Faster IQ4_XS_R4 on Zen4 (#128)

* Faster iq4_xs_r4 on Zen4

The trick is to simply prepare the Q8 block sums for
blocks of 32 as floats. This brings PP-512 up to 254.6 t/s
from 224 t/s.

* Fix broken matrix x vector product on Zen4

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

4 files changed, 54 insertions, 27 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 69bb6d88..974e42b2 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1124,6 +1124,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q8_K16,
         .row_meta_size            = 20,
     },
+    [GGML_TYPE_Q8_K32] = {
+        .type_name                = "q8_K32",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_q8_K),
+        .is_quantized             = true,
+        .from_float               = quantize_row_q8_K32,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_BF16] = {
         .type_name                = "bf16",
         .blck_size                = 1,
@@ -1292,7 +1300,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_iq4_xs_r4,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_xs_r4_ref,
         .vec_dot                  = vec_dot_iq4_xs_r4_q8_k,
-        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K32,
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
@@ -15633,6 +15641,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_Q8_K64:
         case GGML_TYPE_Q8_K16:
+        case GGML_TYPE_Q8_K32:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 69c0b9a4..a9adedfc 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -2917,15 +2917,16 @@ static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn
 
 template <int nrc_y>
 static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
-    GGML_ASSERT(nrc_x%8 == 0);
+    GGML_ASSERT(nrc_x%4 == 0);
     Q8<nrc_y, block_q8_K> q8(info);
     auto m4 = _mm256_set1_epi8(0xf);
 #ifndef HAVE_FANCY_SIMD
     auto m1 = _mm256_set1_epi16(1);
-#endif
     auto values128 = _mm_loadu_si128((const __m128i *)iq4k_values);
     auto values = MM256_SET_M128I(values128, values128);
-    //auto values = load_iq4nl_values_256();
+#else
+    auto values = load_iq4nl_values_256();
+#endif
     int nbl = n / QK_K;
     using helper_t = union { __m256i vec; uint32_t val[8]; };
     helper_t h;
@@ -2969,7 +2970,7 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
                     sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
                     float d8 = q8.scale(iy, ibl);
-                    float m8 = d8 * (q8.y[iy][ibl].bsums[2*ib+0] + q8.y[iy][ibl].bsums[2*ib+1]);
+                    float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
                     acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(d8)), _mm256_cvtepi32_ps(sumi), acc[iy]);
                     acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(m8), acc[iy]);
 #else
@@ -2979,15 +2980,6 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
                                                   _mm256_maddubs_epi16(s4, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xff), qx[3])));
                     auto sumi = _mm256_add_epi32(_mm256_madd_epi16(m1, sumi1), _mm256_madd_epi16(m1, sumi2));
                     acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
-                    //auto sumi1 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00))),
-                    //                              _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55))));
-                    //auto sumi2 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[2], _mm256_shuffle_epi32(y, 0xaa))),
-                    //                              _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[3], _mm256_shuffle_epi32(y, 0xff))));
-                    //auto sumi  = _mm256_add_epi32(sumi1, sumi2);
-                    //float d8 = q8.scale(iy, ibl);
-                    //float m8 = d8 * (q8.y[iy][ibl].bsums[2*ib+0] + q8.y[iy][ibl].bsums[2*ib+1]);
-                    //acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(d8)), _mm256_cvtepi32_ps(sumi), acc[iy]);
-                    //acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(m8), acc[iy]);
 #endif
                 }
             }
@@ -3057,7 +3049,7 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
                     sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
                     float d8 = q8.scale(iy, ibl);
-                    float m8 = d8 * (q8.y[iy][ibl].bsums[2*ib+0] + q8.y[iy][ibl].bsums[2*ib+1]);
+                    float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
                     acc[2*iy+0] = _mm512_fmadd_ps(_mm512_mul_ps(scales, _mm512_set1_ps(d8)), _mm512_cvtepi32_ps(sumi), acc[2*iy+0]);
                     acc[2*iy+1] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(m8), acc[2*iy+1]);
                 }
@@ -5074,7 +5066,7 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) {
             mm.funcs[5] = mul_mat_iq4_xs_r4_q8_k<6>;
             mm.funcs[6] = mul_mat_iq4_xs_r4_q8_k<7>;
             mm.funcs[7] = mul_mat_iq4_xs_r4_q8_k<8>;
-            expected_typeB = GGML_TYPE_Q8_K;
+            expected_typeB = GGML_TYPE_Q8_K32;
             break;
         case GGML_TYPE_Q4_0_R4:
             assert (ne00 % QK4_NL == 0);
diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index 0ea608bc..66cf32bc 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -2469,8 +2469,8 @@ size_t quantize_iq6_k(const float * src, void * dst, int64_t nrows, int64_t n_pe
     return nrows * nblock * sizeof(block_iq6_k);
 }
 
