Refactor f8_t, add bf8_t (#792)

* Refactor f8_t to add bf8_t

* Add check_err impl for f8_t

* Update fp8 test

* Format

* Revert the fix

* Update vector_type implementation

* Add bf8 test

* Add bf8, use BitInt types

* Add bf8 conversion methods

* Update type_convert for fp8/bf8

* Add check_err fp8/bf8 support

* Add subnorm fp8 tests

* Add subnorm bf8 tests

* Fix conversion

* Add bf8 cmake bindings

* Add macros to enable build with disabled fp8/bf8

* Remove is_native method

* Update flag combination for mixed precision instances

* Add more flag checks

* Add another flag to a client example

* Add type traits, decouple f8/bf8 casting

* Clean up

* Decouple fp8 and bf8 flags

* Remove more redundant flags

* Remove leftover comments

test/data_type/CMakeLists.txt

@@ -3,5 +3,12 @@ if (USE_BITINT_EXTENSION_INT4)
   target_link_libraries(test_int4 PRIVATE utility)
 endif()
 
-add_gtest_executable(test_fp8 fp8.cpp)
-target_link_libraries(test_fp8 PRIVATE utility)
+if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
+  add_gtest_executable(test_f8 f8.cpp)
+  target_link_libraries(test_f8 PRIVATE utility)
+endif()
+
+if(DTYPES MATCHES "bf8" OR NOT DEFINED DTYPES)
+  add_gtest_executable(test_bf8 bf8.cpp)
+  target_link_libraries(test_bf8 PRIVATE utility)
+endif()

