From bc61454a0c9c3567e30a2302b52f78ad30720135 Mon Sep 17 00:00:00 2001
From: Sergei Shliapnikov <sergei.shliapnikov@amd.com>
Date: Tue, 30 Dec 2025 01:09:54 -0600
Subject: [PATCH] [HIP] Fix double to E8M0 conversion underflow (avoid unsigned
 exp wraparound)

---
 .../include/hip/amd_detail/amd_hip_fp8.h      | 17 ++++++++++---
 .../catch/unit/deviceLib/fp8_e8m0.cc          | 25 ++++++++++++++++---
 2 files changed, 35 insertions(+), 7 deletions(-)

diff --git a/projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp8.h b/projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp8.h
index 519a8aad6149..fb4a97ca994b 100644
--- a/projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp8.h
+++ b/projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp8.h
@@ -987,11 +987,20 @@ __FP8_HOST_DEVICE_STATIC__ __hip_fp8_storage_t __hip_cvt_double_to_e8m0(
   __hip_fp8_storage_t e8m0;
   if (double_exp == 0x0U) {
     e8m0 = 0x0U;
-  } else if ((double_exp - 0x0380U) > 0x00FF) {  // if double is NaN/Inf/or too large
-    e8m0 = hip_detail::e8m0_NaN;
   } else {
-    e8m0 =
-        double_exp - 0x0380U;  // shift due to bias difference between double and single precision
+    // Use signed arithmetic to avoid unsigned underflow for small normal exponents
+    // 0x0380 is a bias difference between double and single precision (1023 - 127)
+    const int float_exp = static_cast<int>(double_exp) - 0x0380;
+    if (float_exp <= 0) {
+      // Underflow to subnormal/zero in float domain
+      e8m0 = 0x0U;
+    } else if (float_exp > 0x00FF) {
+      // NaN/Inf or overflow beyond 8-bit exponent range
+      e8m0 = hip_detail::e8m0_NaN;
+    } else {
+      // Valid finite range [1..254], 0xFF yields NaN and is later handled by saturation
+      e8m0 = static_cast<__hip_fp8_storage_t>(float_exp);
+    }
   }
 
   // If there is a mantissa and the exp wont overflow round up
diff --git a/projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc b/projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc
index e094ed2bb290..373a3b5e44f0 100644
--- a/projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc
+++ b/projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc
@@ -23,6 +23,7 @@ THE SOFTWARE.
 #include <hip_test_common.hh>
 #include <hip/hip_fp8.h>
 #include <cmath>
+#include <limits>
 
 void host_cvt_bfloat16raw_to_e8m0(const std::vector<__hip_bfloat16>& in,
                                   std::vector<unsigned char>& out, __hip_saturation_t sat,
@@ -249,6 +250,22 @@ TEST_CASE("Unit__hip_cvt_double_to_e8m0") {
                             -8.0,
                             1e38,
                             -1e38,
+                            // tiny underflow example -> underflows to 0x00
+                            // rounding-to-+Inf promotes to 0x01
+                            1e-50,
+                            -1e-50,
+                            // smallest double-subnormal (denorm_min) -> underflows to 0x00;
+                            // rounding-to-+Inf does not promote (magnitude below half-step)
+                            std::numeric_limits<double>::denorm_min(),
+                            -std::numeric_limits<double>::denorm_min(),
+                            // smallest double-normal -> underflows to 0x00;
+                            // rounding-to-+Inf does not promote (fraction == 0)
+                            std::numeric_limits<double>::min(),
+                            -std::numeric_limits<double>::min(),
+                            // smallest float-normal (cast to double) -> E8M0 code 0x01;
+                            // rounding-to-+Inf does not promote (fraction == 0)
+                            static_cast<double>(std::numeric_limits<float>::min()),
+                            -static_cast<double>(std::numeric_limits<float>::min()),
                             std::nan("1"),
                             -std::nan("1"),
                             std::numeric_limits<double>::infinity(),
@@ -259,9 +276,11 @@ TEST_CASE("Unit__hip_cvt_double_to_e8m0") {
   std::vector<unsigned char> out(exp.size());
 
   if (rounding == hipRoundPosInf) {
-    exp = {0x00U, 0x7EU, 0x7FU, 0x81U, 0x82U, 0x82U, 0x7EU, 0x7FU, 0x81U, 0x82U, 0x82U, 0xFE, 0xFE};
+    exp = {0x00U, 0x7EU, 0x7FU, 0x81U, 0x82U, 0x82U, 0x7EU, 0x7FU, 0x81U, 0x82U, 0x82U, 0xFE, 0xFE,
+           0x01U, 0x01U, 0x00U, 0x00U, 0x00U, 0x00U, 0x01U, 0x01U};
   } else {
-    exp = {0x00U, 0x7EU, 0x7EU, 0x81U, 0x81U, 0x82U, 0x7EU, 0x7EU, 0x81U, 0x81U, 0x82U, 0xFD, 0xFD};
+    exp = {0x00U, 0x7EU, 0x7EU, 0x81U, 0x81U, 0x82U, 0x7EU, 0x7EU, 0x81U, 0x81U, 0x82U, 0xFD, 0xFD,
+           0x00U, 0x00U, 0x00U, 0x00U, 0x00U, 0x00U, 0x01U, 0x01U};
   }
   if (saturation == __HIP_NOSAT) {
     std::vector<unsigned char> exp_append = {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF};
@@ -337,7 +356,7 @@ TEST_CASE("Unit__hip_cvt_e8m0_to_bf16raw") {
     if (std::isnan(exp[i])) {
       REQUIRE(std::isnan(out[i]));
     } else {
-      REQUIRE_THAT(out[i], Catch::WithinAbs(exp[i], 1e-6f) || Catch::WithinRel(exp[i], 1e-3f));
+      REQUIRE_THAT(out[i], Catch::Matchers::WithinAbs(exp[i], 1e-6f) || Catch::Matchers::WithinRel(exp[i], 1e-3f));
     }
   }
 }