Merge branch 'develop' into develop

src/fpga/api/fpga_api.h

@@ -14,36 +14,35 @@ limitations under the License. */
 
 #pragma once
 
+#include <stdint.h>
 #include <cstddef>
 #include <iostream>
 #include <limits>
 
 // memory management;
 
-namespace paddle {
-namespace mobile {
+namespace paddle_mobile {
 namespace fpga {
-namespace api {
 
 int open_device();
 int close_device();
 
-void *fpga_malloc(size_t size);
-void fpga_free(void *ptr);
-void fpga_copy(void *dst, const void *src, size_t num);
+void* fpga_malloc(size_t size);
+void fpga_free(void* ptr);
+void fpga_copy(void* dst, const void* src, size_t num);
 
 struct FpgaVersionArgs {
-  void *buf;
+  void* buf;
 };
 
 struct MemoryToPhysicalArgs {
-  const void *src;
+  const void* src;
   uint64_t physical;
 };
 
 struct MemoryCopyArgs {
-  void *src;
-  void *dst;
+  void* src;
+  void* dst;
   size_t size;
 };
 
@@ -51,38 +50,71 @@ struct FpgaQuantArgs {
   float scale;
 };
 
-struct FpgaBNArgs {};
+struct FpgaBNArgs {
+  bool enabled = false;
+  void* bias_addr;
+  void* scale_addr;
+};
+
+struct FpgaKernelArgs {
+  uint32_t width;
+  uint32_t height;
+  uint32_t stride_h;
+  uint32_t stride_w;
+};
+
+struct FpgaImageArgs {
+  uint32_t width;
+  uint32_t height;
+  uint32_t channels;
+  uint32_t pad_h;
+  uint32_t pad_w;
+};
 
 struct FpgaConvArgs {
-  bool enable_BN = false;
-  bool enable_Relu = false;
-  struct FpgaBNParam bn_parm;
+  bool relu_enabled;
+  struct FpgaBNArgs BNargs;
+  void* image_addr;
+  void* filter_addr;
+  void* bias_addr;
+  void* output_addr;
+  float quant_scale;
+  struct FpgaImageArgs image;
+  uint32_t filter_num;
+  uint32_t group_num;
+
+  struct FpgaKernelArgs kernel;
 };
 
 struct FpgaPoolArgs {
-  bool enable_BN = false;
-  struct FpgaBNParam bn_parm;
+  void* image_addr;
+  void* output_addr;
+  struct FpgaImageArgs image;
+  struct FpgaKernelArgs kernel;
 };
 
-struct FpgaEWAddArgs {  // only support X + Y
-  bool enable_Relu = false;
+struct FpgaEWAddArgs {
+  bool relu_enabled;
+  void* image0_addr;
+  void* image1_addr;
+  void* result_addr;
+  uint32_t const0;
+  uint32_t const1;
+  uint32_t data_len;  // aligned element count
 };
 
-int ComputeFpgaConv(struct FpgaConvArgs);
-int ComputeFpgaPool(struct FpgaPoolArgs);
-int ComputeFpgaEWAdd(struct FpgaEWAddArgs);
+int ComputeFpgaConv(struct FpgaConvArgs args);
+int ComputeFpgaPool(struct FpgaPoolArgs args);
+int ComputeFpgaEWAdd(struct FpgaEWAddArgs args);
 
-#define IOCTL_FPGA_MAGIC 'FPGA'
+#define IOCTL_FPGA_MAGIC 'CNN'
 #define IOCTL_VERSION _IOW(IOCTL_FPGA_MAGIC, 1, struct FpgaVersionArgs)
 #define IOCTL_GET_QUANT _IOW(IOCTL_FPGA_MAGIC, 2, struct FpgaQuantArgs)
-#define IOCTL_SET_QUANT _IOW(IOCTL_FPGA_MAGIC, 3, struct FpgaArgs)
+#define IOCTL_SET_QUANT _IOW(IOCTL_FPGA_MAGIC, 3, struct FpgaQuantArgs)
 #define IOCTL_MEM_COPY _IOW(IOCTL_FPGA_MAGIC, 11, struct MemoryCopyArgs)
-#define IOCTL_MEM_TOPHY _IOW(IOCTL_FPGA_MAGIC, 12, struct MemoryToPhysicalArgs)
 #define IOCTL_CONFIG_CONV _IOW(IOCTL_FPGA_MAGIC, 21, struct FpgaConvArgs)
 #define IOCTL_CONFIG_POOLING _IOW(IOCTL_FPGA_MAGIC, 22, struct FpgaPoolArgs)
 #define IOCTL_CONFIG_EW _IOW(IOCTL_FPGA_MAGIC, 23, struct FpgaEWAddArgs)
 
-}  // namespace api
 }  // namespace fpga
-}  // namespace mobile
-}  // namespace paddle
+}  // namespace paddle_mobile

src/operators/op_param.h

@@ -22,6 +22,9 @@ limitations under the License. */
 #include "framework/scope.h"
 #include "framework/tensor.h"
 #include "framework/variable.h"
+#ifdef PADDLE_MOBILE_FPGA
+#include "fpga/api/fpga_api.h"
+#endif
 
 namespace paddle_mobile {
 namespace operators {
@@ -256,6 +259,15 @@ class ElementwiseAddParam : OpParam {
   Tensor *input_y_;
   Tensor *out_;
   int axis_;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::FpgaEWAddArgs fpga_EW_add_args;
+
+ public:
+  const fpga::FpgaEWAddArgs &FpgaArgs() const { return fpga_EW_add_args; }
+  void SetFpgaArgs(const fpga::FpgaEWAddArgs &args) { fpga_EW_add_args = args; }
+#endif
 };
 
 #ifdef FUSION_ELEMENTWISEADDRELU_OP
@@ -450,80 +462,15 @@ class PoolParam : public OpParam {
   vector<int> paddings_;
   bool ceil_mode_;
   bool global_pooling_ = false;
-};
-#endif
-
-#ifdef FUSION_POOLBN_OP
-class FusionPoolBNParam : OpParam {
- public:
-  FusionPoolBNParam(const VariableNameMap &inputs,
-                    const VariableNameMap &outputs, const AttributeMap &attrs,
-                    const Scope &scope) {
-    input_ = InputXFrom<LoDTensor>(inputs, scope);
-    pooling_type_ = GetAttr<string>("pooling_type", attrs);
-    ksize_ = GetAttr<vector<int>>("ksize", attrs);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    ceil_mode_ = GetAttr<bool>("ceil_mode", attrs);
-    global_pooling_ = GetAttr<bool>("global_pooling", attrs);
-    output_y_ = OutputYFrom<LoDTensor>(outputs, scope);
-    input_bias_ = InputBiasFrom<LoDTensor>(inputs, scope);
-    input_mean_ = InputMeanFrom<LoDTensor>(inputs, scope);
-    input_scale_ = InputScaleFrom<LoDTensor>(inputs, scope);
-    input_variance_ = InputVarianceFrom<LoDTensor>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
-  }
-  const Tensor *Input() const { return input_; }
-
-  const string &PoolingType() const { return pooling_type_; }
-
-  const vector<int> &Ksize() const { return ksize_; }
-
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  bool isCeilMode() const { return ceil_mode_; }
-
-  bool isGlobalPooling() const { return global_pooling_; }
-
-  Tensor *OutputY() const { return output_y_; }
-
-  const Tensor *InputBias() const { return input_bias_; }
-
-  const Tensor *InputMean() const { return input_mean_; }
-
-  const Tensor *InputScale() const { return input_scale_; }
-
-  const Tensor *InputVariance() const { return input_variance_; }
-
-  const float &Epsilon() const { return epsilon_; }
-
-  const float &Momentum() const { return momentum_; }
-
-  const bool &IsTest() const { return is_test_; }
-
-  const string &DataFormat() const { return data_format_; }
+#ifdef PADDLE_MOBILE_FPGA
 
  private:
-  Tensor *input_;
-  string pooling_type_;
-  vector<int> ksize_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  bool ceil_mode_;
-  bool global_pooling_ = false;
-  Tensor *output_y_;
-  Tensor *input_bias_;
-  Tensor *input_mean_;
-  Tensor *input_scale_;
-  Tensor *input_variance_;
-  float epsilon_;
-  float momentum_;
-  bool is_test_;
-  string data_format_;
+  fpga::FpgaPoolArgs fpga_pool_args;
+
+ public:
+  const fpga::FpgaPoolArgs &FpgaArgs() const { return fpga_pool_args; }
+  void SetFpgaArgs(const fpga::FpgaPoolArgs &args) { fpga_pool_args = args; }
+#endif
 };
 #endif
 
