Bitnet: 2.25 bpw version

Just scaler and AVX2 for now.
PP-512 is even faster (325 t/s on the Ryzn-7950X, 404 t/s on
Ryzen-5975WX). We lose ~6-7% for TG due to being memory bound and
the model being 10% larger.

3 files changed, 25 insertions, 81 deletions

diff --git a/ggml-common.h b/ggml-common.h
index c7f865e8..f5a35960 100644
--- a/ggml-common.h
+++ b/ggml-common.h
@@ -381,13 +381,14 @@ typedef struct {
 } block_iq1_bn;
 static_assert(sizeof(block_iq1_bn) == sizeof(uint16_t) + QK_IQ1BN/8 + QK_IQ1BN/16, "wrong iq1_bn block size/padding");
 //
-// Bitnet - implemented as 2.0 bpw
+// Bitnet - implemented as 2.25 bpw
 //
 #define QK_IQ2BN 64
 typedef struct {
+    ggml_half d;
     uint8_t qs[QK_IQ2BN/4];
 } block_iq2_bn;
-static_assert(sizeof(block_iq2_bn) == QK_IQ2BN/4, "wrong iq2_bn block size/padding");
+static_assert(sizeof(block_iq2_bn) == sizeof(ggml_half) + QK_IQ2BN/4, "wrong iq2_bn block size/padding");
 
 // Used by IQ1_M quants
 typedef union {
diff --git a/iqk-quantize.cpp b/iqk-quantize.cpp
index 287a7dd3..42e5a264 100644
--- a/iqk-quantize.cpp
+++ b/iqk-quantize.cpp
@@ -159,25 +159,14 @@ void IQ1BNQuantizer::quantize_one_row_2bn(const float * src, block_iq2_bn * y, i
 
     const int nblock = n_per_row/QK_IQ1BN;
 
-    const auto& iq1bn = get_iq1bn_data();
-
     auto max_in_row = row_max(n_per_row, src);
-
-    ggml_half * d = (ggml_half *)y;
-    *d = GGML_FP32_TO_FP16(max_in_row);
-
-    auto ql = (uint8_t *)(d + 2);
-    auto qh = ql + QK_IQ1BN/8;
-    std::memset(ql, 0, QK_IQ1BN/8);
-    std::memset(qh, 0, QK_IQ1BN/16);
-    auto xb = src;
-    auto extra = quantize_one_block_1bn(iq1bn, xb, L, ql, qh);
-    *(uint16_t *)(d + 1) = extra;
+    ggml_half dh = GGML_FP32_TO_FP16(max_in_row);
 
     constexpr int Nj = QK_IQ1BN/4;
 
-    for (int ib = 1; ib < nblock; ++ib) {
-        xb = src + QK_IQ1BN*ib;
+    for (int ib = 0; ib < nblock; ++ib) {
+        y[ib].d = dh;
+        auto xb = src + QK_IQ1BN*ib;
         for (int j = 0; j < QK_IQ1BN; ++j) {
             L[j] = fabsf(xb[j]) < 1e-6f ? 1 : xb[j] < 0 ? 0 : 2;
         }
@@ -258,22 +247,11 @@ void dequantize_row_iq2_bn(const block_iq2_bn * x, float * y, int64_t k) {
     assert(k%QK_IQ1BN == 0);
     int nblock = k / QK_IQ1BN;
 
-    float d = GGML_FP16_TO_FP32(*(const ggml_half *)x);
-    auto * extra_ptr = (const uint16_t *)x;
-    auto extra = extra_ptr[1];
-    auto ql = (const uint8_t *)(extra_ptr + 2);
-    auto qh = ql + QK_IQ1BN/8;
-    for (int l = 0; l < QK_IQ1BN/8; ++l) {
-        uint16_t idx = ql[l] | ((qh[l/2] << (8 - 4*(l%2))) & 0x0f00);
-        uint16_t val = iq1bn_grid_u16[idx];
-        float dls = extra & (1 << l) ? -d : d;
-        for (int j = 0; j < 8; ++j) y[j] = dls * (((val >> 2*j) & 3) - 1);
-        y += 8;
-    }
+    float d = GGML_FP16_TO_FP32(x[0].d);
     auto m = -d;
     auto d1 = d, d2 = d*0.25f, d3 = d2*0.25f, d4 = d3*0.25f;
     constexpr int Nj = QK_IQ1BN/4;
-    for (int i = 1; i < nblock; ++i) {
+    for (int i = 0; i < nblock; ++i) {
         for (int j = 0; j < Nj; ++j) {
             y[j+   0] = d1*(x[i].qs[j] & 0x03) + m;
             y[j+1*Nj] = d2*(x[i].qs[j] & 0x0c) + m;
@@ -396,25 +374,10 @@ void ggml_vec_dot_iq2_bn_q8_K64(int n, float * s, size_t bs, const void * vx, si
 
     float sumf = 0;
 
-    float d = GGML_FP16_TO_FP32(*(const ggml_half *)x);
-    auto * extra_ptr = (const uint16_t *)x;
-    auto extra = extra_ptr[1];
-    auto ql = (const uint8_t *)(extra_ptr + 2);
-    auto qh = ql + QK_IQ1BN/8;
-    auto q8 = y[0].qs;
-    int sumi = 0;
-    for (int k = 0; k < QK_IQ1BN/8; ++k) {
-        uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
-        uint16_t val = iq1bn_grid_u16[idx];
-        int s = 0;
-        for (int j = 0; j < 8; ++j) s += q8[j] * (((val >> 2*j) & 3) - 1);
-        sumi += extra & (1 << k) ? -s : s;
-        q8 += 8;
-    }
-    sumf += y[0].d * sumi;
+    float d = GGML_FP16_TO_FP32(x[0].d);
 
