optimize int8 max pool

mindspore/lite/nnacl/int8/concat_int8.c

@@ -46,15 +46,11 @@ void Int8Concat(int8_t **inputs, int8_t *output, ConcatParameter *para, int axis
         float bias = -input_quant[i].zp_ * scale;
         for (int j = 0; j < in_copy_size; j++) {
           int32_t output_tmp = round(input_ptr[j] * scale + bias) + output_zp;
-          if (output_tmp > max_int8) {
+          output_tmp = output_tmp > min_int8 ? output_tmp : min_int8;
-            output[j] = max_int8;
+          output_tmp = output_tmp < max_int8 ? output_tmp : max_int8;
-          } else if (output_tmp < min_int8) {
-            output[j] = min_int8;
-          } else {
           output[j] = (int8_t)output_tmp;
         }
       }
-      }
       output += in_copy_size;
     }
   }

mindspore/lite/nnacl/int8/pooling_int8.c

@@ -451,118 +451,83 @@ void MaxPoolingOptInt8(const int8_t *input_ptr, int8_t *output_ptr, PoolingParam
   int output_batch = pooling_param->output_batch_;
   int out_plane = output_w * output_h;
   int out_tile_count = UP_DIV(out_plane, TILE_NUM);
-  int thread_num = out_tile_count < pooling_param->thread_num_ ? out_tile_count : pooling_param->thread_num_;
+  int thread_num = MSMIN(out_tile_count, pooling_param->thread_num_);
-  int c16 = channel / 16;
+  int8_t out_array[MAX_MAXPOOL_SIZE];
 
   for (int batch = 0; batch < output_batch; batch++) {
     int in_batch_offset = batch * in_h * in_w * channel;
     int out_batch_offset = batch * output_h * output_w * channel;
     for (int thread_id = task_id; thread_id < out_tile_count; thread_id += thread_num) {
       int cal_start_index = thread_id * TILE_NUM;
-      int real_cal_num = (out_plane - cal_start_index) > TILE_NUM ? TILE_NUM : (out_plane - cal_start_index);
+      int real_cal_num = out_plane - cal_start_index;
+      real_cal_num = MSMIN(real_cal_num, TILE_NUM);
       for (int i = 0; i < real_cal_num; i++) {
         int index = cal_start_index + i;
         int out_w_index = index % output_w;
         int out_h_index = index / output_w;
         int in_w_index = out_w_index * stride_w - pad_w;
         int in_h_index = out_h_index * stride_h - pad_h;
+        int ky_s = 0 > (-in_h_index) ? 0 : (-in_h_index);
+        int ky_e = MSMIN(win_h, in_h - in_h_index);
+        int kx_s = 0 > (-in_w_index) ? 0 : (-in_w_index);
+        int kx_e = MSMIN(win_w, in_w - in_w_index);
+        int input_stride = (in_h_index * in_w + in_w_index) * channel + in_batch_offset;
         int out_plane_offset = out_batch_offset + index * channel;
-        for (int j = 0; j < c16; j++) {
-          int in_channel_offset = in_batch_offset + j * 16;
-          int out_channel_offset = out_plane_offset + j * 16;
-#ifdef ENABLE_NEON
-          int8x16_t tmp_max = vdupq_n_s8(INT8_MIN);
-#else
-          int8_t tmp_max[16];
-          for (int m = 0; m < C16NUM; ++m) {
-            tmp_max[m] = INT8_MIN;
-          }
-#endif
-          for (int h = 0; h < win_h; h++) {
-            for (int w = 0; w < win_w; w++) {
-              if ((in_h_index + h) < 0 || (in_h_index + h) >= in_h || (in_w_index + w) < 0 ||
-                  (in_w_index + w) >= in_w) {
-                continue;
-              } else {
-                int in_offset = in_channel_offset + ((in_h_index + h) * in_w + in_w_index + w) * channel;
-#ifdef ENABLE_NEON
-                tmp_max = vmaxq_s8(tmp_max, vld1q_s8(input_ptr + in_offset));
-#else
-                for (int k = 0; k < C16NUM; ++k) {
-                  tmp_max[k] = MaxInt8(tmp_max[k], *(input_ptr + in_offset + k));
-                }
-#endif
-              }
-            }  // win_w loop
-          }    // win_h loop
-#ifdef ENABLE_NEON
-          vst1q_s8(output_ptr + out_channel_offset, tmp_max);
-#else
-          for (int l = 0; l < C16NUM; ++l) {
-            *(output_ptr + out_channel_offset + l) = tmp_max[l];
-          }
-#endif
-        }  // in_channel loop
 
