an sm share " << nreg_max
+              << " registers" << std::endl;
+
+  } else {
+    std::cout << "Each physical thread has " << nreg_max << " registers"
+              << std::endl;
+  }
 }
 
 int main(int argc, const char** argv) {