api: add fp4_e3m0 support

Author: mgouicem

include/oneapi/dnnl/dnnl.hpp

@@ -854,6 +854,8 @@ struct memory : public handle<dnnl_memory_t> {
     enum class data_type {
         /// Undefined data type (used for empty memory descriptors).
         undef = dnnl_data_type_undef,
+        /// 4-bit float data type with 3-bit exponent and 0 bit mantissa.
+        f4_e3m0 = dnnl_f4_e3m0,
         /// [MX-compliant 4-bit float data type](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) with 2-bit exponent and 1 bit mantissa.
         f4_e2m1 = dnnl_f4_e2m1,
         /// [MX-compliant 8-bit compliant scale data type](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) with 8-bit exponent.

include/oneapi/dnnl/dnnl_common_types.h

@@ -106,6 +106,8 @@ typedef enum {
     dnnl_e8m0 = 13,
     /// [MX-compliant 4-bit float data type](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) with 2-bit exponent and 1 bit mantissa.
     dnnl_f4_e2m1 = 14,
+    /// 4-bit float data type with 3-bit exponent and 0 bit mantissa.
+    dnnl_f4_e3m0 = 15,
 
     /// Parameter to allow internal only data_types without undefined behavior.
     /// This parameter is chosen to be valid for so long as sizeof(int) >= 2.

src/common/c_types_map.hpp

@@ -152,6 +152,7 @@ const alg_kind_t eltwise_stochastic_round
 using data_type_t = dnnl_data_type_t;
 namespace data_type {
 const data_type_t undef = dnnl_data_type_undef;
+const data_type_t f4_e3m0 = dnnl_f4_e3m0;
 const data_type_t f4_e2m1 = dnnl_f4_e2m1;
 const data_type_t e8m0 = dnnl_e8m0;
 const data_type_t f8_e5m2 = dnnl_f8_e5m2;

src/common/dnnl_debug_autogenerated.cpp

@@ -59,6 +59,7 @@ const char *dnnl_dt2str(dnnl_data_type_t v) {
     if (v == dnnl_u4) return "u4";
     if (v == dnnl_e8m0) return "e8m0";
     if (v == dnnl_f4_e2m1) return "f4_e2m1";
+    if (v == dnnl_f4_e3m0) return "f4_e3m0";
     if (v == dnnl_data_type_max) return "data_type_max";
     assert(!"unknown dt");
     return "unknown dt";

src/common/dnnl_traits.hpp

@@ -38,6 +38,10 @@ struct typesize_traits {}; /* ::data_type_size -> f32 */
 template <primitive_kind_t>
 struct pkind_traits {}; /* ::desc_type, ::query_d */
 
+template <>
+struct prec_traits<data_type::f4_e3m0> {
+    typedef float4_e3m0_t type;
+};
 template <>
 struct prec_traits<data_type::f4_e2m1> {
     typedef float4_e2m1_t type;
@@ -95,6 +99,10 @@ struct prec_traits<data_type::boolean> {
     typedef bool type;
 };
 
+template <>
+struct data_traits<float4_e3m0_t> {
+    static constexpr data_type_t data_type = data_type::f4_e3m0;
+};
 template <>
 struct data_traits<float4_e2m1_t> {
     static constexpr data_type_t data_type = data_type::f4_e2m1;

src/common/float4.cpp

@@ -32,7 +32,7 @@ uint8_t float2e2m1(float f) {
     // There is no NaN or infinity in e2m1, for now we just return zero
     // TODO: figure if there is a standard value to return
     uint32_t naninf_mask = 0x7f800000;
-    if ((f_raw & naninf_mask) == naninf_mask) return 0x00000000;
+    if ((f_raw & naninf_mask) == naninf_mask) return 0x00;
 
     // we convert with naive closest value computation out of 8
     float e2m1_val_table[8] = {0.0f, .5f, 1.0f, 1.5f, 2.0f, 3.0f, 4.0f, 6.0f};
@@ -91,5 +91,72 @@ float4_e2m1_t::operator float16_t() const {
     return e2m1_table[raw_bits_];
 }
 