-    for (int i = 1; i < nblock; ++i) {
-        q8 = y[i].qs;
+    for (int i = 0; i < nblock; ++i) {
+        auto q8 = y[i].qs;
         int s0 = 0, s1 = 0, s2 = 0, s3 = 0, s4 = 0;
         for (int j = 0; j < Nj; ++j) {
             s1 += q8[j+   0] * (x[i].qs[j] & 0x03);
diff --git a/iqk_mul_mat.cpp b/iqk_mul_mat.cpp
index 83d3a472..9eb063e9 100644
--- a/iqk_mul_mat.cpp
+++ b/iqk_mul_mat.cpp
@@ -1412,14 +1412,8 @@ IQK_NOINLINE void mul_mat_iq2bn_q8_K64(int n, const void * vx, size_t bx, const
     const int nb = n / QK_IQ1BN;
     Q8_K64<nrc_y> q8(info);
     __m256  accd[nrc_y];
-    __m256i signs[2];
 
     const auto m1_8   = _mm256_set1_epi8(1);
-    const auto shuff1 = _mm256_set_epi64x(0x0808080808080808, 0x0000000000000000, 0x0808080808080808, 0x0000000000000000);
-    const auto shuff2 = _mm256_add_epi8(shuff1, m1_8);
-    const auto shuff3 = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000);
-    const auto shuff4 = _mm256_set_epi64x(0x0707070707070707, 0x0606060606060606, 0x0505050505050505, 0x0404040404040404);
-    const auto mask1  = _mm256_set1_epi64x(0x8040201008040201);
     const auto mask2  = _mm256_set1_epi8(3);
 #if !(defined __AVX512VNNI__ && defined __AVX512VL__)
     const auto m1_16  = _mm256_set1_epi16(1);
@@ -1428,37 +1422,23 @@ IQK_NOINLINE void mul_mat_iq2bn_q8_K64(int n, const void * vx, size_t bx, const
     for (int ix = 0; ix < nrc_x; ++ix) {
 
         const block_iq2_bn * x = (const block_iq2_bn *)((const char *)vx + ix*bx);
-        float d = GGML_FP16_TO_FP32(*(const ggml_half *)x);
-        auto extra_ptr = (const uint16_t *)x;
+        float d = GGML_FP16_TO_FP32(x[0].d);
 
-        auto all_signs = _mm256_set1_epi8(extra_ptr[1]);
-        all_signs = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(all_signs, mask1), mask1), m1_8);
-        signs[0] =  _mm256_shuffle_epi8(all_signs, shuff3);
-        signs[1] =  _mm256_shuffle_epi8(all_signs, shuff4);
-
-        auto ql = (const uint8_t *)(extra_ptr + 2);
-        auto qh = ql + QK_IQ1BN/8;
-        auto aux1 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[3] | ((qh[1] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[2] | ((qh[1] << 8) & 0x0f00)],
-                                      iq1bn_grid_xxx[ql[1] | ((qh[0] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[0] | ((qh[0] << 8) & 0x0f00)]);
-        auto aux2 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[7] | ((qh[3] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[6] | ((qh[3] << 8) & 0x0f00)],
-                                      iq1bn_grid_xxx[ql[5] | ((qh[2] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[4] | ((qh[2] << 8) & 0x0f00)]);
-
-        auto v1 = _mm256_sub_epi8(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff2), mask1), mask1),
-                                  _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff1), mask1), mask1));
-        auto v2 = _mm256_sub_epi8(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff2), mask1), mask1),
-                                  _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff1), mask1), mask1));
-
-        v1 = _mm256_sign_epi8(v1, signs[0]);
-        v2 = _mm256_sign_epi8(v2, signs[1]);
-        for (int iy = 0; iy < nrc_y; ++iy) {
-            auto dot1 = _mm256_sign_epi8(q8.load_quants(iy, 0, 0), v1);
-            auto dot2 = _mm256_sign_epi8(q8.load_quants(iy, 0, 1), v2);
+        {
+            auto q2bits = _mm_loadu_si128((const __m128i *)x[0].qs);
+            auto q2 = MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits);
+            auto v1 = _mm256_sub_epi8(_mm256_and_si256(q2, mask2), m1_8);
+            auto v2 = _mm256_sub_epi8(_mm256_and_si256(_mm256_srli_epi16(q2, 4), mask2), m1_8);
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                auto dot1 = _mm256_sign_epi8(q8.load_quants(iy, 0, 0), v1);
+                auto dot2 = _mm256_sign_epi8(q8.load_quants(iy, 0, 1), v2);
 #if defined __AVX512VNNI__ && defined __AVX512VL__
-            auto dot = _mm256_dpbusd_epi32(_mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot1), m1_8, dot2);
+                auto dot = _mm256_dpbusd_epi32(_mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot1), m1_8, dot2);
 #else
-            auto dot = _mm256_madd_epi16(m1_16, _mm256_add_epi16(_mm256_maddubs_epi16(m1_8, dot1), _mm256_maddubs_epi16(m1_8, dot2)));
+                auto dot = _mm256_madd_epi16(m1_16, _mm256_add_epi16(_mm256_maddubs_epi16(m1_8, dot1), _mm256_maddubs_epi16(m1_8, dot2)));
 #endif
-            accd[iy] = _mm256_mul_ps(_mm256_set1_ps(q8.scale(iy, 0)), _mm256_cvtepi32_ps(dot));
+                accd[iy] = _mm256_mul_ps(_mm256_set1_ps(q8.scale(iy, 0)), _mm256_cvtepi32_ps(dot));
+            }
         }
 
         for (int i = 1; i < nb; ++i) {