Merge branch 'develop' of https://github.com/PaddlePaddle/paddle-mobile into backup

src/fpga/V1/api.cpp

@@ -207,7 +207,7 @@ void format_DWDconv_filter(framework::Tensor *filter_tensor, float *scale_ptr,
   //      framework::make_ddim({num, 1, height, width});
   //  filter_tensor->Resize(dims_new);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int8_t));
+  filter_tensor->set_type(typeid(int16_t));
 }
 
 void format_fc_filter(framework::Tensor *filter_tensor, float max_value) {
@@ -466,24 +466,9 @@ void expand_EW_arg(EWAddArgs *arg) {
   uint64_t image_amount_per_row =
       align_to_x((uint64_t)args.image0.width * (uint64_t)args.image0.channels,
                  IMAGE_ALIGNMENT);
-  //////////////////////////////////////////////////////////
-  // temporary modify for EW and DMA problem
-  uint64_t image_image_pixel = 0;
-  if ((args.image0.width * args.image0.channels) >= 24576) {
-    if ((args.image0.width * args.image0.channels) % 32 != 0) {
-      DLOG << "EW parameter can not be support";
-    } else {
-      image_amount_per_row = image_amount_per_row / 2;
-      image_image_pixel = ((uint64_t)args.image0.channels << 32) |
-                          ((uint64_t)(args.image0.width / 2) << 16) |
-                          (uint64_t)(args.image0.height * 2);
-    }
-  } else {
-    image_image_pixel = ((uint64_t)args.image0.channels << 32) |
+  uint64_t image_image_pixel = ((uint64_t)args.image0.channels << 32) |
                                ((uint64_t)args.image0.width << 16) |
                                (uint64_t)args.image0.height;
-  }
-  //////////////////////////////////////////////////////////
 
   (*arg).driver.image0_address_phy = image0_address_phy;
   (*arg).driver.image1_address_phy = image1_address_phy;

src/fpga/V1/pe.cpp

@@ -13,11 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "fpga/common/pe.h"
+#include "common/enforce.h"
 #include "common/types.h"
 #include "fpga/V1/filter.h"
 #include "fpga/V1/image.h"
 #include "fpga/common/config.h"
 #include "fpga/common/driver.h"
+#include "fpga/common/fpga_common.h"
 #ifdef COST_TIME_PRINT
 #include <sys/time.h>
 #include <time.h>
@@ -252,7 +254,14 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   reg_writeq(
       ((uint64_t)args.kernel.height) | (((uint64_t)args.kernel.width) << 32),
       REG_CONV_FILTER_PIXEL);
-  reg_writeq(args.driver.output_height | (args.driver.output_width << 32),
+
+  uint64_t output_height_fraction =
+      args.driver.output_height / ROW_PARALLEL_NUM;
+  uint64_t output_height_remainder =
+      args.driver.output_height % ROW_PARALLEL_NUM;
+  reg_writeq(args.driver.output_height | (output_height_fraction << 16) |
+                 (output_height_remainder << 26) |
+                 (args.driver.output_width << 32),
              REG_CONV_RESULT_PIXEL);
   reg_writeq(((uint64_t)args.image.pad_height) |
                  (((uint64_t)args.image.pad_width) << 32),
@@ -296,6 +305,7 @@ int ComputeBasicConv(const struct ConvArgs &args) {
     g_fpgainfo.pe_data->pes[PE_IDX_CONV]->status = ERROR;
     ret = -EIO;
     DLOG << "Conv Wait Irq Timeout!";
+    PADDLE_MOBILE_ENFORCE(0, "Conv Wait Irq Timeout");
   }
   output_scale = reg_readq(REG_SCALE_PARAMETER);
   output_scale = (output_scale << 32) | (output_scale >> 32);
@@ -447,6 +457,7 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
     g_fpgainfo.pe_data->pes[PE_IDX_POOLING]->status = ERROR;
     ret = -EIO;
     DLOG << "Pooling Wait Irq Timeout!";
+    PADDLE_MOBILE_ENFORCE(0, "Pooling Wait Irq Timeout!");
   }
   DLOG << "after reg poll";
 