test/data_type/bf8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type_convert.hpp"
+
+using ck::bf8_t;
+using ck::f8_convert_sr;
+using ck::half_t;
+using ck::type_convert;
+
+TEST(BF8, NumericLimits)
+{
+    // constants given for negative zero nan mode
+    EXPECT_EQ(ck::NumericLimits<bf8_t>::Min(), type_convert<bf8_t>(0x04));
+    EXPECT_EQ(ck::NumericLimits<bf8_t>::Max(), type_convert<bf8_t>(0x7F));
+    EXPECT_EQ(ck::NumericLimits<bf8_t>::Lowest(), type_convert<bf8_t>(0xFF));
+    EXPECT_EQ(ck::NumericLimits<bf8_t>::QuietNaN(), type_convert<bf8_t>(0x80));
+}
+
+TEST(BF8, ConvertFP32Nearest)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+    // convert 0 float to bf8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(type_convert<bf8_t>(0.0f)), abs_tol);
+    // convert minimal float to bf8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(type_convert<bf8_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+    // convert maximal bf8_t to float and check if equal to 57344.0
+    ASSERT_NEAR(57344.0f, type_convert<float>(type_convert<bf8_t>(57344.0f)), abs_tol);
+    // convert maximal float to bf8 and back, check if clipped to 57344.0
+    ASSERT_NEAR(57344.0f,
+                type_convert<float>(type_convert<bf8_t>(std::numeric_limits<float>::max())),
+                abs_tol);
+    // convert inf float to bf8_t and check if it is qNan
+    ASSERT_NEAR(type_convert<bf8_t>(0x80),
+                type_convert<bf8_t>(std::numeric_limits<float>::infinity()),
+                abs_tol);
+    // positive norm float value to bf8 and back, check if holds
+    float pos_float = 0.0000762939f;
+    ASSERT_NEAR(pos_float, type_convert<float>(type_convert<bf8_t>(pos_float)), abs_tol);
+    // negative norm float value to bf8 and back, check if holds
+    float neg_float = -0.0000610351f;
+    ASSERT_NEAR(neg_float, type_convert<float>(type_convert<bf8_t>(neg_float)), abs_tol);
+    // positive subnorm float value to bf8 and back, check if holds
+    pos_float = 0.0000305175f;
+    ASSERT_NEAR(pos_float, type_convert<float>(type_convert<bf8_t>(pos_float)), abs_tol);
+    // negative subnorm float value to bf8 and back, check if holds
+    neg_float = -0.0000152587f;
+    ASSERT_NEAR(neg_float, type_convert<float>(type_convert<bf8_t>(neg_float)), abs_tol);
+}
+
+TEST(BF8, ConvertFP32Stochastic)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+    // convert 0 float to bf8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_t>(0.0f)), abs_tol);
+    // convert minimal float to bf8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+    // convert maximal bf8_t to float and check if equal to 57344.0
+    ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_sr<bf8_t>(57344.0f)), abs_tol);
+    // convert maximal float to bf8 and back, check if clipped to 57344.0
+    ASSERT_NEAR(57344.0f,
+                type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::max())),
+                abs_tol);
+    // convert inf float to bf8_t and check if it is qNan
+    ASSERT_NEAR(type_convert<bf8_t>(0x80),
+                f8_convert_sr<bf8_t>(std::numeric_limits<float>::infinity()),
+                abs_tol);
+    // positive norm float value to bf8 and back, check if holds
+    float pos_float = 0.0000762939f;
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
+    // negative norm float value to bf8 and back, check if holds
+    float neg_float = -0.0000610351f;
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
+    // positive subnorm float value to bf8 and back, check if holds
+    pos_float = 0.0000305175f;
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
+    // negative subnorm float value to bf8 and back, check if holds
+    neg_float = -0.0000152587f;
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
+}
+
+TEST(BF8, ConvertFP16Nearest)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-3;
+    // convert 0 fp16 to bf8 and back, check if holds
+    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(type_convert<bf8_t>(half_t{0.0})), abs_tol);
+    // convert minimal fp16 to bf8 and back, check if holds
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(type_convert<bf8_t>(ck::NumericLimits<half_t>::Min())),
+                abs_tol);
+    // convert maximal bf8_t to fp16 and check if equal to 57344.0
+    ASSERT_NEAR(
+        half_t{57344.0}, type_convert<half_t>(type_convert<bf8_t>(half_t{57344.0})), abs_tol);
+    // convert maximal fp16 to bf8 and back, check if clipped to 57344.0
+    ASSERT_NEAR(half_t{57344.0},
+                type_convert<half_t>(type_convert<bf8_t>(ck::NumericLimits<half_t>::Max())),
+                abs_tol);
+    // convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
+    ASSERT_NEAR(type_convert<bf8_t>(0x80),
+                type_convert<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+                abs_tol);
+    // positive norm fp16 value to bf8 and back, check if holds
+    half_t pos_half = half_t{0.0000762939};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<bf8_t>(pos_half)), abs_tol);
+    // negative norm fp16 value to bf8 and back, check if holds
+    half_t neg_half = half_t{-0.0000610351};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<bf8_t>(neg_half)), abs_tol);
+    // positive subnorm fp16 value to bf8 and back, check if holds
+    pos_half = half_t{0.0000305175};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<bf8_t>(pos_half)), abs_tol);
+    // negative subnorm fp16 value to bf8 and back, check if holds
+    neg_half = half_t{-0.0000152587};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<bf8_t>(neg_half)), abs_tol);
+}
+
+TEST(BF8, ConvertFP16Stochastic)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-3;
+    // convert 0 fp16 to bf8 and back, check if holds
+    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{0.0})), abs_tol);
+    // convert minimal fp16 to bf8 and back, check if holds
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Min())),
+                abs_tol);
+    // convert maximal bf8_t to fp16 and check if equal to 57344.0
+    ASSERT_NEAR(
+        half_t{57344.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{57344.0})), abs_tol);
+    // convert maximal fp16 to bf8 and back, check if clipped to 57344.0
+    ASSERT_NEAR(half_t{57344.0},
+                type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Max())),
+                abs_tol);
+    // convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
+    ASSERT_NEAR(type_convert<bf8_t>(0x80),
+                f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+                abs_tol);
+    // positive norm fp16 value to bf8 and back, check if holds
+    half_t pos_half = half_t{0.0000762939};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
+    // negative norm fp16 value to bf8 and back, check if holds
+    half_t neg_half = half_t{-0.0000610351};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
+    // positive subnorm fp16 value to bf8 and back, check if holds
+    pos_half = half_t{0.0000305175};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
+    // negative subnorm fp16 value to bf8 and back, check if holds
+    neg_half = half_t{-0.0000152587};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
+}

test/data_type/fp8.cpp → test/data_type/f8.cpp

@@ -12,10 +12,11 @@ using ck::type_convert;
 
 TEST(FP8, NumericLimits)
 {
-    EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), 0x08);
-    EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), 0x77);
-    EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), 0xF7);
-    EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), 0x80);
+    // constants given for negative zero nan mode
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), type_convert<f8_t>(0x08));
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), type_convert<f8_t>(0x7F));
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), type_convert<f8_t>(0xFF));
+    EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), type_convert<f8_t>(0x80));
 }
 