@@ -704,7 +651,7 @@ class MultiClassNMSParam : public OpParam {
 class FeedParam : public OpParam {
  public:
   FeedParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
-            const AttributeMap &attrs, Scope &scope) {
+            const AttributeMap &attrs, Scope const &scope) {
     input_x_ = InputXFrom<LoDTensor>(inputs, scope);
     out_ = OutFrom<LoDTensor>(outputs, scope);
     auto var = scope.Var("batch_size");
@@ -983,6 +930,15 @@ class FusionFcParam : public OpParam {
   int x_num_col_dims_;
   int y_num_col_dims_;
   int axis_;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::FpgaConvArgs fpga_conv_args;
+
+ public:
+  const fpga::FpgaConvArgs &FpgaArgs() const { return fpga_conv_args; }
+  void SetFpgaArgs(const fpga::FpgaConvArgs &args) { fpga_conv_args = args; }
+#endif
 };
 
 #ifdef FUSION_FCRELU_OP
@@ -1032,6 +988,15 @@ class FusionConvAddParam : public OpParam {
   vector<int> paddings_;
   vector<int> dilations_;
   int groups;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::FpgaConvArgs fpga_conv_args;
+
+ public:
+  const fpga::FpgaConvArgs &FpgaArgs() const { return fpga_conv_args; }
+  void SetFpgaArgs(const fpga::FpgaConvArgs &args) { fpga_conv_args = args; }
+#endif
 };
 
 Print &operator<<(Print &printer, const FusionConvAddParam &conv_param);
@@ -1128,6 +1093,15 @@ class FusionConvAddBNReluParam : public OpParam {
   bool is_test_;
   Tensor *new_bias_;
   Tensor *new_scale_;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::FpgaConvArgs fpga_conv_args;
+
+ public:
+  const fpga::FpgaConvArgs &FpgaArgs() const { return fpga_conv_args; }
+  void SetFpgaArgs(const fpga::FpgaConvArgs &args) { fpga_conv_args = args; }
+#endif
 };
 #endif
 
@@ -1213,6 +1187,15 @@ class FusionConvAddBNParam : public OpParam {
   bool is_test_;
   Tensor *new_bias_;
   Tensor *new_scale_;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::FpgaConvArgs fpga_conv_args;
+
+ public:
+  const fpga::FpgaConvArgs &FpgaArgs() const { return fpga_conv_args; }
+  void SetFpgaArgs(const fpga::FpgaConvArgs &args) { fpga_conv_args = args; }
+#endif
 };
 #endif
 
@@ -1426,9 +1409,5 @@ class DropoutParam : public OpParam {
 };
 #endif
 
-#ifdef REGION_OP
-class RegionParam : public OpParam {};
-#endif
-
 }  // namespace operators
 }  // namespace paddle_mobile

test/common/test_gemm.cpp

@@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include <cstdlib>
+#include <ctime>
 #include <iostream>
 #include "../test_helper.h"
 #include "common/log.h"
@@ -20,13 +22,21 @@ limitations under the License. */
 
 #define a(i, j) a[(i)*lda + (j)]
 #define b(i, j) b[(i)*ldb + (j)]
+#define c(i, j) c[(i)*ldc + (j)]
 #define c1(i, j) c1[(i)*ldc + (j)]
 
-#define m 62
-#define n 63
-#define k 74
+void print_matirx(int m, int n, int ldc, float *c) {
+  for (int i = 0; i < m; ++i) {
+    std::cout << c(i, 0);
+    for (int j = 1; j < n; ++j) {
+      std::cout << " | " << c(i, j);
+    }
+    std::cout << std::endl;
+  }
+  std::cout << std::endl;
+}
 
-int main() {
+int do_sgemm(int m, int n, int k, bool relu, int t1, int t2, int pr) {
   int lda = k;
   int ldb = n;
   int ldc = n;
@@ -39,44 +49,88 @@ int main() {
       static_cast<float *>(paddle_mobile::memory::Alloc(sizeof(float) * m * n));
   float *c1 =
       static_cast<float *>(paddle_mobile::memory::Alloc(sizeof(float) * m * n));
+  float *scale =
+      static_cast<float *>(paddle_mobile::memory::Alloc(sizeof(float) * m));
+  float *bias =
+      static_cast<float *>(paddle_mobile::memory::Alloc(sizeof(float) * m));
 
