From 26bd889ff8b4cf83117cef87797e35a8d81a4e2c Mon Sep 17 00:00:00 2001
From: "liu.shen" <liu.shen@trs.com.cn>
Date: Sun, 9 Mar 2025 19:26:12 +0800
Subject: [PATCH] =?UTF-8?q?fix=20#829:=20=E5=85=BC=E5=AE=B9Intel=20Cascade?=
 =?UTF-8?q?=20Lake=E6=9E=B6=E6=9E=84=E7=9A=84CPU?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 third_party/llamafile/iqk_mul_mat.inc | 25 +++++++++++++++++++++++--
 1 file changed, 23 insertions(+), 2 deletions(-)

diff --git a/third_party/llamafile/iqk_mul_mat.inc b/third_party/llamafile/iqk_mul_mat.inc
index 99ee537..35bc7b8 100644
--- a/third_party/llamafile/iqk_mul_mat.inc
+++ b/third_party/llamafile/iqk_mul_mat.inc
@@ -2385,7 +2385,12 @@ struct SimpleBits {
     __m256i values[4];
 };
 
-
+// fix for #829: 添加对 AVX512VPOPCNTDQ 的检测
+#if defined(HAVE_FANCY_SIMD) && defined(__AVX512VPOPCNTDQ__)
+#define HAVE_AVX512_POPCNT 1
+#else
+#define HAVE_AVX512_POPCNT 0
+#endif
 
 struct EvenSignHelper {
     #if defined HAVE_FANCY_SIMD
@@ -2396,7 +2401,23 @@ struct EvenSignHelper {
         };
         IQK_ALWAYS_INLINE void sign_2_values(__m256i aux, __m256i * values) const {
             aux = _mm256_and_si256(_mm256_srlv_epi32(aux, shifts), mask);
-            auto pcnt = _mm256_popcnt_epi32(aux);
+            
+            // fix for #829: 兼容Intel Cascade Lake架构的CPU，如果不支持AVX512VPOPCNTDQ扩展，则使用替代实现
+            #if HAVE_AVX512_POPCNT
+                auto pcnt = _mm256_popcnt_epi32(aux);
+                
+            #else
+                // 提供替代实现，使用标准的位计数方法
+                __m256i pcnt;
+                int* pcnt_ptr = reinterpret_cast<int*>(&pcnt);
+                int* aux_ptr = reinterpret_cast<int*>(&aux); // 直接获取 aux 的地址，避免不必要的复制
+                
+                #pragma unroll 8  // 提示编译器展开循环，提高 SIMD 计算吞吐量
+                for (int i = 0; i < 8; i++) {
+                    pcnt_ptr[i] = __builtin_popcount(aux_ptr[i]); // 使用编译器内置 popcount
+                }
+            #endif
+            
             sbits_t sbits;
             sbits.vec = _mm256_cvtepi32_epi8(_mm256_or_si256(aux, _mm256_slli_epi32(_mm256_and_si256(pcnt, mone), 7)));
             values[0] = _mm256_mask_sub_epi8(values[0], sbits.mask[0], _mm256_setzero_si256(), values[0]);