@@ -529,6 +540,7 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
     g_fpgainfo.pe_data->pes[PE_IDX_EW]->status = ERROR;
     ret = -EIO;
     DLOG << "EW Wait Irq Timeout!";
+    PADDLE_MOBILE_ENFORCE(0, "EW Wait Irq Timeout!");
   }
 
   output_scale = reg_readq(REG_SCALE_PARAMETER);
@@ -561,6 +573,7 @@ int PerformBypass(const struct BypassArgs &args) {
        << "   out_scale_address:" << args.output.scale_address;
 #endif
 #ifdef PADDLE_MOBILE_ZU5
+  uint64_t bypass_interrupt = reg_readq(REG_INTERRUPT);
   uint64_t output_scale = 0;
   uint64_t timer_cnt = 0;
   uint64_t cmd = 0;
@@ -666,12 +679,12 @@ int PerformBypass(const struct BypassArgs &args) {
   reg_writeq(output_address_phy, REG_CONVERT_DST_ADDR);
   reg_writeq(datalen, REG_CONVERT_LENGTH);
   reg_writeq(cmd, REG_CONVERT_CMD);
-
   DLOG << "before reg poll";
   if (0 != fpga_regpoll(REG_INTERRUPT, INTERRUPT_BYPASS, PE_IRQ_TIMEOUT)) {
     g_fpgainfo.pe_data->pes[PE_IDX_BYPASS]->status = ERROR;
     ret = -EIO;
     DLOG << "BYPASS Wait Irq Timeout!";
+    PADDLE_MOBILE_ENFORCE(0, "BYPASS Wait Irq Timeout!");
   }
   DLOG << "after reg poll";
 
@@ -1052,6 +1065,7 @@ int ComputeDWConv(const struct DWconvArgs &args) {
     g_fpgainfo.pe_data->pes[PE_IDX_POOLING]->status = ERROR;
     ret = -EIO;
     DLOG << "Pooling Wait Irq Timeout!";
+    PADDLE_MOBILE_ENFORCE(0, "DWConv Wait Irq Timeout");
   }
   DLOG << "after reg poll";
 

src/fpga/common/driver.cpp

@@ -51,11 +51,7 @@ int open_memdevice() {
   return g_fpgainfo.fd_mem;
 }
 
-void pl_reset() {
-  // DLOG << "PL RESET";
-
-  usleep(100 * 1000);
-}
+void pl_reset() { usleep(100 * 1000); }
 
 void setup_pe(struct pe_data_s *pe_data, struct fpga_pe *pe,
               char const *type_name, int pe_idx) {
@@ -77,7 +73,7 @@ void pl_init() {
 
   pe_data = (struct pe_data_s *)malloc(sizeof(struct pe_data_s));
   if (pe_data == nullptr) {
-    DLOG << "pe_data malloc error!";
+    std::cout << "pe_data malloc error!" << std::endl;
     return;
   }
   memset(pe_data, 0, sizeof(struct pe_data_s));
@@ -165,7 +161,7 @@ uint64_t vaddr_to_paddr_driver(void *address) {
   if (iter != g_fpgainfo.fpga_vaddr2paddr_map.end()) {
     paddr = iter->second;
   } else {
-    DLOG << "Invalid pointer: " << address;
+    std::cout << "Invalid pointer: " << address << std::endl;
   }
 
   return paddr;
@@ -191,7 +187,7 @@ void *fpga_reg_free(void *ptr) {
     g_fpgainfo.fpga_addr2size_map.erase(iter);
     munmap(ptr, size);
   } else {
-    DLOG << "Invalid pointer" << ptr;
+    std::cout << "Invalid pointer" << ptr << std::endl;
   }
 }
 