-
-void iqk_quantize_row_q8_K(const float * x, void * vy, int64_t k) {
+template <bool is_K32>
+void iqk_quantize_row_q8_K_T(const float * x, void * vy, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
     block_q8_K * y = (block_q8_K *)vy;
@@ -2505,8 +2505,14 @@ void iqk_quantize_row_q8_K(const float * x, void * vy, int64_t k) {
             __m256i i1 = _mm256_cvtps_epi32(v1);
             __m256i i2 = _mm256_cvtps_epi32(v2);
             __m256i i3 = _mm256_cvtps_epi32(v3);
-            y[i].bsums[2*ib+0] = hsum_i32_8(_mm256_add_epi32(i0, i1));
-            y[i].bsums[2*ib+1] = hsum_i32_8(_mm256_add_epi32(i2, i3));
+            if constexpr (is_K32) {
+                int bsum = hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+                auto bs = (float *)y[i].bsums;
+                bs[ib] = d*bsum;
+            } else {
+                y[i].bsums[2*ib+0] = hsum_i32_8(_mm256_add_epi32(i0, i1));
+                y[i].bsums[2*ib+1] = hsum_i32_8(_mm256_add_epi32(i2, i3));
+            }
             i0 = _mm256_packs_epi32( i0, i1 );
             i2 = _mm256_packs_epi32( i2, i3 );
             i0 = _mm256_packs_epi16( i0, i2 );
@@ -2539,12 +2545,24 @@ void iqk_quantize_row_q8_K(const float * x, void * vy, int64_t k) {
             int v = nearest_int(iscale*x[j]);
             y[i].qs[j] = MIN(127, v);
         }
-        for (int j = 0; j < QK_K/16; ++j) {
-            int sum = 0;
-            for (int ii = 0; ii < 16; ++ii) {
-                sum += y[i].qs[j*16 + ii];
+        if constexpr (is_K32) {
+            auto bs = (float *)y[i].bsums;
+            float d = 1/iscale;
+            for (int j = 0; j < QK_K/32; ++j) {
+                int sum = 0;
+                for (int ii = 0; ii < 32; ++ii) {
+                    sum += y[i].qs[j*32 + ii];
+                }
+                bs[j] = d*sum;
+            }
+        } else {
+            for (int j = 0; j < QK_K/16; ++j) {
+                int sum = 0;
+                for (int ii = 0; ii < 16; ++ii) {
+                    sum += y[i].qs[j*16 + ii];
+                }
+                y[i].bsums[j] = sum;
             }
-            y[i].bsums[j] = sum;
         }
         y[i].d = 1/iscale;
         x += QK_K;
@@ -2553,6 +2571,14 @@ void iqk_quantize_row_q8_K(const float * x, void * vy, int64_t k) {
 
 }
 
+void iqk_quantize_row_q8_K(const float * x, void * vy, int64_t k) {
+    iqk_quantize_row_q8_K_T<false>(x, vy, k);
+}
+
+void quantize_row_q8_K32(const float * x, void * vy, int64_t k) {
+    iqk_quantize_row_q8_K_T<true>(x, vy, k);
+}
+
 namespace {
 static void quantize_row_iq4_k_impl_bs128(const int super_block_size, const int block_size,
         int n_per_row, const float * x, char * cy,
diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h
index f5bbd3dc..c370c8e8 100644
--- a/ggml/src/iqk/iqk_quantize.h
+++ b/ggml/src/iqk/iqk_quantize.h
@@ -61,8 +61,6 @@ size_t quantize_iq2_ks(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst
 void   dequantize_row_iq2_ks(const block_iq2_ks  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 void   vec_dot_iq2_ks_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
-void iqk_quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
-
 void   quantize_row_iq4_nl_r4_ref(const float * GGML_RESTRICT x, block_iq4_nl_r4  * GGML_RESTRICT y, int64_t k);
 void   quantize_row_iq4_nl_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 size_t quantize_iq4_nl_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
@@ -111,9 +109,11 @@ void   dequantize_row_iq2_bn_r4(const block_iq2_bn * GGML_RESTRICT x, float * GG
 size_t quantize_iq2_bn_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 void   vec_dot_iq2_bn_r4_q8_K64(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+void iqk_quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void quantize_row_q8_K64_ref(const float * GGML_RESTRICT x, block_q8_K64 * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_K64(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_K16(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K32(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 #ifdef __cplusplus
 }