implement requested changes

Author: Timen

src/layer/x86/avx_usability.h

@@ -20,7 +20,7 @@
 
 static inline __m256 loadfp16(const unsigned short* ptr)
 {
-    return _mm256_cvtph_ps(_mm_load_si128((__m128i*)(ptr)));
+    return _mm256_cvtph_ps(_mm_lddqu_si128((__m128i*)(ptr)));
 }
 static inline __m256 _mm256_fmadd_1_ps(__m256 a, __m256 b, float c)
 {

src/layer/x86/convolution_3x3_pack8.h

@@ -65,8 +65,10 @@ static void conv3x3s1_winograd64_transform_kernel_pack8_avx(const Mat& kernel, M
     }
     // interleave
     // src = 64-inch-outch
-    // dst = 4b-4a-inch/4a-64-outch/4b;
-    kernel_tm_pack8.create(inch / 8, 64, (outch / 4) / 2 + (outch / 4) % 2, (size_t)4u * 64, 64);
+    // dst = 8b-8a-inch/8a-64-outch/8b;
+    kernel_tm_pack8.create(inch / 8, 64, outch / 8, (size_t)4u * 64, 64);
+
+    // kernel_tm_pack8.create(inch / 8, 64, (outch / 4) / 2 + (outch / 4) % 2, (size_t)4u * 64, 64);
     int q = 0;
     for (; q + 7 < outch; q += 8)
     {

src/layer/x86/convolution_3x3_pack8_fp16.h

@@ -65,8 +65,8 @@ static void conv3x3s1_winograd64_transform_kernel_fp16_pack8_avx(const Mat& kern
     }
     // interleave
     // src = 64-inch-outch
-    // dst = 8b-8a-inch/8a-84-outch/8b;
-    kernel_tm_pack8.create(inch / 8, 64, (outch / 4) / 2 + (outch / 4) % 2, (size_t)2u * 64, 64);
+    // dst = 8b-8a-inch/8a-64-outch/8b;
+    kernel_tm_pack8.create(inch / 8, 64, outch / 8, (size_t)2u * 64, 64);
     int q = 0;
     for (; q + 7 < outch; q += 8)
     {