[Refactor] Remove dead kernel code (#57621)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-09-20 10:39:23 -07:00
committed by GitHub
parent 49ee12d742
commit 4868312128
12 changed files with 4 additions and 559 deletions
+1 -241
View File
@@ -67,104 +67,6 @@ using fp8_type = __hip_fp8_e4m3_fnuz;
using fp8x2_type = __hip_fp8x2_e4m3_fnuz;
#endif
// fp8 -> half
template <>
__inline__ __device__ uint16_t
vec_conversion<uint16_t, uint8_t>(const uint8_t& a) {
return __hip_cvt_fp8_to_halfraw(a, fp8_type::__default_interpret).x;
}
// fp8x2 -> half2
template <>
__inline__ __device__ uint32_t
vec_conversion<uint32_t, uint16_t>(const uint16_t& a) {
union {
__half2_raw h2r;
uint32_t ui32;
} tmp;
tmp.h2r = __hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
return tmp.ui32;
}
// fp8x4 -> half2x2
template <>
__inline__ __device__ uint2 vec_conversion<uint2, uint32_t>(const uint32_t& a) {
union {
uint2 u32x2;
uint32_t u32[2];
} tmp;
tmp.u32[0] = vec_conversion<uint32_t, uint16_t>((uint16_t)a);
tmp.u32[1] = vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U));
return tmp.u32x2;
}
// fp8x8 -> half2x4
template <>
__inline__ __device__ uint4 vec_conversion<uint4, uint2>(const uint2& a) {
union {
uint4 u64x2;
uint2 u64[2];
} tmp;
tmp.u64[0] = vec_conversion<uint2, uint32_t>(a.x);
tmp.u64[1] = vec_conversion<uint2, uint32_t>(a.y);
return tmp.u64x2;
}
using __nv_bfloat16 = __hip_bfloat16;
// fp8 -> __nv_bfloat16
template <>
__inline__ __device__ __nv_bfloat16
vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a) {
fp8_type f8;
f8.__x = a;
return __float2bfloat16(static_cast<float>(f8));
}
using __nv_bfloat162 = __hip_bfloat162;
// fp8x2 -> __nv_bfloat162
template <>
__inline__ __device__ __nv_bfloat162
vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a) {
__nv_bfloat162 res;
res.x = vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)a);
res.y = vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)(a >> 8U));
return res;
}
// fp8x4 -> bf16_4_t
template <>
__inline__ __device__ bf16_4_t
vec_conversion<bf16_4_t, uint32_t>(const uint32_t& a) {
bf16_4_t res;
res.x = vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)a);
res.y = vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)(a >> 16U));
return res;
}
// fp8x8 -> bf16_8_t
template <>
__inline__ __device__ bf16_8_t vec_conversion<bf16_8_t, uint2>(const uint2& a) {
bf16_4_t tmp1, tmp2;
tmp1 = vec_conversion<bf16_4_t, uint32_t>(a.x);
tmp2 = vec_conversion<bf16_4_t, uint32_t>(a.y);
bf16_8_t res;
res.x = tmp1.x;
res.y = tmp1.y;
res.z = tmp2.x;
res.w = tmp2.y;
return res;
}
// fp8 -> float
template <>
__inline__ __device__ float vec_conversion<float, uint8_t>(const uint8_t& a) {
fp8_type f8;
f8.__x = a;
return static_cast<float>(f8);
}
// fp8x2 -> float2
template <>
__inline__ __device__ float2
@@ -193,138 +95,6 @@ vec_conversion<float4, uint32_t>(const uint32_t& a) {
return res;
}
// fp8x8 -> float8
template <>
__inline__ __device__ Float8_ vec_conversion<Float8_, uint2>(const uint2& a) {
Float4_ tmp1, tmp2;
tmp1 = vec_conversion<Float4_, uint32_t>(a.x);
tmp2 = vec_conversion<Float4_, uint32_t>(a.y);
Float8_ res;
res.x = tmp1.x;
res.y = tmp1.y;
res.z = tmp2.x;
res.w = tmp2.y;
return res;
}
// half -> fp8
template <>
__inline__ __device__ uint8_t
vec_conversion<uint8_t, uint16_t>(const uint16_t& a) {
__half_raw tmp;
tmp.x = a;
return __hip_cvt_halfraw_to_fp8(tmp, fp8_type::__default_saturation,
fp8_type::__default_interpret);
}
template <>
__inline__ __device__ uint16_t
vec_conversion<uint16_t, uint32_t>(const uint32_t& a) {
union {
uint32_t ui32;
__half2_raw h2r;
} tmp;
tmp.ui32 = a;
return __hip_cvt_halfraw2_to_fp8x2(tmp.h2r, fp8_type::__default_saturation,
fp8_type::__default_interpret);
