Merge branch 'add_b2i_benchmark' into 'master'

add b2i and i2b benchmark

See merge request !1136

test/ccbenchmark/mace/ops/buffer_to_image_benchmark.cc

@@ -14,8 +14,8 @@
 
 #ifdef MACE_ENABLE_OPENCL
 
-#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/benchmark_utils/test_benchmark.h"
+#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/ops/opencl/buffer_transformer.h"
 #include "mace/ops/ops_test_util.h"
 
@@ -25,29 +25,23 @@ namespace test {
 
 namespace {
 template <DeviceType D, typename T>
-void FilterBufferToImage(int iters,
-                         int out_channel, int in_channel,
-                         int height, int width) {
+void FilterBufferToImage(
+    int iters, int out_channel, int in_channel, int height, int width) {
   mace::testing::StopTiming();
 
   OpsTestNet net;
-  OpContext context(net.ws(),
-                    OpTestContext::Get()->GetDevice(DeviceType::GPU));
+  OpContext context(net.ws(), OpTestContext::Get()->GetDevice(DeviceType::GPU));
 
   // Add input data
-  net.AddRandomInput<D, T>("Input",
-                           {out_channel, in_channel, height, width});
+  net.AddRandomInput<D, T>("Input", {out_channel, in_channel, height, width});
   // Create output
   Tensor *b2i_output = net.ws()->CreateTensor(
       "B2IOutput", context.device()->allocator(), DataTypeToEnum<T>::value);
 
   auto transform_func = [&]() {
     OpenCLBufferTransformer(MemoryType::GPU_BUFFER, MemoryType::GPU_IMAGE)
-        .Transform(&context,
-                   net.ws()->GetTensor("Input"),
-                   OpenCLBufferType::IN_OUT_CHANNEL,
-                   MemoryType::GPU_IMAGE,
-                   0,
+        .Transform(&context, net.ws()->GetTensor("Input"),
+                   OpenCLBufferType::CONV2D_FILTER, MemoryType::GPU_IMAGE, 0,
                    b2i_output);
   };
 
