minor fixes

vortexgpgpu · Dec 6, 2024 · 115ff2b · 115ff2b
1 parent 896c593
commit 115ff2b
Show file tree

Hide file tree

Showing 5 changed files with 308 additions and 177 deletions.
diff --git a/ci/regression.sh.in b/ci/regression.sh.in
@@ -363,8 +363,8 @@ scope()
 {
     echo "begin scope tests..."
 
-    SCOPE_DEPTH=256 ./ci/blackbox.sh --driver=opae --app=demo --args="-n1" --scope
-    SCOPE_DEPTH=256 ./ci/blackbox.sh --driver=xrt --app=demo --args="-n1" --scope
+    SCOPE_DEPTH=128 ./ci/blackbox.sh --driver=opae --app=demo --args="-n1" --scope
+    SCOPE_DEPTH=128 ./ci/blackbox.sh --driver=xrt --app=demo --args="-n1" --scope
 
     echo "debugging scope done!"
 }
@@ -385,7 +385,7 @@ synthesis()
     echo "begin synthesis tests..."
 
     PREFIX=build_base make -C hw/syn/yosys clean
-    PREFIX=build_base CONFIGS="-DDPI_DISABLE -DEXT_F_DISABLE" make -C hw/syn/yosys synthesis
+    PREFIX=build_base CONFIGS="-DDPI_DISABLE -DEXT_F_DISABLE -DNUM_WARPS=2 -DNUM_THREADS=2" make -C hw/syn/yosys synthesis
 
     echo "synthesis tests done!"
 }

diff --git a/perf/cache/cache_perf.log b/perf/cache/cache_perf.log
diff --git a/perf/cache/run.sh b/perf/cache/run.sh
@@ -10,17 +10,17 @@ sgemm()
 {
 echo "begin cache tests"
 
-CONFIGS="-DICACHE_NUM_WAYS=2" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' > ./perf/cache/cache_perf.log
-echo -e "\n**************************************\n" >> ./perf/cache/cache_perf.log
-CONFIGS="-DDCACHE_NUM_WAYS=2" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> ./perf/cache/cache_perf.log
-echo -e "\n**************************************\n" >> ./perf/cache/cache_perf.log 
-CONFIGS="-DICACHE_NUM_WAYS=4" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> ./perf/cache/cache_perf.log
-echo -e "\n**************************************\n" >> ./perf/cache/cache_perf.log
-CONFIGS="-DDCACHE_NUM_WAYS=4" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> ./perf/cache/cache_perf.log
-echo -e "\n**************************************\n" >> ./perf/cache/cache_perf.log
-CONFIGS="-DICACHE_NUM_WAYS=8" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> ./perf/cache/cache_perf.log
-echo -e "\n**************************************\n" >> ./perf/cache/cache_perf.log
-CONFIGS="-DDCACHE_NUM_WAYS=8" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> ./perf/cache/cache_perf.log
+CONFIGS="-DICACHE_NUM_WAYS=2" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' > cache_perf.log
+echo -e "\n**************************************\n" >> cache_perf.log
+CONFIGS="-DDCACHE_NUM_WAYS=2" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> cache_perf.log
+echo -e "\n**************************************\n" >> cache_perf.log
+CONFIGS="-DICACHE_NUM_WAYS=4" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> cache_perf.log
+echo -e "\n**************************************\n" >> cache_perf.log
+CONFIGS="-DDCACHE_NUM_WAYS=4" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> cache_perf.log
+echo -e "\n**************************************\n" >> cache_perf.log
+CONFIGS="-DICACHE_NUM_WAYS=8" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> cache_perf.log
+echo -e "\n**************************************\n" >> cache_perf.log
+CONFIGS="-DDCACHE_NUM_WAYS=8" ./ci/blackbox.sh --driver=rtlsim --app=sgemm --args="-n64" --perf=1 | grep 'PERF' >> cache_perf.log
 
 echo "cache tests done!"
 }
@@ -36,6 +36,6 @@ case $1 in
     -h | --help ) usage
                     ;;
     * ) sgemm
-        ;;             
+        ;;
 esac
 shift
diff --git a/sim/common/softfloat_ext.cpp b/sim/common/softfloat_ext.cpp
@@ -148,10 +148,9 @@ static inline uint64_t rsqrte7(uint64_t val, int e, int s, bool sub) {
       59,  58,  57,  56,  56,  55,  54,  53};
 
   if (sub) {
-    while (extract64(sig, s - 1, 1) == 0) {
-      exp--;
-      sig <<= 1;
-    }
+    while (extract64(sig, s - 1, 1) == 0)
+      exp--, sig <<= 1;
+
     sig = (sig << 1) & make_mask64(0, s);
   }
 
@@ -358,9 +357,9 @@ float16_t f16_recip7(float16_t in) {
     [[fallthrough]];
   default: // +- normal
     uA.ui = recip7(uA.ui, 5, 10, softfloat_roundingMode, sub, &round_abnormal);
-    if (round_abnormal)
-      softfloat_exceptionFlags |=
-          softfloat_flag_inexact | softfloat_flag_overflow;
+    if (round_abnormal) {
+      softfloat_exceptionFlags |= softfloat_flag_inexact | softfloat_flag_overflow;
+    }
     break;
   }
 
@@ -401,9 +400,9 @@ float32_t f32_recip7(float32_t in) {
     [[fallthrough]];
   default: // +- normal
     uA.ui = recip7(uA.ui, 8, 23, softfloat_roundingMode, sub, &round_abnormal);
-    if (round_abnormal)
-      softfloat_exceptionFlags |=
-          softfloat_flag_inexact | softfloat_flag_overflow;
+    if (round_abnormal) {
+      softfloat_exceptionFlags |= softfloat_flag_inexact | softfloat_flag_overflow;
+    }
     break;
   }
 
@@ -444,9 +443,9 @@ float64_t f64_recip7(float64_t in) {
     [[fallthrough]];
   default: // +- normal
     uA.ui = recip7(uA.ui, 11, 52, softfloat_roundingMode, sub, &round_abnormal);
-    if (round_abnormal)
-      softfloat_exceptionFlags |=
-          softfloat_flag_inexact | softfloat_flag_overflow;
+    if (round_abnormal) {
+      softfloat_exceptionFlags |= softfloat_flag_inexact | softfloat_flag_overflow;
+    }
     break;
   }