From 0dab160fb2177d178eef3148c6a529e0855009e9 Mon Sep 17 00:00:00 2001
From: Christian Sarofeen <csarofeen@nvidia.com>
Date: Sat, 27 Aug 2022 10:28:01 -0400
Subject: [PATCH] Fix softmax bwd sizes. (#1890)

---
 benchmarks/cpp/nvfuser/softmax_backward.cpp | 32 ++++++++++-----------
 1 file changed, 16 insertions(+), 16 deletions(-)

diff --git a/benchmarks/cpp/nvfuser/softmax_backward.cpp b/benchmarks/cpp/nvfuser/softmax_backward.cpp
index 8fb35083c6dc7..be8dfda275d0e 100644
--- a/benchmarks/cpp/nvfuser/softmax_backward.cpp
+++ b/benchmarks/cpp/nvfuser/softmax_backward.cpp
@@ -177,13 +177,13 @@ NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp32)
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -201,13 +201,13 @@ NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp16)
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Outer_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -225,13 +225,13 @@ NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp32)
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -249,13 +249,13 @@ NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp16)
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 NVFUSER_BENCHMARK_RUN(NvFuserScheduler_Softmax_BWD_Inner_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -275,13 +275,13 @@ BENCHMARK(Baseline_Softmax_BWD_Outer_fp32)
 
 BENCHMARK(Baseline_Softmax_BWD_Outer_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 BENCHMARK(Baseline_Softmax_BWD_Outer_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -299,13 +299,13 @@ BENCHMARK(Baseline_Softmax_BWD_Outer_fp16)
 
 BENCHMARK(Baseline_Softmax_BWD_Outer_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 BENCHMARK(Baseline_Softmax_BWD_Outer_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -323,13 +323,13 @@ BENCHMARK(Baseline_Softmax_BWD_Inner_fp32)
 
 BENCHMARK(Baseline_Softmax_BWD_Inner_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 BENCHMARK(Baseline_Softmax_BWD_Inner_fp32)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
@@ -347,13 +347,13 @@ BENCHMARK(Baseline_Softmax_BWD_Inner_fp16)
 
 BENCHMARK(Baseline_Softmax_BWD_Inner_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{32768, 32 * 1024 * 1024}, {2, 16}})
+    ->Ranges({{32768, 16 * 1024 * 1024}, {2, 16}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();
 
 BENCHMARK(Baseline_Softmax_BWD_Inner_fp16)
     // ->RangeMultiplier(2)
-    ->Ranges({{2, 16}, {32768, 32 * 1024 * 1024}})
+    ->Ranges({{2, 16}, {32768, 16 * 1024 * 1024}})
     ->Unit(benchmark::kMicrosecond)
     ->UseManualTime();