Fix QB4 segfault on Sandy Bridge/Ivy Bridge

- Replace aligned vksum load with unaligned for QB4 PiperOrigin-RevId: 724769160
google · Feb 9, 2025 · e55b399 · e55b399
1 parent 3acb27f
commit e55b399
Show file tree

Hide file tree

Showing 25 changed files with 25 additions and 30 deletions.
diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-avx-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-avx-ld128.c
@@ -60,7 +60,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__avx_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
 
     __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-avx-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-avx-ld64.c
@@ -60,7 +60,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__avx_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
 
     __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse2-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse2-ld128.c
@@ -56,7 +56,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse2_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
     vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse2-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse2-ld64.c
@@ -56,7 +56,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse2_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
     vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse41-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse41-ld128.c
@@ -56,7 +56,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse41_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
     vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse41-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-minmax-sse41-ld64.c
@@ -56,7 +56,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse41_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
     vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-avx-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-avx-ld128.c
@@ -66,7 +66,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__avx_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-avx-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-avx-ld64.c
@@ -66,7 +66,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__avx_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
 

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse2-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse2-ld128.c
@@ -62,7 +62,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__sse2_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse2-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse2-ld64.c
@@ -62,7 +62,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__sse2_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse41-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse41-ld128.c
@@ -62,7 +62,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__sse41_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse41-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x4c8-minmax-sse41-ld64.c
@@ -62,7 +62,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x4c8__sse41_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-avx-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-avx-ld128.c
@@ -72,7 +72,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__avx_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
     const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-avx-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-avx-ld64.c
@@ -72,7 +72,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__avx_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
     const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse2-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse2-ld128.c
@@ -68,7 +68,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__sse2_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse2-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse2-ld64.c
@@ -68,7 +68,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__sse2_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse41-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse41-ld128.c
@@ -68,7 +68,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__sse41_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse41-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-minmax-sse41-ld64.c
@@ -68,7 +68,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x4c8__sse41_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-avx-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-avx-ld128.c
@@ -78,7 +78,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__avx_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
     const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-avx-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-avx-ld64.c
@@ -78,7 +78,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__avx_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
     const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
     const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse2-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse2-ld128.c
@@ -74,7 +74,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__sse2_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse2-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse2-ld64.c
@@ -74,7 +74,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__sse2_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse41-ld128.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse41-ld128.c
@@ -74,7 +74,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__sse41_ld128(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse41-ld64.c b/src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-sse41-ld64.c
@@ -74,7 +74,7 @@ void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x4c8__sse41_ld64(
   XNN_FORCE_REALIZATION(vmask);
 
   do {
-    const __m128 vksum = _mm_load_ps((const float*) w);
+    const __m128 vksum = _mm_loadu_ps((const float*) w);
     const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
     const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
     const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));

diff --git a/src/qs8-gemm/MRx4c8-sse.c.in b/src/qs8-gemm/MRx4c8-sse.c.in
@@ -198,7 +198,7 @@ void xnn_${DATATYPE_SPEC}_gemm${GEMM_SUFFIX}_minmax${REQUANTIZATION_SPEC}_ukerne
           __m128i vacc${M}x2 = _mm_unpacklo_epi64(vksum23${M}, vzero);
           __m128i vacc${M}x3 = _mm_unpackhi_epi64(vksum23${M}, vzero);
     $elif DATATYPE == "QB4":
-      const __m128 vksum = _mm_load_ps((const float*) w);
+      const __m128 vksum = _mm_loadu_ps((const float*) w);
       $if AVX:
         $for M in range(MR):
           const __m128i vinput_zero_point${M} = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[${M}].zero_point));
@@ -404,11 +404,6 @@ void xnn_${DATATYPE_SPEC}_gemm${GEMM_SUFFIX}_minmax${REQUANTIZATION_SPEC}_ukerne
               $# SSE 2 only
               $else:
                 ${_}const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8);
-          $elif VARIANT == "EXTENDED":
-            $if N == 0:
-              ${_}const __m128i vxb${N} = _mm_load_si128((const __m128i*) w);
-            $else:
-              ${_}const __m128i vxb${N} = _mm_load_si128((const __m128i*) ((const int16_t*) w + ${N * 8}));
 
           $for M in range(MR):
             ${_}vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N}));