Kernel 10: Autotuning

scripts/kernel_10_autotuner.sh

@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+
+set -u 
+
+# Define the range of values for each parameter
+BK_VALUES=(8 16 32 64)
+BM_VALUES=(64 128 256)
+BN_VALUES=(64 128 256)
+WM_VALUES=(32 64 128 256)
+WN_VALUES=(32 64 128 256)
+WNITER_VALUES=(1 2 4 8)
+TM_VALUES=(4 8 16 32)
+TN_VALUES=(4 8 16 32)
+NUM_THREADS_VALUES=(128 256)
+
+cd "$(dirname "$0")"
+cd "../build"
+
+RUNNER="../src/runner.cu"
+OUTPUT="../benchmark_results/kernel_10_autotune_results.txt"
+
+# Clear the output file
+echo "" > $OUTPUT
+
+# Set GPU to use
+export DEVICE="0"
+WARPSIZE=32
+
+
+TOTAL_CONFIGS="$(( ${#BK_VALUES[@]} * ${#BM_VALUES[@]} * ${#BN_VALUES[@]} * ${#WM_VALUES[@]} * ${#WN_VALUES[@]} * ${#WNITER_VALUES[@]} * ${#TM_VALUES[@]} * ${#TN_VALUES[@]} * ${#NUM_THREADS_VALUES[@]} ))"
+CONFIG_NUM=0
+
+# Loop through all combinations of parameters
+for BK in "${BK_VALUES[@]}"; do
+for BM in "${BM_VALUES[@]}"; do
+for BN in "${BN_VALUES[@]}"; do
+for WM in "${WM_VALUES[@]}"; do
+for WN in "${WN_VALUES[@]}"; do
+for WN_ITER in "${WNITER_VALUES[@]}"; do
+for TM in "${TM_VALUES[@]}"; do
+for TN in "${TN_VALUES[@]}"; do
+for NUM_THREADS in "${NUM_THREADS_VALUES[@]}"; do
+echo ""
+CONFIG_NUM=$(( CONFIG_NUM + 1 ))
+# skip configurations that don't fullfil preconditions
+NUM_WARPS=$(( NUM_THREADS / 32 ))
+if ! (( BN % WN == 0 && BM % WM == 0 )); then
+  echo "Error: BN % WN must be 0 and BM % WM must be 0."
+  continue
+fi
+if ! (( (BN / WN) * (BM / WM) == NUM_WARPS )); then
+  echo "Error: (BN / WN) * (BM / WM) must be equal to NUM_WARPS."
+  continue
+fi
+if ! (( (WM * WN) % (WARPSIZE * TM * TN * WN_ITER) == 0 )); then
+  echo "Error: (WM * WN) % (WARPSIZE * TM * TN * WN_ITER) must be 0."
+  continue
+fi
+WM_ITER=$(( (WM * WN) / (WARPSIZE * TM * TN * WN_ITER) ))
+if ! (( WM % WM_ITER == 0 && WN % WN_ITER == 0 )); then
+  echo "Error: WM % WM_ITER must be 0 and WN % WN_ITER must be 0."
+  continue
+fi
+if ! (( (NUM_THREADS * 4) % BK == 0 )); then
+  echo "Error: (NUM_THREADS * 4) % BK must be 0."
+  continue
+fi
+if ! (( (NUM_THREADS * 4) % BN == 0 )); then
+  echo "Error: (NUM_THREADS * 4) % BN must be 0."
+  continue
+fi
+if ! (( BN % (16 * TN) == 0 )); then
+  echo "Error: BN must be a multiple of 16 * TN."
+  continue
+fi
+if ! (( BM % (16 * TM) == 0 )); then
+  echo "Error: BM must be a multiple of 16 * TM."
+  continue
+fi
+if ! (( (BM * BK) % (4 * NUM_THREADS) == 0 )); then
+  echo "Error: (BM * BK) % (4 * NUM_THREADS) must be 0."
+  continue
+fi
+if ! (( (BN * BK) % (4 * NUM_THREADS) == 0 )); then
+  echo "Error: (BN * BK) % (4 * NUM_THREADS) must be 0."
+  continue
+fi
+
+# Update the parameters in the source code
+sed -i "s/const uint K10_NUM_THREADS = .*/const uint K10_NUM_THREADS = $NUM_THREADS;/" $RUNNER
+sed -i "s/const uint K10_BN = .*/const uint K10_BN = $BN;/" $RUNNER
+sed -i "s/const uint K10_BM = .*/const uint K10_BM = $BM;/" $RUNNER
+sed -i "s/const uint K10_BK = .*/const uint K10_BK = $BK;/" $RUNNER
+sed -i "s/const uint K10_WM = .*/const uint K10_WM = $WM;/" $RUNNER
+sed -i "s/const uint K10_WN = .*/const uint K10_WN = $WN;/" $RUNNER
+sed -i "s/const uint K10_WNITER = .*/const uint K10_WNITER = $WN_ITER;/" $RUNNER
+sed -i "s/const uint K10_TM = .*/const uint K10_TM = $TM;/" $RUNNER
+sed -i "s/const uint K10_TN = .*/const uint K10_TN = $TN;/" $RUNNER
+
+# Rebuild the program
+ninja
+
+echo "($CONFIG_NUM/$TOTAL_CONFIGS): BK=$BK BM=$BM BN=$BN WM=$WM WN=$WN WN_ITER=$WN_ITER TM=$TM TN=$TN NUM_THREADS=$NUM_THREADS" |& tee -a $OUTPUT
+# Run the benchmark and get the result
+# Kill the program after 4 seconds if it doesn't finish
+timeout -v 8 ./sgemm 10 | tee -a $OUTPUT
+done
+done
+done
+done
+done
+done
+done
+done
+done

src/runner.cu

@@ -325,15 +325,13 @@ void runSgemmWarptiling(int M, int N, int K, float alpha, float *A, float *B,
                         float beta, float *C) {
   const uint K10_NUM_THREADS = 128;
   const uint K10_BN = 128;
-  const uint K10_BM = 128;
-  const uint K10_BK = 16;
+  const uint K10_BM = 64;
+  const uint K10_BK = 8;
   const uint K10_WN = 64;
-  const uint K10_WM = 64;
-  const uint K10_WMITER = 2;
-  const uint K10_TN = 8;
+  const uint K10_WM = 32;
+  const uint K10_WNITER = 2;
+  const uint K10_TN = 4;
   const uint K10_TM = 4;
-  constexpr uint K10_WNITER =
-      (K10_WM * K10_WN) / (32 * K10_TM * K10_TN * K10_WMITER);
   dim3 blockDim(K10_NUM_THREADS);
 
   constexpr uint NUM_WARPS = K10_NUM_THREADS / 32;
@@ -342,11 +340,13 @@ void runSgemmWarptiling(int M, int N, int K, float alpha, float *A, float *B,
   static_assert((K10_BN % K10_WN == 0) and (K10_BM % K10_WM == 0));
   static_assert((K10_BN / K10_WN) * (K10_BM / K10_WM) == NUM_WARPS);
 
-  // warpsubtile in warptile
-  static_assert((K10_WM % K10_WMITER == 0) and (K10_WN % K10_WNITER == 0));
   // threads in warpsubtile
   static_assert((K10_WM * K10_WN) % (WARPSIZE * K10_TM * K10_TN * K10_WNITER) ==
                 0);
+  constexpr uint K10_WMITER =
+      (K10_WM * K10_WN) / (32 * K10_TM * K10_TN * K10_WNITER);
+  // warpsubtile in warptile
+  static_assert((K10_WM % K10_WMITER == 0) and (K10_WN % K10_WNITER == 0));
 
   static_assert((K10_NUM_THREADS * 4) % K10_BK == 0,
                 "NUM_THREADS*4 must be multiple of K9_BK to avoid quantization "