Revert "AVX implementations (#1370)"

This reverts commit 948d124.

Author: ggerganov

SHA256SUMS

@@ -1,19 +1,24 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
 666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ae89af479ab4d31c4e555ad8cc1dc9bf1f68d617186158cc381cd5a0fccd10bd  models/7B/ggml-model-q4_0.bin
-862072e2036a1bdb1a01ec2e159381f332a9e2357b886031c075fb7efa86db9b  models/7B/ggml-model-q4_1.bin
-0bef7cefa880a67a0b6d2a7e4559ded235823535ad616808dd8b5e47ff0a202f  models/7B/ggml-model-q5_0.bin
-97b9c38b2b8aed0c0aa90e0a975570ce3455c47d62128b382c55acbf6e2035f6  models/7B/ggml-model-q5_1.bin
+99aeb35f26b577fa2732716cca4d8b5ada39a78ea9b2dca2651fc632b5d101b6  models/7B/ggml-model-q4_0.bin
+cc061458339a3eb8bcecbf0a825e9924fb7d1a8150f63cd5d091caa99215aafe  models/7B/ggml-model-q4_1.bin
+25b050337a87344da687a7f2adddc03bd99b7f6c140450e836649f3585fb6496  models/7B/ggml-model-q4_2.bin
 7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265  models/7B/params.json
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
 2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
+eecb575d325d935157761172e2bf05984dad216eb2b06777b73463cf9b818bab  models/13B/ggml-model-q4_0.bin
+d9581b5b88e5622532fe897c9f9b0e67a317d22dd27a6f90fa4ab8c6d23ccdbb  models/13B/ggml-model-q4_1.bin
+75a218a47df03f5f96354656329864613abcb67779412b9bc2282b28c1c3cbaa  models/13B/ggml-model-q4_2.bin
 4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f  models/13B/params.json
 e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/consolidated.00.pth
 4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff  models/30B/consolidated.01.pth
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
 7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
+517b9e525742c42b5478a6280a4b41ec66f46298c57aba7f0453d491682fe42d  models/30B/ggml-model-q4_0.bin
+7b75ac615fa369ee593493a7e6ef87542bf0350255db928b22c5a24f6d598bcd  models/30B/ggml-model-q4_1.bin
+aadbc9cf806313a55be570f62884eed289d30c313fac3b7838717e01bd553204  models/30B/ggml-model-q4_2.bin
 2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb  models/30B/params.json
 135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe  models/65B/consolidated.00.pth
 9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde  models/65B/consolidated.01.pth
@@ -24,5 +29,8 @@ a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/con
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
 60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
+01672072136f8be6ca9d7cebe5f86ed316e8b85851b9fe3de951809233cea4f2  models/65B/ggml-model-q4_0.bin
+4743a28aac3e5f32a6e838a815f51d3779de44fbbe251d745251e66c23c5950f  models/65B/ggml-model-q4_1.bin
+1b6f6588d0e2ecfe6c4d849088e48e5e3083466b962daa32e3261363e21fc5e9  models/65B/ggml-model-q4_2.bin
 999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b  models/65B/params.json
 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347  models/tokenizer.model

ggml.c

@@ -473,16 +473,23 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 //
 
 #if __AVX__ || __AVX2__ || __AVX512F__
-// multiply int8_t, add results pairwise twice
-static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
-    // Get absolute values of x vectors
-    const __m128i ax = _mm_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = _mm_sign_epi8(y, x);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = _mm_maddubs_epi16(ax, sy);
-    const __m128i ones = _mm_set1_epi16(1);
-    return _mm_madd_epi16(ones, dot);
+// Unpack 16 4-bit fields into 16 bytes
+// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval
+static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi)
+{
+    // Load 8 bytes from memory
+    __m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi );
+
+    // Expand bytes into uint16_t values
+    __m128i bytes = _mm_cvtepu8_epi16( tmp );
+
+    // Unpack values into individual bytes
+    const __m128i lowMask = _mm_set1_epi8( 0xF );
+    __m128i high = _mm_andnot_si128( lowMask, bytes );
+    __m128i low = _mm_and_si128( lowMask, bytes );
+    high = _mm_slli_epi16( high, 4 );
+    bytes = _mm_or_si128( low, high );
+    return bytes;
 }
 
 // horizontally add 8 floats
@@ -529,10 +536,19 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) {
 // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
 static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
 {
-    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
-    const __m256i bytes = _mm256_set_m128i(_mm_srli_epi16(tmp, 4), tmp);
+    // Load 16 bytes from memory
+    __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi );
+
+    // Expand bytes into uint16_t values
+    __m256i bytes = _mm256_cvtepu8_epi16( tmp );
+
+    // Unpack values into individual bytes
     const __m256i lowMask = _mm256_set1_epi8( 0xF );
-    return _mm256_and_si256(lowMask, bytes);
+    __m256i high = _mm256_andnot_si256( lowMask, bytes );
+    __m256i low = _mm256_and_si256( lowMask, bytes );
+    high = _mm256_slli_epi16( high, 4 );
+    bytes = _mm256_or_si256( low, high );
+    return bytes;
 }
 
 // add int16_t pairwise and return as float vector
@@ -2109,23 +2125,31 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         // Compute combined scale for the block
         const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
 
-        const __m128i lowMask = _mm_set1_epi8(0xF);
-        const __m128i off = _mm_set1_epi8(8);
+        __m128i i32[2];
+        for (int j = 0; j < 2; ++j) {
+            // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes
+            __m128i bx = bytes_from_nibbles_16(x[i].qs + 8*j);
+            __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j));
+
+            // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+            const __m128i off = _mm_set1_epi8( 8 );
+            bx = _mm_sub_epi8( bx, off );
 
-        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);
+            // Get absolute values of x vectors
+            const __m128i ax = _mm_sign_epi8(bx, bx);
 
-        __m128i bx = _mm_and_si128(lowMask, tmp);
-        __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs);
-        bx = _mm_sub_epi8(bx, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx, by);
+            // Sign the values of the y vectors
+            const __m128i sy = _mm_sign_epi8(by, bx);
 
-        bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
-        by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
-        bx = _mm_sub_epi8(bx, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx, by);
+            // Perform multiplication and create 16-bit values
+            const __m128i dot = _mm_maddubs_epi16(ax, sy);
+
+            const __m128i ones = _mm_set1_epi16(1);
+            i32[j] = _mm_madd_epi16(ones, dot);
+        }
 
         // Convert int32_t to float
-        __m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1));
+        __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] ));
         // Apply the scale, and accumulate
         acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
     }
@@ -2472,8 +2496,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         int sumi = 0;
 
         for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
 
             const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
             const int32_t x1 = ((x[i].qs[j] >>   4) | xh_1) - 16;
@@ -2698,8 +2722,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         int sumi = 0;
 
         for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
 
             const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0;
             const int32_t x1 = (x[i].qs[j] >>  4) | xh_1;