 TEST(FP8, ConvertFP32Nearest)
@@ -35,12 +36,20 @@ TEST(FP8, ConvertFP32Nearest)
                 type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::max())),
                 abs_tol);
     // convert inf float to f8_t and check if it is qNan
-    ASSERT_NEAR(0x80, type_convert<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
-    // positive float value to fp8 and back, check if holds
-    float pos_float = 0.0078125f;
+    ASSERT_NEAR(type_convert<f8_t>(0x80),
+                type_convert<f8_t>(std::numeric_limits<float>::infinity()),
+                abs_tol);
+    // positive norm float value to fp8 and back, check if holds
+    float pos_float = 0.017578125f;
+    ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
+    // negative norm float value to fp8 and back, check if holds
+    float neg_float = -0.015625f;
+    ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
+    // positive subnorm float value to fp8 and back, check if holds
+    pos_float = 0.00390625f;
     ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
-    // negative float value to fp8 and back, check if holds
-    float neg_float = -0.0156250f;
+    // negative subnorm float value to fp8 and back, check if holds
+    neg_float = -0.001953125f;
     ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
 }
 
@@ -61,12 +70,20 @@ TEST(FP8, ConvertFP32Stochastic)
                 type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::max())),
                 abs_tol);
     // convert inf float to f8_t and check if it is qNan
-    ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
-    // positive float value to fp8 and back, check if holds
-    float pos_float = 0.0078125f;
+    ASSERT_NEAR(type_convert<f8_t>(0x80),
+                f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()),
+                abs_tol);
+    // positive norm float value to fp8 and back, check if holds
+    float pos_float = 0.017578125f;
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
+    // negative norm float value to fp8 and back, check if holds
+    float neg_float = -0.015625f;
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
+    // positive subnorm float value to fp8 and back, check if holds
+    pos_float = 0.00390625f;
     ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
-    // negative float value to fp8 and back, check if holds
-    float neg_float = -0.0156250f;
+    // negative subnorm float value to fp8 and back, check if holds
+    neg_float = -0.001953125f;
     ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
 }
 
@@ -87,12 +104,20 @@ TEST(FP8, ConvertFP16Nearest)
                 type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Max())),
                 abs_tol);
     // convert QuietNaN fp16 to f8_t and check if it is QuietNaN
-    ASSERT_NEAR(0x80, type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
-    // positive fp16 value to fp8 and back, check if holds
-    half_t pos_half = half_t{0.0078125};
+    ASSERT_NEAR(type_convert<f8_t>(0x80),
+                type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+                abs_tol);
+    // positive norm fp16 value to fp8 and back, check if holds
+    half_t pos_half = half_t{0.017578125};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
+    // negative norm fp16 value to fp8 and back, check if holds
+    half_t neg_half = half_t{-0.015625};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
+    // positive subnorm fp16 value to fp8 and back, check if holds
+    pos_half = half_t{0.00390625};
     ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
-    // negative fp16 value to fp8 and back, check if holds
-    half_t neg_half = half_t{-0.0156250};
+    // negative subnorm fp16 value to fp8 and back, check if holds
+    neg_half = half_t{-0.001953125};
     ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
 }
 
@@ -113,11 +138,19 @@ TEST(FP8, ConvertFP16Stochastic)
                 type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Max())),
                 abs_tol);
     // convert QuietNaN fp16 to f8_t and check if it is QuietNaN
-    ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
-    // positive fp16 value to fp8 and back, check if holds
-    half_t pos_half = half_t{0.0078125};
+    ASSERT_NEAR(type_convert<f8_t>(0x80),
+                f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
+                abs_tol);
+    // positive norm fp16 value to fp8 and back, check if holds
+    half_t pos_half = half_t{0.017578125};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
+    // negative norm fp16 value to fp8 and back, check if holds
+    half_t neg_half = half_t{-0.015625};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
+    // positive subnorm fp16 value to fp8 and back, check if holds
+    pos_half = half_t{0.00390625};
     ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
-    // negative fp16 value to fp8 and back, check if holds
-    half_t neg_half = half_t{-0.0156250};
+    // negative subnorm fp16 value to fp8 and back, check if holds
+    neg_half = half_t{-0.001953125};
     ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
 }