From d5c05821f3c3d6cabe8ac45776fe0ecb0da13eca Mon Sep 17 00:00:00 2001
From: junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
Date: Thu, 30 May 2024 17:30:10 +0800
Subject: [PATCH] ggml : fix loongarch build (O2 issue) (#7636)

---
 ggml-quants.c | 20 ++++++++++++++------
 ggml.c        |  2 +-
 2 files changed, 15 insertions(+), 7 deletions(-)

diff --git a/ggml-quants.c b/ggml-quants.c
index 4f2c7224c..1128d66e2 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -6828,6 +6828,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         int bit = 0;
         int is  = 0;
+        __m256i xvbit;
 
         const uint8_t * restrict q3 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -6836,21 +6837,25 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             // load low 2 bits
             const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             // prepare low and high bits
             const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
-            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
-            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
-            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit);
             const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
-            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2);
+            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
             ++bit;
 
             // load Q8 quants
@@ -8033,6 +8038,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         __m256i sumi = __lasx_xvldi(0);
 
         int bit = 0;
+        __m256i xvbit;
 
         for (int j = 0; j < QK_K/64; ++j) {
 
@@ -8041,13 +8047,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
             const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
 
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
             const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
-            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4);
+            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
             const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
             hmask = __lasx_xvslli_h(hmask, 1);
 
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
             const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
-            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4);
+            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
             const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
             hmask = __lasx_xvslli_h(hmask, 1);
 
diff --git a/ggml.c b/ggml.c
index b2b725f65..f3a90ff2c 100644
--- a/ggml.c
+++ b/ggml.c
@@ -1580,7 +1580,7 @@ do {                                                              \
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
 #define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
 
-static inline __m256 __lasx_f32cx8_load(ggml_fp16_t *x) {
+static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {