FP16 unittest tests expect less error values near zero.
Reduce minimum allowed error from 1e-3 to 1e-4
PiperOrigin-RevId: 332284507
diff --git a/test/binary-elementwise-operator-tester.h b/test/binary-elementwise-operator-tester.h
index 2b2867b..7764fe6 100644
--- a/test/binary-elementwise-operator-tester.h
+++ b/test/binary-elementwise-operator-tester.h
@@ -482,7 +482,7 @@
for (size_t n = 0; n < output_dims[5]; n++) {
const size_t index =
i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5];
- ASSERT_NEAR(fp16_ieee_to_fp32_value(output[index]), output_ref[index], std::max(1.0e-3f, std::abs(output_ref[index]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(output[index]), output_ref[index], std::max(1.0e-4f, std::abs(output_ref[index]) * 1.0e-2f))
<< "(i, j, k, l, m, n) = (" << i << ", " << j << ", " << k << ", " << l << ", " << m << ", " << n << ")";
}
}
diff --git a/test/convolution-operator-tester.h b/test/convolution-operator-tester.h
index f0df7ff..8d66225 100644
--- a/test/convolution-operator-tester.h
+++ b/test/convolution-operator-tester.h
@@ -1151,7 +1151,7 @@
// << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
// ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max)
// << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
- ASSERT_NEAR(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-3f, std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
+ ASSERT_NEAR(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-4f, std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
}
}
@@ -1936,7 +1936,7 @@
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max)
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
- ASSERT_NEAR(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-3f, std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
+ ASSERT_NEAR(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-4f, std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
}
}
@@ -2015,7 +2015,7 @@
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max)
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
- ASSERT_NEAR(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-3f, std::abs(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
+ ASSERT_NEAR(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), std::max(1.0e-4f, std::abs(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c]) * 1.0e-2f))
<< "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c;
}
}
diff --git a/test/dwconv-microkernel-tester.h b/test/dwconv-microkernel-tester.h
index 94ca1ef..a38d3e5 100644
--- a/test/dwconv-microkernel-tester.h
+++ b/test/dwconv-microkernel-tester.h
@@ -477,7 +477,7 @@
<< "x = " << x << ", channel = " << c;
ASSERT_LE(fp16_ieee_to_fp32_value(output[x * output_stride() + c]), output_max)
<< "x = " << x << ", channel = " << c;
- ASSERT_NEAR(output_ref[x * channels() + c], fp16_ieee_to_fp32_value(output[x * output_stride() + c]), std::max(1.0e-3f, std::abs(output_ref[x * channels() + c]) * 1.0e-2f))
+ ASSERT_NEAR(output_ref[x * channels() + c], fp16_ieee_to_fp32_value(output[x * output_stride() + c]), std::max(1.0e-4f, std::abs(output_ref[x * channels() + c]) * 1.0e-2f))
<< "x = " << x << ", channel = " << c;
}
}
diff --git a/test/gavgpool-microkernel-tester.h b/test/gavgpool-microkernel-tester.h
index de041bf..0264ef6 100644
--- a/test/gavgpool-microkernel-tester.h
+++ b/test/gavgpool-microkernel-tester.h
@@ -507,7 +507,7 @@
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
ASSERT_GE(fp16_ieee_to_fp32_value(output[c]), output_min)
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
- ASSERT_NEAR(fp16_ieee_to_fp32_value(output[c]), output_ref[c], std::max(1.0e-3f, std::abs(output_ref[c]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(output[c]), output_ref[c], std::max(1.0e-4f, std::abs(output_ref[c]) * 1.0e-2f))
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
}
}
diff --git a/test/gemm-microkernel-tester.h b/test/gemm-microkernel-tester.h
index 7b9ec53..20ec6d4 100644
--- a/test/gemm-microkernel-tester.h
+++ b/test/gemm-microkernel-tester.h
@@ -811,7 +811,7 @@
// Validate micro-kernel outputs.