}
// bf16 -> fp8
template <>
__inline__ __device__ uint8_t
vec_conversion<uint8_t, __nv_bfloat16>(const __nv_bfloat16& a) {
return __hip_cvt_float_to_fp8(__bfloat162float(a),
fp8_type::__default_saturation,
fp8_type::__default_interpret);
}
// float -> fp8
template <>
__inline__ __device__ uint8_t vec_conversion<uint8_t, float>(const float& a) {
return __hip_cvt_float_to_fp8(a, fp8_type::__default_saturation,
fp8_type::__default_interpret);
}
// float2 -> half2
template <>
__inline__ __device__ uint32_t
vec_conversion<uint32_t, float2>(const float2& a) {
union {
half2 float16;
uint32_t uint32;
};
float16 = __float22half2_rn(a);
return uint32;
}
// Float4 -> half2x2
template <>
__inline__ __device__ uint2 vec_conversion<uint2, Float4_>(const Float4_& a) {
uint2 b;
float2 val;
val.x = a.x.x;
val.y = a.x.y;
b.x = vec_conversion<uint32_t, float2>(val);
val.x = a.y.x;
val.y = a.y.y;
b.y = vec_conversion<uint32_t, float2>(val);
return b;
}
// Float4 -> float4
template <>
__inline__ __device__ float4 vec_conversion<float4, Float4_>(const Float4_& a) {
float4 b;
b.x = a.x.x;
b.y = a.x.y;
b.z = a.y.x;
b.w = a.y.y;
return b;
}
// Float8 -> half2x4
template <>
__inline__ __device__ uint4 vec_conversion<uint4, Float8_>(const Float8_& a) {
uint4 b;
b.x = vec_conversion<uint32_t, float2>(a.x);
b.y = vec_conversion<uint32_t, float2>(a.y);
b.z = vec_conversion<uint32_t, float2>(a.z);
b.w = vec_conversion<uint32_t, float2>(a.w);
return b;
}
// float2 -> bfloat162
template <>
__inline__ __device__ __nv_bfloat162
vec_conversion<__nv_bfloat162, float2>(const float2& a) {
__nv_bfloat162 b = __float22bfloat162_rn(a);
return b;
}
// Float4 -> bfloat162x2
template <>
__inline__ __device__ bf16_4_t
vec_conversion<bf16_4_t, Float4_>(const Float4_& a) {
bf16_4_t b;
b.x = __float22bfloat162_rn(a.x);
b.y = __float22bfloat162_rn(a.y);
return b;
}
// Float8 -> bfloat162x4
template <>
__inline__ __device__ bf16_8_t
vec_conversion<bf16_8_t, Float8_>(const Float8_& a) {
bf16_8_t b;
b.x = __float22bfloat162_rn(a.x);
b.y = __float22bfloat162_rn(a.y);
b.z = __float22bfloat162_rn(a.z);
b.w = __float22bfloat162_rn(a.w);
return b;
}
/* Scaled and vectorized conversions, for data exchange between high and low
precision domains
@@ -335,6 +105,7 @@ vec_conversion<bf16_8_t, Float8_>(const Float8_& a) {
*/
using __nv_bfloat16 = __hip_bfloat16;
using __nv_bfloat162 = __hip_bfloat162;
// fp8 -> __nv_bfloat16
template <>
@@ -613,17 +384,6 @@ scaled_vec_conversion<uint32_t, float4>(const float4& a, float scale) {
}
#endif // ENABLE_FP8
template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
__inline__ __device__ Tout convert(const Tin& x) {
#ifdef ENABLE_FP8
if constexpr (kv_dt == Fp8KVCacheDataType::kFp8E4M3) {
return vec_conversion<Tout, Tin>(x);
}
#endif
assert(false);
return {}; // Squash missing return statement warning
}
template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
__inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) {
#ifdef ENABLE_FP8
@@ -282,19 +282,6 @@ __inline__ __device__ float4 scaled_vec_conversion<float4, uint32_t>(
}
#endif // ENABLE_FP8
template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
__inline__ __device__ Tout convert(const Tin& x) {
#if 0 // Disable the following code to reduce the binary size.
if constexpr (kv_dt == Fp8KVCacheDataType::kFp8E4M3) {
return vec_conversion<Tout, Tin>(x, __NV_E4M3);
} else if constexpr (kv_dt == Fp8KVCacheDataType::kFp8E5M2) {
return vec_conversion<Tout, Tin>(x, __NV_E5M2);
}
#endif
assert(false);
__builtin_unreachable(); // Suppress missing return statement warning
}
template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
__inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) {
#ifdef ENABLE_FP8