!8767 [MD] Use neon instruction set to accelerate the ConvertTo of lite cv

From: @jiangzhiwen8
Reviewed-by: 
Signed-off-by:

mindspore/ccsrc/minddata/dataset/kernels/image/lite_cv/image_process.cc

@@ -21,6 +21,13 @@
 #include <cmath>
 #include <vector>
 
+#ifdef ENABLE_ANDROID
+#if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || defined(_M_ARM64)
+#define USE_NEON
+#include <arm_neon.h>
+#endif
+#endif
+
 namespace mindspore {
 namespace dataset {
 
@@ -423,25 +430,59 @@ bool ConvertTo(const LiteMat &src, LiteMat &dst, double scale) {
   if (src.data_type_ != LDataType::UINT8) {
     return false;
   }
+
   if (scale < 0.0 || scale > 100) {
     return false;
   }
+
   if (dst.IsEmpty()) {
-    (void)dst.Init(src.width_, src.height_, src.channel_, LDataType::FLOAT32);
-  } else if (dst.height_ != src.height_ || dst.width_ != src.width_ || dst.channel_ != src.channel_) {
+    dst.Init(src.width_, src.height_, src.channel_, LDataType::FLOAT32);
+  } else if (src.width_ != dst.width_ || src.height_ != dst.height_ || src.channel_ != dst.channel_) {
     return false;
   } else if (dst.data_type_ != LDataType::FLOAT32) {
     return false;
   }
-  const unsigned char *src_start_p = src;
-  float *dst_start_p = dst;
-  for (int h = 0; h < src.height_; h++) {
-    for (int w = 0; w < src.width_; w++) {
-      uint32_t index = (h * src.width_ + w) * src.channel_;
-      for (int c = 0; c < src.channel_; c++) {
-        dst_start_p[index + c] = (static_cast<float>(src_start_p[index + c] * scale));
-      }
-    }
+
+  const uint8_t *src_ptr = (const uint8_t *)src;
+  float *dst_ptr = reinterpret_cast<float *>(dst.data_ptr_);
+  int64_t total_size = src.height_ * src.width_ * src.channel_;
+  int64_t x = 0;
+#ifdef USE_NEON
+  float32x4_t v_scale = vdupq_n_f32(static_cast<float>(scale));
+  float32x4_t v_c = vdupq_n_f32(0.0f);
+  const int64_t step = 16;
+  for (; x <= total_size - step; x += step) {
+    uint8x16_t v_src = vld1q_u8(src_ptr + x);
+    uint8x16_t v_dst;
+
+    uint16x8_t v_l_16x8 = vmovl_u8(vget_low_u8(v_src));
+    uint16x8_t v_h_16x8 = vmovl_u8(vget_high_u8(v_src));
+
+    float32x4_t v_ll_f32x4 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_l_16x8)));
+    float32x4_t v_lh_f32x4 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_l_16x8)));
+    float32x4_t v_hl_f32x4 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_h_16x8)));
+    float32x4_t v_hh_f32x4 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_h_16x8)));
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+    v_ll_f32x4 = vfmaq_f32(v_c, v_ll_f32x4, v_scale);
+    v_lh_f32x4 = vfmaq_f32(v_c, v_lh_f32x4, v_scale);
+    v_hl_f32x4 = vfmaq_f32(v_c, v_hl_f32x4, v_scale);
+    v_hh_f32x4 = vfmaq_f32(v_c, v_hh_f32x4, v_scale);
+#else
+    v_ll_f32x4 = vmlaq_f32(v_c, v_ll_f32x4, v_scale);
+    v_lh_f32x4 = vmlaq_f32(v_c, v_lh_f32x4, v_scale);
+    v_hl_f32x4 = vmlaq_f32(v_c, v_hl_f32x4, v_scale);
+    v_hh_f32x4 = vmlaq_f32(v_c, v_hh_f32x4, v_scale);
+#endif
+
+    vst1q_f32(dst_ptr + x, v_ll_f32x4);
+    vst1q_f32(dst_ptr + x + 4, v_lh_f32x4);
+    vst1q_f32(dst_ptr + x + 8, v_hl_f32x4);
+    vst1q_f32(dst_ptr + x + 12, v_hh_f32x4);
+  }
+#endif
+  for (; x < total_size; x++) {
+    dst_ptr[x] = static_cast<float>(src_ptr[x] * scale);
   }
   return true;
 }

mindspore/ccsrc/minddata/dataset/kernels/image/lite_cv/lite_mat.cc

@@ -323,6 +323,10 @@ inline void SubtractImpl(const uint32_t *src0, const uint32_t *src1, uint32_t *d
 }
 
 bool Subtract(const LiteMat &src_a, const LiteMat &src_b, LiteMat *dst) {
+  if (dst == NULL) {
+    return false;
+  }
+
   if (src_a.width_ != src_b.width_ || src_a.height_ != src_b.height_ || src_a.channel_ != src_b.channel_) {
     return false;
   }
@@ -479,6 +483,10 @@ inline void DivideImpl(const uint32_t *src0, const uint32_t *src1, uint32_t *dst
 }
 
 bool Divide(const LiteMat &src_a, const LiteMat &src_b, LiteMat *dst) {
+  if (dst == NULL) {
+    return false;
+  }
+
   if (src_a.width_ != src_b.width_ || src_a.height_ != src_b.height_ || src_a.channel_ != src_b.channel_) {
     return false;
   }
@@ -576,6 +584,10 @@ inline void MultiplyImpl(const uint32_t *src0, const uint32_t *src1, uint32_t *d
 }
 
 bool Multiply(const LiteMat &src_a, const LiteMat &src_b, LiteMat *dst) {
+  if (dst == NULL) {
+    return false;
+  }
+
   if (src_a.width_ != src_b.width_ || src_a.height_ != src_b.height_ || src_a.channel_ != src_b.channel_) {
     return false;
   }