-        // 8 channel
+        int c = 0;
-        int tmp_c = c16 * 16;
+        for (; c < channel; c += MAX_MAXPOOL_SIZE) {
-        int c8 = (channel - c16 * 16) / 8;
+          int real_channel = channel - c;
-        for (int k = 0; k < c8; k++) {
+          real_channel = MSMIN(real_channel, MAX_MAXPOOL_SIZE);
-          int in_channel_offset = in_batch_offset + tmp_c + k * 8;
+          memset(out_array, INT8_MIN, real_channel);
-          int out_channel_offset = out_plane_offset + tmp_c + k * 8;
+          int8_t *out_data = output_ptr + out_plane_offset + c;
-#ifdef ENABLE_NEON
+          for (int h = ky_s; h < ky_e; ++h) {
-          int8x8_t tmp_max = vdup_n_s8(INT8_MIN);
+            int in_h_offset = input_stride + h * in_w * channel + c;
-#else
+            for (int w = kx_s; w < kx_e; ++w) {
-          int8_t tmp_max[8];
+              const int8_t *in_data = input_ptr + in_h_offset + w * channel;
-          for (int m = 0; m < C8NUM; ++m) {
+              int j = 0;
-            tmp_max[m] = INT8_MIN;
-          }
-#endif
-          for (int h = 0; h < win_h; h++) {
-            for (int w = 0; w < win_w; w++) {
-              if ((in_h_index + h) < 0 || (in_h_index + h) >= in_h || (in_w_index + w) < 0 ||
-                  (in_w_index + w) >= in_w) {
-                continue;
-              } else {
-                int in_offset = in_channel_offset + ((in_h_index + h) * in_w + in_w_index + w) * channel;
 #ifdef ENABLE_NEON
-                tmp_max = vmax_s8(tmp_max, vld1_s8(input_ptr + in_offset));
+              int c16 = real_channel / 16 * 16;
-#else
+              int c8 = real_channel / 8 * 8;
-                for (int l = 0; l < C8NUM; ++l) {
+              for (; j < c16; j += 16) {
-                  tmp_max[l] = MaxInt8(tmp_max[l], *(input_ptr + in_offset + l));
+                int8x16_t ori_in = vld1q_s8(in_data);
-                }
+                int8x16_t out_array16 = vld1q_s8(out_array + j);
+                in_data += 16;
+                out_array16 = vmaxq_s8(ori_in, out_array16);
+                vst1q_s8(out_array + j, out_array16);
+              }  // 16 channel loop
+
+              for (; j < c8; j += 8) {
+                int8x8_t ori_in = vld1_s8(in_data);
+                int8x8_t out_array8 = vld1_s8(out_array + j);
+                in_data += 8;
+                out_array8 = vmax_s8(ori_in, out_array8);
+                vst1_s8(out_array + j, out_array8);
+              }  // 8 channel loop
 #endif
+              for (; j < real_channel; ++j) {
+                out_array[j] = out_array[j] > in_data[j] ? out_array[j] : in_data[j];
               }
-            }  // win_w loop
+            }  // kw loop
-          }    // win_h loop
+          }    // kh loop
+
+          int j = 0;
 #ifdef ENABLE_NEON
-          vst1_s8(output_ptr + out_channel_offset, tmp_max);
+          int c16 = real_channel / 16 * 16;
-#else
+          int c8 = real_channel / 8 * 8;
-          for (int l = 0; l < C8NUM; ++l) {
+          for (; j < c16; j += 16) {
-            *(output_ptr + out_channel_offset + l) = tmp_max[l];
+            vst1q_s8(out_data, vld1q_s8(out_array + j));
-          }
+            out_data += 16;
-#endif
+          }  // 16 channel loop
-        }  // 8 channel loop
 
-        // res channel
+          for (; j < c8; j += 8) {
-        int channel_s = c16 * 16 + c8 * 8;
+            vst1_s8(out_data, vld1_s8(out_array + j));
-        for (int k = channel_s; k < channel; k++) {
+            out_data += 8;
-          int in_channel_offset = in_batch_offset + k;
+          }  // 8 channel loop
-          int out_channel_offset = out_plane_offset + k;
+#endif
-          int8_t tmp_max = INT8_MIN;
+          for (; j < real_channel; ++j) {
-          for (int h = 0; h < win_h; h++) {
+            out_data[j] = out_array[j];
-            for (int w = 0; w < win_w; w++) {
-              if ((in_h_index + h) < 0 || (in_h_index + h) >= in_h || (in_w_index + w) < 0 ||
-                  (in_w_index + w) >= in_w) {
-                continue;
-              } else {
-                int in_offset = in_channel_offset + ((in_h_index + h) * in_w + in_w_index + w) * channel;
-                tmp_max = MaxInt8(tmp_max, *(input_ptr + in_offset));
           }
-            }  // win_w loop
+        }  // 256 channel loop
-          }    // win_h loop
-          *(output_ptr + out_channel_offset) = tmp_max;
-        }  // channel_res loop
       }    // out_plane loop
     }      // out_batch loop
   }

mindspore/lite/nnacl/int8/pooling_int8.h

@@ -26,6 +26,8 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
+#define MAX_MAXPOOL_SIZE 256
+
 int AvgPoolingInt8(const int8_t *input_ptr, int8_t *output_ptr, PoolingParameter *pooling_param, int task_id);
 
 int AvgPoolingOptInt8(const int8_t *input_ptr, int8_t *output_ptr, PoolingParameter *pooling_param, int task_id);