ggml : IQ4_NL sgemm + Q4_0 AVX optimization (llama/9422)

* squashed

readd my iq4_nl sgemm PR https://github.com/ggerganov/llama.cpp/pull/8049

have ggml_vec_dot_q4_0 do two blocks per loop for avx

try out f16c ggml_vec_dot_iq4_nl, but it's not really faster. as per https://github.com/ggerganov/llama.cpp/pull/8549 we can calculate several blocks at a time with no issue

* shuffle

* remove f16c iq4_nl as i cant make it faster than before

ggml/src/ggml-quants.c

@@ -230,6 +230,12 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
 
     return _mm_packus_epi16( bytes1, bytes2);
 }
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
 #endif
 #elif defined(__SSSE3__)
 // horizontally add 4x4 floats
@@ -4206,37 +4212,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
     sumf = hsum_float_8(acc);
 #elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        const __m128i lowMask = _mm_set1_epi8(0xF);
-        const __m128i off = _mm_set1_epi8(8);
-
-        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[ib].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
-        by_0 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0);
+    const __m128i mone = _mm_set1_epi16(1);
 
-        // Convert int32_t to float
-        __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
 
-        // Apply the scale, and accumulate
-        acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
+        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
     }
 
-    sumf = hsum_float_8(acc);
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
 #elif defined(__SSSE3__)
     // set constants
     const __m128i lowMask = _mm_set1_epi8(0xF);
@@ -11819,15 +11825,6 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 #endif
 }
 
-
-#if defined(__AVX__)
-static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
-    const __m128i ax = _mm_sign_epi8(x, x);
-    const __m128i sy = _mm_sign_epi8(y, x);
-    return _mm_maddubs_epi16(ax, sy);
-}
-#endif
-
 #if defined(__AVX2__)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
     const __m256i ax = _mm256_sign_epi8(x, x);