+uint8_t float2e3m0(float f) {
+    uint32_t f_raw = float2int(f);
+    uint32_t sign = f_raw & 0x80000000;
+
+    // There is no NaN or infinity in e3m0, we just return maxval
+    uint32_t naninf_mask = 0x7f800000;
+    if ((f_raw & naninf_mask) == naninf_mask) return 0x7;
+
+    // we convert with naive closest value computation out of 8
+    float e3m0_val_table[8] = {0.0f, .25f, .5f, 1.0f, 2.0f, 4.0f, 8.0f, 16.0f};
+
+    float abs_f = int2float(f_raw ^ sign);
+
+    int idx = 0;
+    float min_diff = ::fabsf(e3m0_val_table[idx] - abs_f);
+    uint8_t raw_bits = idx;
+    for (++idx; idx < 8; ++idx) {
+        float diff = ::fabsf(e3m0_val_table[idx] - abs_f);
+        if (diff < min_diff) {
+            min_diff = diff;
+            raw_bits = idx;
+        }
+        // Special case for midpoint, we round to even (so even index)
+        if ((diff == min_diff) && !(idx & 1)) raw_bits = idx;
+    }
+    assert(raw_bits < 8);
+    // reapply sign
+    if (sign) raw_bits = raw_bits | 0x08;
+    assert(raw_bits < 16);
+    return raw_bits;
+}
+
+float4_e3m0_t &float4_e3m0_t::operator=(bfloat16_t f) {
+    float f32 = f;
+    raw_bits_ = float2e3m0(f32);
+    return *this;
+}
+
+float4_e3m0_t &float4_e3m0_t::operator=(float16_t f) {
+    float f32 = f;
+    raw_bits_ = float2e3m0(f32);
+    return *this;
+}
+
+float4_e3m0_t &float4_e3m0_t::operator=(float f) {
+    raw_bits_ = float2e3m0(f);
+    return *this;
+}
+
+float4_e3m0_t::operator float() const {
+    // List of e3m0 values. The index of each value maps to its encoding.
+    static const float e3m0_table[16]
+            = {0.0f, .25f, .5f, 1.0f, 2.0f, 4.0f, 8.0f, 16.0f, -0.0f, -.25f,
+                    -.5f, -1.0f, -2.0f, -4.0f, -8.0f, -16.0f};
+    assert(raw_bits_ < 16);
+    return e3m0_table[raw_bits_];
+}
+
+float4_e3m0_t::operator float16_t() const {
+    // List of e3m0 values. The index of each value maps to its encoding.
+    static const float16_t e3m0_table[16]
+            = {0.0f, .25f, .5f, 1.0f, 2.0f, 4.0f, 8.0f, 16.0f, -0.0f, -.25f,
+                    -.5f, -1.0f, -2.0f, -4.0f, -8.0f, -16.0f};
+    assert(raw_bits_ < 16);
+    return e3m0_table[raw_bits_];
+}
+
 } // namespace impl
 } // namespace dnnl

src/common/float4.hpp

@@ -49,6 +49,29 @@ struct float4_e2m1_t {
 };
 static_assert(sizeof(float4_e2m1_t) == 1, "float4_e2m1_t must be 1 byte");
 
+struct float4_e3m0_t {
+    uint8_t raw_bits_;
+    float4_e3m0_t() = default;
+    constexpr float4_e3m0_t(uint8_t r, bool = true) : raw_bits_(r) {}
+    float4_e3m0_t(float f) { (*this) = f; }
+    float4_e3m0_t(float16_t f) { (*this) = f; }
+    float4_e3m0_t(bfloat16_t f) { (*this) = f; }
+
+    float4_e3m0_t DNNL_API &operator=(float f);
+    float4_e3m0_t DNNL_API &operator=(float16_t f);
+    float4_e3m0_t DNNL_API &operator=(bfloat16_t f);
+
+    DNNL_API operator float() const;
+    DNNL_API operator float16_t() const;
+    DNNL_API operator bfloat16_t() const;
+
+    float4_e3m0_t &operator+=(const float a) {
+        (*this) = float {*this} + a;
+        return *this;
+    }
+};
+static_assert(sizeof(float4_e3m0_t) == 1, "float4_e3m0_t must be 1 byte");
+
 } // namespace impl
 } // namespace dnnl
 

src/common/nstl.hpp

@@ -157,6 +157,22 @@ struct numeric_limits<int8_t> : public std::numeric_limits<int8_t> {};
 template <>
 struct numeric_limits<uint8_t> : public std::numeric_limits<uint8_t> {};
 
+template <>
+struct numeric_limits<float4_e3m0_t> {
+    static constexpr float4_e3m0_t lowest() { return float4_e3m0_t(0xf, true); }
+    // Min normal is equal to the value 1.0
+    static constexpr float4_e3m0_t min() { return float4_e3m0_t(0x1, true); }
+    // Max normal is equal to the value 6.0
+    static constexpr float4_e3m0_t max() { return float4_e3m0_t(0x7, true); }
+
+    static constexpr int bias = 0x3;
+    static constexpr int digits = 1; // 1 implicit bit
+
+    static constexpr float4_e3m0_t epsilon() {
+        return float4_e3m0_t(0x3, true);
+    }
+};
+
 template <>
 struct numeric_limits<float4_e2m1_t> {
     static constexpr float4_e2m1_t lowest() { return float4_e2m1_t(0xf, true); }

src/common/type_helpers.hpp

@@ -93,6 +93,7 @@ namespace types {
 inline size_t data_type_size(data_type_t data_type) {
     using namespace data_type;
     switch ((int)data_type) {
+        case f4_e3m0: return sizeof(prec_traits<f4_e3m0>::type);
         case f4_e2m1: return sizeof(prec_traits<f4_e2m1>::type);
         case e8m0: return sizeof(prec_traits<e8m0>::type);
         case f8_e5m2: return sizeof(prec_traits<f8_e5m2>::type);
@@ -139,6 +140,7 @@ inline T min_value(data_type_t data_type) {
     case x: \
         return static_cast<T>(nstl::numeric_limits<prec_traits<x>::type>::min())
     switch (data_type) {
+        CASE(f4_e3m0);
         CASE(f4_e2m1);
         CASE(e8m0);
         CASE(f8_e5m2);
@@ -166,6 +168,7 @@ inline T max_value(data_type_t data_type) {
     case x: \
         return static_cast<T>(nstl::numeric_limits<prec_traits<x>::type>::max())
     switch (data_type) {
+        CASE(f4_e3m0);
         CASE(f4_e2m1);
         CASE(e8m0);
         CASE(f8_e5m2);
@@ -195,6 +198,7 @@ inline float max_value(data_type_t data_type) {
         return static_cast<float>( \
                 nstl::numeric_limits<prec_traits<x>::type>::max())
     switch (data_type) {
+        CASE(f4_e3m0);
         CASE(f4_e2m1);
         CASE(e8m0);
         CASE(f8_e5m2);
@@ -233,6 +237,7 @@ inline T lowest_value(data_type_t data_type) {
         return static_cast<T>( \
                 nstl::numeric_limits<prec_traits<x>::type>::lowest())
     switch (data_type) {
+        CASE(f4_e3m0);
         CASE(f4_e2m1);
         CASE(e8m0);
         CASE(f8_e5m2);
@@ -261,6 +266,7 @@ inline T digits(data_type_t data_type) {
         return static_cast<T>( \
                 nstl::numeric_limits<prec_traits<x>::type>::digits)
     switch (data_type) {
+        CASE(f4_e3m0);
         CASE(f4_e2m1);
         CASE(e8m0);
         CASE(f8_e5m2);
@@ -419,6 +425,7 @@ inline data_type_t default_accum_data_type(
     // true
     if (one_of(src_dt, s8, u8, u4, s4) && (dst_dt != f32 || strict)) return s32;
 
+    if (one_of(f4_e3m0, src_dt, dst_dt)) return f32;
     if (one_of(f4_e2m1, src_dt, dst_dt)) return f32;
     if (one_of(f8_e5m2, src_dt, dst_dt)) return f32;
     if (one_of(f8_e4m3, src_dt, dst_dt)) return f32;
@@ -461,6 +468,7 @@ inline data_type_t default_accum_data_type(data_type_t src_dt,
             return f32;
     }
 
+    if (one_of(f4_e3m0, src_dt, wei_dt, dst_dt)) return f32;
     if (one_of(f4_e2m1, src_dt, wei_dt, dst_dt)) return f32;
     if (one_of(f8_e5m2, src_dt, wei_dt, dst_dt)) return f32;
     if (one_of(f8_e4m3, src_dt, wei_dt, dst_dt)) return f32;
@@ -1256,8 +1264,8 @@ inline bool memory_desc_sanity_check(int ndims, const dims_t dims,
     if (ndims == 0) return true;
 
     bool ok = dims != nullptr && 0 < ndims && ndims <= DNNL_MAX_NDIMS
-            && utils::one_of(data_type, f4_e2m1, e8m0, f8_e5m2, f8_e4m3, f16,
-                    bf16, f32, f64, s32, s8, u8, s4, u4);
+            && utils::one_of(data_type, f4_e3m0, f4_e2m1, e8m0, f8_e5m2,
+                    f8_e4m3, f16, bf16, f32, f64, s32, s8, u8, s4, u4);
     if (!ok) return false;
 
     bool has_runtime_dims = false;

tests/benchdnn/dnnl_debug_autogenerated.cpp

@@ -50,6 +50,7 @@ dnnl_data_type_t str2dt(const char *str) {
     CASE(u4);
     CASE(e8m0);
     CASE(f4_e2m1);
+    CASE(f4_e3m0);
     CASE(data_type_max);
 #undef CASE
     if (!strcmp("undef", str) || !strcmp("dnnl_data_type_undef", str))