+  srand(unsigned(time(0)));
   for (int i = 0; i < m * k; ++i) {
-    a[i] = 2;
+    a[i] = t1 + rand() % t2;
   }
   for (int i = 0; i < k * n; ++i) {
-    b[i] = 2;
+    b[i] = t1 + rand() % t2;
   }
-  for (int i = 0; i < m * n; ++i) {
-    c[i] = 2;
-    c1[i] = 2;
+  for (int i = 0; i < m; ++i) {
+    scale[i] = t1 + rand() % t2;
   }
-
-  auto time1 = time();
-  //  paddle_mobile::operators::math::Sgemm(m, n, k, 0.9, a, lda, b, ldb, 0.3,
-  //  c,
-  //                                        ldc);
-  auto time2 = time();
-  DLOG << "gemm cost :" << time_diff(time1, time2) << "ms\n";
-  for (int i = 0; i < m * n; ++i) {
-    std::cout << c[i] << " | ";
-    if (i % n == (n - 1)) {
-      std::cout << std::endl;
-    }
+  for (int i = 0; i < m; ++i) {
+    bias[i] = t1 + rand() % t2;
   }
-  for (int j = 0; j < n; ++j) {
-    for (int i = 0; i < m; ++i) {
-      c1(i, j) *= 0.3;
-      for (int p = 0; p < k; ++p) {
-        c1(i, j) += 0.9 * a(i, p) * b(p, j);
+
+  for (int i = 0; i < m; ++i) {
+    for (int j = 0; j < n; ++j) {
+      float r = 0;
+      for (int p = 0; p < k; p++) {
+        r += a(i, p) * b(p, j);
       }
+      r *= scale[i];
+      r += bias[i];
+      if (relu && (r < 0)) {
+        r = 0;
+      }
+      c1(i, j) = r;
     }
   }
-  std::cout << "正确结果对比:" << std::endl;
+
+  paddle_mobile::operators::math::SgemmWithBn(m, n, k, 0.9, a, lda, b, ldb, 0.3,
+                                              c, ldc, relu, scale, bias);
+  int eq = 0;
+  int neq = 0;
   for (int i = 0; i < m * n; ++i) {
-    std::cout << c1[i] << " | ";
-    if (i % n == (n - 1)) {
-      std::cout << std::endl;
+    if (static_cast<int>(c[i]) == static_cast<int>(c1[i])) {
+      ++eq;
+    } else {
+      ++neq;
     }
   }
+
+  if (pr > 0) {
+    std::cout << "A:" << std::endl;
+    print_matirx(m, k, lda, a);
+    std::cout << "B:" << std::endl;
+    print_matirx(k, n, ldb, b);
+    std::cout << "C:" << std::endl;
+    print_matirx(m, n, ldc, c);
+    std::cout << "C1:" << std::endl;
+    print_matirx(m, n, ldc, c1);
+  }
+
+  std::cout << "mnk=" << m << " " << n << " " << k << " relu=" << relu
+            << "   eq=" << eq << " neq=" << neq << std::endl;
+
+  paddle_mobile::memory::Free(a);
+  paddle_mobile::memory::Free(b);
+  paddle_mobile::memory::Free(c);
+  paddle_mobile::memory::Free(c1);
+  paddle_mobile::memory::Free(scale);
+  paddle_mobile::memory::Free(bias);
+
+  return 0;
+}
+
+int main() {
+  do_sgemm(9, 9, 9, true, 10, 10, 10);
+  do_sgemm(10, 6, 12, false, 10, 10, 0);
+  do_sgemm(512, 256, 384, false, 10, 10, 0);
+  do_sgemm(1366, 768, 256, false, 10, 10, 0);
+  do_sgemm(1255, 755, 333, false, 10, 10, 0);
+  do_sgemm(555, 777, 999, false, 10, 10, 0);
+
+  do_sgemm(10, 6, 12, true, -4, 10, 0);
+  do_sgemm(512, 256, 384, true, -4, 10, 0);
+  do_sgemm(1366, 768, 256, true, -4, 10, 0);
+  do_sgemm(1255, 755, 333, true, -4, 10, 0);
+  do_sgemm(555, 777, 999, true, -4, 10, 0);
   return 0;
 }

tools/op.cmake

@@ -75,11 +75,9 @@ if ("FPGAnets" IN_LIST NET)
   set(FUSION_CONVADDRELU_OP ON)
   set(FUSION_CONVADDBNRELU_OP ON)
   set(FUSION_CONVADDBN_OP ON)
-  set(FUSION_POOLBN_OP ON)
   set(FUSION_ELEMENTWISEADDRELU_OP ON)
   set(FUSION_FC_OP ON)
   set(FUSION_FCRELU_OP ON)
-  set(REGION_OP ON)
   set(POOL_OP ON)
   set(CONCAT_OP ON)
   set(SOFTMAX_OP ON)