From 46f21a5956f85462208ff005e6cbc86f142f8046 Mon Sep 17 00:00:00 2001
From: Ma Mingfei <mingfei.ma@intel.com>
Date: Thu, 4 Dec 2025 12:34:05 +0800
Subject: [PATCH] use faster covnersion from float8_e4m3fn to bfloat16 (#12316)

---
 sgl-kernel/csrc/cpu/gemm_fp8.cpp | 14 +++++++++-----
 sgl-kernel/csrc/cpu/vec.h        | 24 ++++++++++++++++++++++++
 2 files changed, 33 insertions(+), 5 deletions(-)

diff --git a/sgl-kernel/csrc/cpu/gemm_fp8.cpp b/sgl-kernel/csrc/cpu/gemm_fp8.cpp
index 008f83298..b2821982a 100644
--- a/sgl-kernel/csrc/cpu/gemm_fp8.cpp
+++ b/sgl-kernel/csrc/cpu/gemm_fp8.cpp
@@ -56,7 +56,8 @@ inline void unpack_B(
   const int K2 = K >> 1;
   const int ldb2 = ldb;  // ldb * 2 >> 1;
   const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(packed_B);
-  const __m512 vd = _mm512_set1_ps(scale);
+  const __m512 vexp = _mm512_castsi512_ps(_mm512_set1_epi32(kFP8_BIAS));
+  const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(scale), vexp);
 
   constexpr int BLOCK_N = block_size_n();
   static_assert(BLOCK_N == 32);
@@ -74,8 +75,8 @@ inline void unpack_B(
     __m256i b8_0 = _mm512_extracti32x8_epi32(b8, 0);
     __m256i b8_1 = _mm512_extracti32x8_epi32(b8, 1);
 
-    __m512bh bf16_0 = CVT_FP8_TO_BF16(b8_0);
-    __m512bh bf16_1 = CVT_FP8_TO_BF16(b8_1);
+    __m512bh bf16_0 = CVT_FP8_TO_BF16_EXT(b8_0);
+    __m512bh bf16_1 = CVT_FP8_TO_BF16_EXT(b8_1);
 
     // Apply scale
     __m512 f0_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 0));
@@ -147,6 +148,8 @@ struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BL
     // block quant scale
     __m512 vscale;
 
+    const __m512 vexp = _mm512_castsi512_ps(_mm512_set1_epi32(kFP8_BIAS));
+
     auto loadc = [&](auto i) {
       constexpr int col = i % COLS;
       if constexpr (has_bias) {
@@ -178,8 +181,8 @@ struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BL
           if constexpr (PREFETCH_SIZE_K > 0) {
             _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
           }
-          vb[col + 0] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 0));
-          vb[col + 1] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 1));
+          vb[col + 0] = CVT_FP8_TO_BF16_EXT(_mm512_extracti32x8_epi32(b8, 0));
+          vb[col + 1] = CVT_FP8_TO_BF16_EXT(_mm512_extracti32x8_epi32(b8, 1));
         }
       }
       vsum[i] = _mm512_dpbf16_ps(vsum[i], va, vb[col]);
@@ -191,6 +194,7 @@ struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BL
       int kb_end = std::min(K >> 1, kb_start + BLOCK_K2);
       // 1. load scale vector
       vscale = _mm512_set1_ps(scale[kb]);
+      vscale = _mm512_mul_ps(vscale, vexp);
       if constexpr (PREFETCH_SIZE_KB > 0) {
         _mm_prefetch(scale + kb + PREFETCH_SIZE_KB, _MM_HINT_T0);
       }
diff --git a/sgl-kernel/csrc/cpu/vec.h b/sgl-kernel/csrc/cpu/vec.h
index d0ec53d83..486ff260a 100644
--- a/sgl-kernel/csrc/cpu/vec.h
+++ b/sgl-kernel/csrc/cpu/vec.h
@@ -121,6 +121,30 @@ inline __m512bh CVT_FP8_TO_BF16(__m256i a) {
 #endif
 }
 
+// faster version of float8_e4m3fn conversion to bfloat16
+//
+// we mapped cuda implementation from below link and vectorized with avx512:
+// https://github.com/thu-pacman/chitu/blob/1ed2078ec26581ebdca05b7306d4385f86edaa7c/csrc/cuda/marlin/marlin_gemm/dequant.h#L387
+//
+inline __attribute__((always_inline)) __m512bh CVT_FP8_TO_BF16_EXT(__m256i a) {
+  const __m512i mask0 = _mm512_set1_epi16(0x80);  // sign bit
+  const __m512i mask1 = _mm512_set1_epi16(0x7F);  // exponent and mantissa
+  const __m512i mask2 = _mm512_set1_epi16(0x4000);
+
+  __m512i x = _mm512_cvtepu8_epi16(a);
+  __m512i vsign = _mm512_and_si512(x, mask0);
+  vsign = _mm512_slli_epi16(vsign, 8);
+
+  __m512i vexp_and_mant = _mm512_and_si512(x, mask1);
+  vexp_and_mant = _mm512_slli_epi16(vexp_and_mant, 4);
+
+  // _MM_TERNLOG_A | _MM_TERNLOG_B | _MM_TERNLOG_C: 0b11111110
+  return (__m512bh)(_mm512_ternarylogic_epi32(vsign, mask2, vexp_and_mant, 0b11111110));
+}
+
+// bias for conversion of fp8 to bf16 1/256 in float32
+#define kFP8_BIAS 0x3b800000
+
 #endif
 
 // vector to scalar reduction