[ hdot ] Use precision-enhanced hdot

- Previous hdot was using full-fp16.
- Since this is also one of dimension-shrinking computation, should use inter-fp32 values to enhance precision.
- This has not been detected due to small dimension Tensor usage in unittest. Add higher dimension test case accordingly.

**Self evaluation:**
1. Build test:     [X]Passed [ ]Failed [ ]Skipped
2. Run test:     [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: skykongkong8 <ss.kong@samsung.com>

diff --git a/nntrainer/tensor/blas_neon.cpp b/nntrainer/tensor/blas_neon.cpp
index 3609b6b8b53513d6292e4213e068ecc2b6eec092..6f02978e1f4a604a624555eb816a61d39d28f933 100644 (file)
--- a/nntrainer/tensor/blas_neon.cpp
+++ b/nntrainer/tensor/blas_neon.cpp
@@ -1192,51 +1192,29 @@ void haxpy(const unsigned int N, const float alpha, const __fp16 *X,
 }
 
 __fp16 hdot(const unsigned int N, const __fp16 *X, const __fp16 *Y) {
-
-  float16x8_t accX8 = vmovq_n_f16(0);
-  float16x4_t accX4 = vmov_n_f16(0);
+  float32x4_t accX0_3 = vmovq_n_f32(0.F);
+  float32x4_t accX4_7 = vmovq_n_f32(0.F);
 
   unsigned int idx = 0;
-  __fp16 ret = 0;
+  unsigned int N8 = (N >> 3) << 3;
+  float ret = 0;
 
-  // processing batch of 8
-  for (; (N - idx) >= 8; idx += 8) {
+  // Adaptive loop for batch size of 8
+  for (; idx < N8; idx += 8) {
     float16x8_t x = vld1q_f16(&X[idx]);
     float16x8_t y = vld1q_f16(&Y[idx]);
 
-    // x*y + accX8 -> accX8
-    accX8 = vfmaq_f16(accX8, x, y);
-  }
-
-  // check at least one batch of 8 is processed
-  if (N - 8 >= 0) {
-    __fp16 result[8];
-    vst1q_f16(result, accX8);
-    for (unsigned int i = 0; i < 8; i++)
-      ret += result[i];
-  }
-
-  // processing remaining batch of 4
-  for (; (N - idx) >= 4; idx += 4) {
-    float16x4_t x = vld1_f16(&X[idx]);
-    float16x4_t y = vld1_f16(&Y[idx]);
-
-    // x*y + accX4 -> accX4
-    accX4 = vfma_f16(accX4, x, y);
-  }
-
-  // check at least one batch of 4 is processed
-  if (N % 8 >= 4) {
-    __fp16 result[4];
-    vst1_f16(result, accX4);
-    ret += result[0] + result[1] + result[2] + result[3];
+    x = vmulq_f16(x, y);
+    accX0_3 = vaddq_f32(accX0_3, vcvt_f32_f16(vget_low_f16(x)));
+    accX4_7 = vaddq_f32(accX4_7, vcvt_f32_f16(vget_high_f16(x)));
   }
+  ret += vaddvq_f32(accX0_3) + vaddvq_f32(accX4_7);
 
-  // pocessing remaining values
+  // Loop for remaining indices
   for (; idx < N; idx++)
     ret += X[idx] * Y[idx];
 
-  return ret;
+  return static_cast<__fp16>(ret);
 }
 
 __fp16 hnrm2(const unsigned int N, const __fp16 *X) {

diff --git a/test/unittest/unittest_nntrainer_tensor_neon_fp16.cpp b/test/unittest/unittest_nntrainer_tensor_neon_fp16.cpp
index e02eac1786d17f048aea13ed08e4125fc0c47b28..7b3cdae9fa232c197fa6651e4d319075f7ba2bce 100644 (file)
--- a/test/unittest/unittest_nntrainer_tensor_neon_fp16.cpp
+++ b/test/unittest/unittest_nntrainer_tensor_neon_fp16.cpp
@@ -120,6 +120,70 @@ TEST(nntrainer_Tensor, dot) {
   EXPECT_IN_RANGE((float)cosSimNeon, 0.99, 1);
 }
 
+TEST(nntrainer_Tensor, hdot_768) {
+
+  nntrainer::TensorDim::TensorType t_type_nchw_fp16 = {
+    nntrainer::Tformat::NCHW, nntrainer::Tdatatype::FP16};
+
+  nntrainer::TensorDim::TensorType t_type_nchw_fp32 = {
+    nntrainer::Tformat::NCHW, nntrainer::Tdatatype::FP32};
+
+  // conditions for fp16 hdot call:
+  // this->(batch * channel * height) = arg->(width) = 1;
+  size_t batch = 1;
+  size_t channel = 1;
+  size_t height = 1;
+  size_t width = 768;
+
+  nntrainer::Tensor input(
+    nntrainer::TensorDim(1, 1, 1, width, t_type_nchw_fp16));
+
+  nntrainer::Tensor input_2(
+    nntrainer::TensorDim(1, 1, width, 1, t_type_nchw_fp16));
+
+  nntrainer::Tensor input_fp32(
+    nntrainer::TensorDim(1, 1, 1, width, t_type_nchw_fp32));
+
+  nntrainer::Tensor input_fp32_2(
+    nntrainer::TensorDim(1, 1, width, 1, t_type_nchw_fp32));
+
+  const float alpha = 1e-1;
+  const int MOD = 10;
+
+  GEN_TEST_INPUT(input, ((i * j * (batch * height * channel) +
+                          j * (batch * height) + k * (width) + l + 1) %
+                         MOD) *
+                          alpha);
+  GEN_TEST_INPUT(input_fp32, ((i * j * (batch * height * channel) +
+                               j * (batch * height) + k * (width) + l + 1) %
+                              MOD) *
+                               alpha);
+  GEN_TEST_INPUT(input_2, ((i * k * (batch * height * channel) +
+                            j * (batch * height) + k * (width) + l + 1) %
+                           MOD) *
+                            alpha);
+  GEN_TEST_INPUT(input_fp32_2, ((i * k * (batch * height * channel) +
+                                 j * (batch * height) + k * (width) + l + 1) %
+                                MOD) *
+                                 alpha);
+
+  nntrainer::Tensor result_neon = input.dot(input_2, false, false);
+  nntrainer::Tensor result_fp32 = input_fp32.dot(input_fp32_2, false, false);
+
+  float mseErrorNeon =
+    mse<__fp16>(result_neon.getData<__fp16>(), result_fp32.getData<float>(),
+                result_neon.size());
+
+  double cosSimNeon =
+    cosine_similarity<__fp16>(result_neon.getData<__fp16>(),
+                              result_fp32.getData<float>(), result_neon.size());
+
+  const float epsilon = 1e-3;
+
+  EXPECT_IN_RANGE(mseErrorNeon, 0, epsilon);
+  EXPECT_IN_RANGE((float)cosSimNeon, 0.99, 1);
+}
+
 TEST(nntrainer_Tensor, l2norm) {
 
   nntrainer::TensorDim::TensorType t_type_nchw_fp16 = {