for (size_t i = 0; i < m(); i++) {
for (size_t j = 0; j < n(); j++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]), c_ref[i * n() + j], std::max(1.0e-3f, std::abs(c_ref[i * n() + j]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]), c_ref[i * n() + j], std::max(1.0e-4f, std::abs(c_ref[i * n() + j]) * 1.0e-2f))
<< "at " << i << ", " << j << ": reference = " << c_ref[i * n() + j]
<< ", optimized = " << fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]) << ", Mr x Nr x Kr = " << mr() << " x " << nr()
<< " x " << kr() << ", M x N x K = " << m() << " x " << n() << " x " << k();
@@ -930,7 +930,7 @@
<< "at " << i << ", " << i << ": reference = " << c_ref[i * n() + j]
<< ", optimized = " << fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]) << ", Mr x Nr x Kr = " << mr() << " x " << nr()
<< " x " << kr() << ", M x N x KC x KS = " << m() << " x " << n() << " x " << k() << " x " << ks();
- ASSERT_NEAR(fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]), c_ref[i * n() + j], std::max(1.0e-3f, std::abs(c_ref[i * n() + j]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]), c_ref[i * n() + j], std::max(1.0e-4f, std::abs(c_ref[i * n() + j]) * 1.0e-2f))
<< "at " << i << ", " << i << ": reference = " << c_ref[i * n() + j]
<< ", optimized = " << fp16_ieee_to_fp32_value(c[i * cm_stride() + (j / nr()) * cn_stride() + j % nr()]) << ", Mr x Nr x Kr = " << mr() << " x " << nr()
<< " x " << kr() << ", M x N x KC x KS = " << m() << " x " << n() << " x " << k() << " x " << ks();
diff --git a/test/global-average-pooling-operator-tester.h b/test/global-average-pooling-operator-tester.h
index dff7d61..efb620d 100644
--- a/test/global-average-pooling-operator-tester.h
+++ b/test/global-average-pooling-operator-tester.h
@@ -360,7 +360,7 @@
for (size_t c = 0; c < channels(); c++) {
ASSERT_LE(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_max);
ASSERT_GE(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_min);
- ASSERT_NEAR(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_ref[i * channels() + c], std::max(1.0e-3f, std::abs(output_ref[i * channels() + c]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_ref[i * channels() + c], std::max(1.0e-4f, std::abs(output_ref[i * channels() + c]) * 1.0e-2f))
<< "at batch index " << i << " / " << batch_size()
<< ", channel " << c << " / " << channels();
}
diff --git a/test/spmm-microkernel-tester.h b/test/spmm-microkernel-tester.h
index e37bf4d..20885e4 100644
--- a/test/spmm-microkernel-tester.h
+++ b/test/spmm-microkernel-tester.h
@@ -443,7 +443,7 @@
// Validate micro-kernel outputs.
for (size_t pxb = 0; pxb < n(); pxb++) {
for (size_t oc = 0; oc < m(); oc++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(c[pxb * m() + oc]), c_ref[pxb * m() + oc], std::max(1.0e-3f, std::abs(c_ref[pxb * m() + oc]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(c[pxb * m() + oc]), c_ref[pxb * m() + oc], std::max(1.0e-4f, std::abs(c_ref[pxb * m() + oc]) * 1.0e-2f))
<< "at " << pxb << ", " << oc
<< ": Mr x Nr x Kr = " << mr() << " x " << nr()
<< ", M x N x K = " << m() << " x " << n() << " x " << k();
diff --git a/test/vbinary-microkernel-tester.h b/test/vbinary-microkernel-tester.h
index ca992e4..220c089 100644
--- a/test/vbinary-microkernel-tester.h
+++ b/test/vbinary-microkernel-tester.h
@@ -150,7 +150,7 @@
// Verify results.
for (size_t i = 0; i < batch_size(); i++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f))
<< "at " << i << " / " << batch_size();
}
}
@@ -230,7 +230,7 @@
// Verify results.
for (size_t i = 0; i < batch_size(); i++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f))
<< "at " << i << " / " << batch_size();
}
}
diff --git a/test/vbinaryc-microkernel-tester.h b/test/vbinaryc-microkernel-tester.h
index 451d5e3..c51ecc5 100644
--- a/test/vbinaryc-microkernel-tester.h
+++ b/test/vbinaryc-microkernel-tester.h
@@ -146,7 +146,7 @@
// Verify results.
for (size_t i = 0; i < batch_size(); i++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f))
<< "at " << i << " / " << batch_size();
}
}
@@ -229,7 +229,7 @@
// Verify results.
for (size_t i = 0; i < batch_size(); i++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f))
<< "at " << i << " / " << batch_size();
}
}
diff --git a/test/vmulcaddc-microkernel-tester.h b/test/vmulcaddc-microkernel-tester.h
index 566da63..5abed00 100644
--- a/test/vmulcaddc-microkernel-tester.h
+++ b/test/vmulcaddc-microkernel-tester.h
@@ -181,7 +181,7 @@
// Verify results.
for (size_t i = 0; i < rows(); i++) {
for (size_t j = 0; j < channels(); j++) {
- ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i * output_stride() + j]), y_ref[i * channels() + j], std::max(1.0e-3f, std::abs(y_ref[i * channels() + j]) * 1.0e-2f))
+ ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i * output_stride() + j]), y_ref[i * channels() + j], std::max(1.0e-4f, std::abs(y_ref[i * channels() + j]) * 1.0e-2f))
<< "at pixel " << i << " / " << rows()
<< ", channel = " << j << " / " << channels();
}
diff --git a/test/vunary-microkernel-tester.h b/test/vunary-microkernel-tester.h
index 100aef8..4d47c16 100644
--- a/test/vunary-microkernel-tester.h
+++ b/test/vunary-microkernel-tester.h
@@ -237,7 +237,7 @@
// Verify results.
for (size_t i = 0; i < batch_size(); i++) {
- ASSERT_NEAR(y[i], y_ref[i], std::max(1.0e-5 * std::abs(y_ref[i]), 5.0e-6))
+ ASSERT_NEAR(y[i], y_ref[i], std::max(5.0e-6, std::abs(y_ref[i]) * 1.0e-5))
<< "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i];
}
}