Merge branch 'sgemm_conv' into 'master'

Replace gemm to sgemm

See merge request !794

mace/core/buffer.h

@@ -469,6 +469,7 @@ class ScratchBuffer: public Buffer {
 
   MaceStatus GrowSize(index_t size) {
     if (size > size_) {
+      MACE_CHECK(offset_ == 0, "scratch is being used, cannot grow size");
       return Resize(size);
     }
     return MaceStatus::MACE_SUCCESS;
@@ -487,8 +488,12 @@ class ScratchBuffer: public Buffer {
     return slice;
   }
 
-  void Rewind() {
-    offset_ = 0;
+  void Rewind(index_t offset = 0) {
+    offset_ = offset;
+  }
+
+  index_t offset() const {
+    return offset_;
   }
 
  private:

mace/kernels/arm/conv_2d_neon.h

@@ -16,6 +16,7 @@
 #define MACE_KERNELS_ARM_CONV_2D_NEON_H_
 
 #include "mace/core/types.h"
+#include "mace/kernels/sgemm.h"
 
 namespace mace {
 namespace kernels {
@@ -27,7 +28,9 @@ void Conv2dNeonK1x1S1(const float *input,
                       const index_t width,
                       const index_t in_channels,
                       const index_t out_channels,
-                      float *output);
+                      float *output,
+                      SGemm *sgemm,
+                      ScratchBuffer *scratch_buffer);
 
 void Conv2dNeonK3x3S1(const float *input,
                       const float *filter,

mace/kernels/arm/conv_2d_neon_1x1.cc

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #include "mace/kernels/arm/conv_2d_neon.h"
-#include "mace/kernels/gemm.h"
 
 namespace mace {
 namespace kernels {
@@ -25,11 +24,23 @@ void Conv2dNeonK1x1S1(const float *input,
                       const index_t width,
                       const index_t in_channels,
                       const index_t out_channels,
-                      float *output) {
+                      float *output,
+                      SGemm *sgemm,
+                      ScratchBuffer *scratch_buffer) {
   for (index_t b = 0; b < batch; ++b) {
-    Gemm(filter, input + b * in_channels * height * width, 1, out_channels,
-         in_channels, height * width,
-         output + b * out_channels * height * width);
+    sgemm->Run(filter,
+               input + b * in_channels * height * width,
+               1,
+               out_channels,
+               in_channels,
+               in_channels,
+               height * width,
+               false,
+               false,
+               true,
+               false,
+               output + b * out_channels * height * width,
+               scratch_buffer);
   }
 }
 

mace/kernels/arm/conv_winograd.cc

@@ -12,13 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <math.h>
 #include <algorithm>
 
 #include "mace/kernels/arm/conv_winograd.h"
 #include "mace/kernels/gemm.h"
-#include "mace/utils/logging.h"
-#include "mace/utils/utils.h"
 
 namespace mace {
 namespace kernels {
@@ -247,30 +244,38 @@ void BatchGemm(const float *input,
                index_t out_channels,
                index_t tile_count,
                int out_tile_size,
-               float *output) {
-  const index_t filter_stride = out_channels * in_channels;
+               float *output,
+               SGemm *sgemm,
+               ScratchBuffer *scratch_buffer) {
   const int in_tile_area = (out_tile_size + 2) * (out_tile_size + 2);
   const index_t in_batch_size = in_tile_area * in_channels * tile_count;
-  const index_t in_stride = in_channels * tile_count;
   const index_t out_batch_size = in_tile_area * out_channels * tile_count;
-  const index_t out_stride = out_channels * tile_count;
 
-  if (batch == 1) {
-    Gemm(filter, input, in_tile_area, out_channels, in_channels, tile_count,
-         output);
-  } else {
-#pragma omp parallel for collapse(2)
-    for (int b = 0; b < batch; ++b) {
-      for (int i = 0; i < in_tile_area; ++i) {
-        const float *in_ptr = input + b * in_batch_size + i * in_stride;
-        const float *filter_ptr = filter + i * filter_stride;
-        float *out_ptr = output + b * out_batch_size + i * out_stride;
-        Gemm(filter_ptr, in_ptr, 1, out_channels, /* rows */
-             in_channels,                         /* K */
-             tile_count,                          /* cols */
-             out_ptr);
-      }
+  index_t scratch_buffer_offset = 0;
+  if (scratch_buffer) {
+    scratch_buffer_offset = scratch_buffer->offset();
   }
+  // 'batch' is not gemm batch, 'in_tile_area' is. gemm is not thread safe,
+  // so we loop batch using single thread.
+  // Scratch buffer should be rewind to the initial position to use same
+  // scratch memory for each batch.
+  for (int b = 0; b < batch; ++b) {
+    if (scratch_buffer) {
+      scratch_buffer->Rewind(scratch_buffer_offset);
+    }
+    sgemm->Run(filter,
+               input + b * in_batch_size,
+               in_tile_area,
+               out_channels,
+               in_channels,
+               in_channels,
+               tile_count,
+               false,
+               false,
+               true,
+               false,
+               output + b * out_batch_size,
+               scratch_buffer);
   }
 }
 
@@ -613,7 +618,9 @@ void WinoGradConv3x3s1(const float *input,
                        const int out_tile_size,
                        float *transformed_input,
                        float *transformed_output,
-                       float *output) {
+                       float *output,
+                       SGemm *sgemm,
+                       ScratchBuffer *scratch_buffer) {
   index_t out_height = in_height - 2;
   index_t out_width = in_width - 2;
   index_t tile_height_count =
@@ -636,7 +643,8 @@ void WinoGradConv3x3s1(const float *input,
   }
 