@@ -64,36 +58,148 @@ void FilterBufferToImage(int iters,
 }
 }  // namespace
 
-#define MACE_BM_B2I_MACRO(O, I, H, W, TYPE, DEVICE)                  \
-  static void MACE_BM_B2I_##O##_##I##_##H##_##W##_##TYPE##_##DEVICE( \
+#define MACE_BM_FILTER_B2I_MACRO(O, I, H, W, TYPE, DEVICE)                  \
+  static void MACE_BM_FILTER_B2I_##O##_##I##_##H##_##W##_##TYPE##_##DEVICE( \
       int iters) {                                                          \
     const int64_t tot = static_cast<int64_t>(iters) * O * I * H * W;        \
     mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                     \
     FilterBufferToImage<DEVICE, TYPE>(iters, O, I, H, W);                   \
   }                                                                         \
-  MACE_BENCHMARK(MACE_BM_B2I_##O##_##I##_##H##_##W##_##TYPE##_##DEVICE)
-
-#define MACE_BM_B2I(O, I, H, W)              \
-  MACE_BM_B2I_MACRO(O, I, H, W, float, GPU); \
-  MACE_BM_B2I_MACRO(O, I, H, W, half, GPU);
-
-MACE_BM_B2I(5, 3, 3, 3);
-MACE_BM_B2I(5, 3, 7, 7);
-MACE_BM_B2I(32, 16, 1, 1);
-MACE_BM_B2I(32, 16, 3, 3);
-MACE_BM_B2I(32, 16, 5, 5);
-MACE_BM_B2I(32, 16, 7, 7);
-MACE_BM_B2I(64, 32, 1, 1);
-MACE_BM_B2I(64, 32, 3, 3);
-MACE_BM_B2I(64, 32, 5, 5);
-MACE_BM_B2I(64, 32, 7, 7);
-MACE_BM_B2I(128, 64, 1, 1);
-MACE_BM_B2I(128, 64, 3, 3);
-MACE_BM_B2I(128, 32, 1, 1);
-MACE_BM_B2I(128, 32, 3, 3);
-MACE_BM_B2I(256, 32, 1, 1);
-MACE_BM_B2I(256, 32, 3, 3);
+  MACE_BENCHMARK(MACE_BM_FILTER_B2I_##O##_##I##_##H##_##W##_##TYPE##_##DEVICE)
+
+#define MACE_BM_FILTER_B2I(O, I, H, W)              \
+  MACE_BM_FILTER_B2I_MACRO(O, I, H, W, float, GPU); \
+  MACE_BM_FILTER_B2I_MACRO(O, I, H, W, half, GPU);
+
+MACE_BM_FILTER_B2I(5, 3, 3, 3);
+MACE_BM_FILTER_B2I(5, 3, 7, 7);
+MACE_BM_FILTER_B2I(32, 16, 1, 1);
+MACE_BM_FILTER_B2I(32, 16, 3, 3);
+MACE_BM_FILTER_B2I(32, 16, 5, 5);
+MACE_BM_FILTER_B2I(32, 16, 7, 7);
+MACE_BM_FILTER_B2I(64, 32, 1, 1);
+MACE_BM_FILTER_B2I(64, 32, 3, 3);
+MACE_BM_FILTER_B2I(64, 32, 5, 5);
+MACE_BM_FILTER_B2I(64, 32, 7, 7);
+MACE_BM_FILTER_B2I(128, 64, 1, 1);
+MACE_BM_FILTER_B2I(128, 64, 3, 3);
+MACE_BM_FILTER_B2I(128, 32, 1, 1);
+MACE_BM_FILTER_B2I(128, 32, 3, 3);
+MACE_BM_FILTER_B2I(256, 32, 1, 1);
+MACE_BM_FILTER_B2I(256, 32, 3, 3);
+
+namespace {
+template <DeviceType D, typename T>
+void InOutBufferToImage(
+    int iters, int batch, int height, int width, int channel) {
+  mace::testing::StopTiming();
+
+  OpsTestNet net;
+  OpContext context(net.ws(), OpTestContext::Get()->GetDevice(DeviceType::GPU));
+
+  // Add input data
+  net.AddRandomInput<D, T>("Input", {batch, height, width, channel});
+  // Create output
+  Tensor *b2i_output = net.ws()->CreateTensor(
+      "B2IOutput", context.device()->allocator(), DataTypeToEnum<T>::value);
+
+  auto transform_func = [&]() {
+    OpenCLBufferTransformer(MemoryType::GPU_BUFFER, MemoryType::GPU_IMAGE)
+        .Transform(&context, net.ws()->GetTensor("Input"),
+                   OpenCLBufferType::IN_OUT_CHANNEL, MemoryType::GPU_IMAGE, 0,
+                   b2i_output);
+  };
+
+  for (int i = 0; i < 5; ++i) {
+    transform_func();
+  }
+  net.Sync();
+
+  mace::testing::StartTiming();
+  while (iters--) {
+    transform_func();
+  }
+  net.Sync();
+}
+}  // namespace
+
+#define MACE_BM_IN_OUT_B2I_MACRO(N, H, W, C, TYPE, DEVICE)                  \
+  static void MACE_BM_IN_OUT_B2I_##N##_##H##_##W##_##C##_##TYPE##_##DEVICE( \
+      int iters) {                                                          \
+    const int64_t tot = static_cast<int64_t>(iters) * N * H * W * C;        \
+    mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                     \
+    InOutBufferToImage<DEVICE, TYPE>(iters, N, H, W, C);                    \
+  }                                                                         \
+  MACE_BENCHMARK(MACE_BM_IN_OUT_B2I_##N##_##H##_##W##_##C##_##TYPE##_##DEVICE)
+
+#define MACE_BM_IN_OUT_B2I(N, H, W, C)              \
+  MACE_BM_IN_OUT_B2I_MACRO(N, H, W, C, float, GPU); \
+  MACE_BM_IN_OUT_B2I_MACRO(N, H, W, C, half, GPU);
+
+MACE_BM_IN_OUT_B2I(256, 1, 1, 32);
+MACE_BM_IN_OUT_B2I(256, 3, 3, 32);
+MACE_BM_IN_OUT_B2I(1, 4096, 4096, 3);
+
+namespace {
+template <DeviceType D, typename T>
+void InOutImageToBuffer(
+    int iters, int batch, int height, int width, int channel) {
+  mace::testing::StopTiming();
+
+  OpsTestNet net;
+  OpContext context(net.ws(), OpTestContext::Get()->GetDevice(DeviceType::GPU));
+
+  // Add input data
+  net.AddRandomInput<D, T>("Input", {batch, height, width, channel});
+  // Create output
+  Tensor *b2i_output = net.ws()->CreateTensor(
+      "B2IOutput", context.device()->allocator(), DataTypeToEnum<T>::value);
+
+  auto transform_func_b2i = [&]() {
+    OpenCLBufferTransformer(MemoryType::GPU_BUFFER, MemoryType::GPU_IMAGE)
+        .Transform(&context, net.ws()->GetTensor("Input"),
+                   OpenCLBufferType::IN_OUT_CHANNEL, MemoryType::GPU_IMAGE, 0,
+                   b2i_output);
+  };
+  transform_func_b2i();
+
+  Tensor *i2b_output = net.ws()->CreateTensor(
+      "I2BOutput", context.device()->allocator(), DataTypeToEnum<T>::value);
+  auto transform_func_i2b = [&]() {
+    OpenCLBufferTransformer(MemoryType::GPU_IMAGE, MemoryType::GPU_BUFFER)
+        .Transform(&context, b2i_output, OpenCLBufferType::IN_OUT_CHANNEL,
+                   MemoryType::GPU_BUFFER, 0, i2b_output);
+  };
+
+  for (int i = 0; i < 5; ++i) {
+    transform_func_i2b();
+  }
+  net.Sync();
+
+  mace::testing::StartTiming();
+  while (iters--) {
+    transform_func_i2b();
+  }
+  net.Sync();
+}
+}  // namespace
+
+#define MACE_BM_IN_OUT_I2B_MACRO(N, H, W, C, TYPE, DEVICE)                  \
+  static void MACE_BM_IN_OUT_I2B_##N##_##H##_##W##_##C##_##TYPE##_##DEVICE( \
+      int iters) {                                                          \
+    const int64_t tot = static_cast<int64_t>(iters) * N * H * W * C;        \
+    mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                     \
+    InOutImageToBuffer<DEVICE, TYPE>(iters, N, H, W, C);                    \
+  }                                                                         \
+  MACE_BENCHMARK(MACE_BM_IN_OUT_I2B_##N##_##H##_##W##_##C##_##TYPE##_##DEVICE)
+
+#define MACE_BM_IN_OUT_I2B(N, H, W, C)              \
+  MACE_BM_IN_OUT_I2B_MACRO(N, H, W, C, float, GPU); \
+  MACE_BM_IN_OUT_I2B_MACRO(N, H, W, C, half, GPU);
 
+MACE_BM_IN_OUT_I2B(256, 1, 1, 32);
+MACE_BM_IN_OUT_I2B(256, 3, 3, 32);
+MACE_BM_IN_OUT_I2B(1, 4096, 4096, 3);
 }  // namespace test
 }  // namespace ops
 }  // namespace mace