@@ -205,9 +201,6 @@ void *fpga_malloc_driver(size_t size) {
   int i = 0;
   struct MemoryVM2PHYArgs args;
   struct MemoryCacheArgs args_c;
-
-  // memory_request(g_fpgainfo.memory_info, size, &phy_addr);
-
   ret = mmap64(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED,
                g_fpgainfo.fd_mem, FPGA_MEM_PHY_ADDR);
   PADDLE_MOBILE_ENFORCE(ret != (void *)-1, "Should not be -1");
@@ -233,16 +226,12 @@ void fpga_free_driver(void *ptr) {
     size = iter->second;
     g_fpgainfo.fpga_addr2size_map.erase(iter);
     munmap(ptr, size);
-
-    // p_addr = vaddr_to_paddr_driver(ptr);
-    // pos = (p_addr - g_fpgainfo.memory_info->mem_start) / FPGA_PAGE_SIZE;
-
     auto iter = g_fpgainfo.fpga_vaddr2paddr_map.find(ptr);
     if (iter != g_fpgainfo.fpga_vaddr2paddr_map.end()) {
       g_fpgainfo.fpga_vaddr2paddr_map.erase(iter);
     }
   } else {
-    DLOG << "Invalid pointer" << ptr;
+    std::cout << "Invalid pointer" << ptr << std::endl;
   }
 }
 
@@ -295,10 +284,7 @@ int open_device_driver() {
 
   g_fpgainfo.FpgaRegVirAddr =
       (uint64_t *)fpga_reg_malloc(FPGA_REG_SIZE);  // NOLINT
-  // fpga_memory_add();
-
   pl_init();
-
   return ret;
 }
 
@@ -306,7 +292,6 @@ int close_device_driver() {
   pl_destroy();
   fpga_reg_free(g_fpgainfo.FpgaRegVirAddr);
   memory_release(g_fpgainfo.memory_info);
-
   return 0;
 }
 

src/fpga/common/fpga_common.cpp

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "fpga/common/fpga_common.h"
 #include <algorithm>
 #include <map>
+#include <utility>
 #include "fpga/common/config.h"
 #include "fpga/common/driver.h"
 
@@ -199,8 +200,8 @@ uint64_t vaddr_to_paddr(void *address) {
 }
 
 uint32_t paddle_mobile_version() {
-  uint32_t v_master = 34;
-  uint32_t v_slave = 34;
+  uint32_t v_master = 35;
+  uint32_t v_slave = 35;
 
   uint32_t first = 1, second = 2, fourth_master = 1, fourth_slave = 2;
   uint32_t master = first << 24 | second << 16 | v_master << 8 | fourth_master;

src/operators/kernel/fpga/V1/anchor_generator_kernel.cpp

@@ -14,6 +14,9 @@ limitations under the License. */
 
 #ifdef ANCHOR_GENERATOR_OP
 
+#include <string.h>
+#include <iostream>
+#include <utility>
 #include <vector>
 #include "operators/kernel/detection_kernel.h"
 
@@ -29,11 +32,23 @@ bool AnchorGeneratorKernel<FPGA, float>::Init(
   auto stride = param->stride_;
   auto feature_width = input->dims()[3], feature_height = input->dims()[2];
   auto stride_width = stride[0], stride_height = stride[1];
+  auto offset = param->offset_;
 
   int anchors_offset[] = {-2,  -2,   18,   18,  -10, -9,   26,   25,   -23,
                           -20, 39,   36,   -43, -34, 59,   49,   -63,  -54,
                           79,  69,   -96,  -77, 112, 93,   -137, -118, 153,
                           134, -204, -188, 220, 204, -281, -395, 296,  441};
+
+  int anchors_offset2[] = {0, 0, 51, 77, 0, 0, 30, 35, 0, 0, 81, 103,
+                           0, 0, 20, 21, 0, 0, 36, 44, 0, 0, 43, 58,
+                           0, 0, 34, 68, 0, 0, 24, 28, 0, 0, 19, 46};
+
+  if (offset > 0.6) {
+    memcpy(anchors_offset, anchors_offset2, sizeof(anchors_offset));
+    std::cout << "anchor generator marker" << std::endl;
+  } else {
+    std::cout << "anchor generator rfcn" << std::endl;
+  }
   int num_anchors = sizeof(anchors_offset) / (sizeof(int) * 4);
 
   //  DLOG << "feature_height: " << feature_height;