   BatchGemm(transformed_input, transformed_filter, batch, in_channels,
-            out_channels, tile_count, out_tile_size, transformed_output);
+            out_channels, tile_count, out_tile_size, transformed_output,
+            sgemm, scratch_buffer);
 
   switch (out_tile_size) {
     case 2:
@@ -660,7 +668,9 @@ void WinoGradConv3x3s1(const float *input,
                        const index_t in_channels,
                        const index_t out_channels,
                        const int out_tile_size,
-                       float *output) {
+                       float *output,
+                       SGemm *sgemm,
+                       ScratchBuffer *scratch_buffer) {
   index_t out_height = in_height - 2;
   index_t out_width = in_width - 2;
   index_t tile_height_count =
@@ -692,7 +702,7 @@ void WinoGradConv3x3s1(const float *input,
 
   WinoGradConv3x3s1(input, transformed_filter, batch, in_height, in_width,
                     in_channels, out_channels, out_tile_size, transformed_input,
-                    transformed_output, output);
+                    transformed_output, output, sgemm, scratch_buffer);
 
   delete[] transformed_input;
   delete[] transformed_filter;

mace/kernels/arm/conv_winograd.h

@@ -20,6 +20,7 @@
 #endif
 
 #include "mace/core/types.h"
+#include "mace/kernels/sgemm.h"
 
 namespace mace {
 namespace kernels {
@@ -42,7 +43,9 @@ void WinoGradConv3x3s1(const float *input,
                        const index_t in_channels,
                        const index_t out_channels,
                        const int out_tile_size,
-                       float *output);
+                       float *output,
+                       SGemm *sgemm,
+                       ScratchBuffer *scratch_buffer);
 
 void WinoGradConv3x3s1(const float *input,
                        const float *transformed_filter,
@@ -54,7 +57,9 @@ void WinoGradConv3x3s1(const float *input,
                        const int out_tile_size,
                        float *transformed_input,
                        float *transformed_output,
-                       float *output);
+                       float *output,
+                       SGemm *sgemm,
+                       ScratchBuffer *scratch_buffer);
 
 void ConvRef3x3s1(const float *input,
                   const float *filter,

mace/kernels/arm/conv_winograd_test.cc

@@ -65,9 +65,10 @@ TEST(ConvWinogradTest, winograd) {
   kernels::ConvRef3x3s1(input_data, filter_data, batch, in_height, in_width,
                         in_channels, out_channels, output_data_ref);
 
+  SGemm sgemm;
   kernels::WinoGradConv3x3s1(input_data, filter_data, batch, in_height,
                              in_width, in_channels, out_channels, 6,
-                             output_data);
+                             output_data, &sgemm, nullptr);
 
   // test
   for (index_t i = 0; i < output_size; ++i) {

mace/kernels/conv_2d.h

@@ -483,6 +483,16 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
           * sizeof(float);
       total_scratch_size += padded_output_size;
     }
+    // scratch for sgemm
+    if (use_neon_1x1_s1) {
+      total_scratch_size +=
+          (input_batch * input_height * input_width
+              * (input_channels + channels)) * sizeof(float);
+    } else if (use_winograd) {
+      total_scratch_size +=
+          (transformed_input_size + transformed_output_size) * sizeof(float);
+    }
+
     // Init scratch buffer
     scratch_->Rewind();
     scratch_->GrowSize(total_scratch_size);
@@ -547,7 +557,9 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
                           winograd_out_tile_size,
                           transformed_input_data,
                           transformed_output_data,
-                          pad_output);
+                          pad_output,
+                          &sgemm_,
+                          scratch_);
       };
     } else if (use_neon_3x3_s1) {
       conv_func = [=](const float *pad_input, float *pad_output) {
@@ -574,7 +586,9 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
                          extra_input_width,
                          input_channels,
                          channels,
-                         pad_output);
+                         pad_output,
+                         &sgemm_,
+                         scratch_);
       };
     } else if (use_neon_5x5_s1) {
       conv_func = [=](const float *pad_input, float *pad_output) {
@@ -722,6 +736,7 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
   Tensor transformed_filter_;
   bool is_filter_transformed_;
   ScratchBuffer *scratch_;
+  SGemm sgemm_;
 };
 
 template<>

mace/kernels/matmul.h

@@ -89,6 +89,9 @@ struct MatMulFunctor : OpKernel {
     const index_t height_b = B->dim(rank - 2);
     const index_t width_b = B->dim(rank - 1);
 
+    auto scratch_buffer = context_->workspace()->GetScratchBuffer(D);
+    scratch_buffer->Rewind();
+
     sgemm_.Run(a_ptr_base,
                b_ptr_base,
                batch,
@@ -101,7 +104,7 @@ struct MatMulFunctor : OpKernel {
                A->is_weight(),
                B->is_weight(),
                c_ptr_base,
-               context_->workspace()->GetScratchBuffer(D));
+               scratch_buffer);
     return MACE_SUCCESS;
   }
 

mace/kernels/sgemm.cc

@@ -44,7 +44,6 @@ void SGemm::operator()(const MatrixMap<const float> &lhs,
   }
 
   if (scratch_buffer != nullptr) {
-    scratch_buffer->Rewind();
     index_t total_size = result->size();
     if (!lhs.is_const()) {
       total_size += lhs.size();
@@ -54,7 +53,6 @@ void SGemm::operator()(const MatrixMap<const float> &lhs,
     }
     scratch_buffer->GrowSize(total_size * sizeof(float));
 
-    scratch_buffer->Rewind();
     if (!lhs.is_const()) {
       packed_lhs_.reset(new Tensor(scratch_buffer->Scratch(
           lhs.size() * sizeof(float)), DT_FLOAT));