whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -69,6 +69,10 @@
 #define VECTOR_REGISTERS 16
 #endif
 
+#if defined(__riscv_v_intrinsic)
+#define LMUL 4
+#endif
+
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
 namespace {
@@ -176,6 +180,46 @@ inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <>
+inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
+    return __riscv_vfmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
+    return __riscv_vfmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
+    return __riscv_vfmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
+    return __riscv_vfmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_zvfbfwma)
+inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM
 
@@ -228,6 +272,25 @@ inline float hsum(__m512 x) {
 }
 #endif // __AVX512F__
 
+#if defined(__riscv_zvfh)
+inline float hsum(vfloat32m1_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m1_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m1()));
+}
+inline float hsum(vfloat32m2_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m2_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m2()));
+}
+inline float hsum(vfloat32m4_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m4_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m4()));
+}
+inline float hsum(vfloat32m8_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m8_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m8()));
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED MEMORY LOADING
 
@@ -316,6 +379,88 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
+}
+template <> inline vfloat32m1_t load(const float *p) {
+    return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t load(const float *p) {
+    return __riscv_vle32_v_f32m2(p, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t load(const float *p) {
+    return __riscv_vle32_v_f32m4(p, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t load(const float *p) {
+    return __riscv_vle32_v_f32m8(p, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_zvfbfwma)
+template <> inline vbfloat16mf2_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16mf2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vbfloat16m1_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m1(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vbfloat16m2_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m2());
+}
+#endif
+
+#if defined(__riscv_zvfh)
+template <typename T> T set_zero();
+
+template <> inline vfloat16mf2_t set_zero() {
+    return __riscv_vfmv_v_f_f16mf2(0, __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t set_zero() {
+    return __riscv_vfmv_v_f_f16m1(0, __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t set_zero() {
+    return __riscv_vfmv_v_f_f16m2(0, __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t set_zero() {
+    return __riscv_vfmv_v_f_f16m4(0, __riscv_vsetvlmax_e16m4());
+}
+template <> inline vfloat32m1_t set_zero() {
+    return __riscv_vfmv_v_f_f32m1(0.0f, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t set_zero() {
+    return __riscv_vfmv_v_f_f32m2(0, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t set_zero() {
+    return __riscv_vfmv_v_f_f32m4(0, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t set_zero() {
+    return __riscv_vfmv_v_f_f32m8(0, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_v_intrinsic)
+template <typename T> size_t vlmax() {
+    if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    else if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
+    else if constexpr (std::is_same_v<T, vfloat32m2_t>) { return  __riscv_vsetvlmax_e32m2(); }
+    else if constexpr (std::is_same_v<T, vfloat32m4_t>) { return  __riscv_vsetvlmax_e32m4(); }
+    else if constexpr (std::is_same_v<T, vfloat32m8_t>) { return  __riscv_vsetvlmax_e32m8(); }
+    return 0;
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -388,7 +533,7 @@ class tinyBLAS {
         if constexpr (RN > 1) {
             return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
         } else {
-            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
             GGML_ASSERT(false); // we have miss something.
         }
     }
@@ -489,6 +634,573 @@ class tinyBLAS {
     const int64_t ldc;
 };
 
+#if defined(__riscv_v_intrinsic)
+template <typename D, typename V, typename TA, typename TB, typename TC>
+class tinyBLAS_RVV {
+  public:
+    tinyBLAS_RVV(const ggml_compute_params * params, int64_t k,
+             const TA *A, int64_t lda,
+             const TB *B, int64_t ldb,
+             TC *C, int64_t ldc)
+        : params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
+    }
+
+    bool matmul(int64_t m, int64_t n) {
+        if (k % vlmax<V>() != 0) {
+            return false;
+        }
+
+#if LMUL == 1
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 4>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 2>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 1>(m, n, SIZE_N, 12);
+            return true;
+        }
+#elif LMUL == 2
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 4>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 2>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 1>(m, n, SIZE_N, 24);
+            return true;
+        }
+#else // LMUL = 4
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 8>(m, n, SIZE_N, 36);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 4>(m, n, SIZE_N, 36);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 2>(m, n, SIZE_N, 36);
+            return true;
+        }
+#endif
+        return false;
+    }
+
+  private:
+    template<int RM, int RN, int BM>
+    inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
+        if (SIZE_N == RN) {
+            return gemm<RM, RN, BM>(m, n, BN);
+        }
+        if constexpr (RN > 1) {
+            return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
+        } else {
+            GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
+            GGML_ASSERT(false); // we have miss something.
+        }
+    }
+
+    inline void gemm_bloc_4x6(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+        D Cv40 = set_zero<D>();
+        D Cv41 = set_zero<D>();
+        D Cv42 = set_zero<D>();
+        D Cv43 = set_zero<D>();
+        D Cv50 = set_zero<D>();
+        D Cv51 = set_zero<D>();
+        D Cv52 = set_zero<D>();
+        D Cv53 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            V Bv4 = load<V>(B + ldb * (jj + 4) + l);
+            V Bv5 = load<V>(B + ldb * (jj + 5) + l);
+
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv40 = madd(Av0, Bv4, Cv40);
+            Cv50 = madd(Av0, Bv5, Cv50);
+
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv41 = madd(Av1, Bv4, Cv41);
+            Cv51 = madd(Av1, Bv5, Cv51);
+
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv42 = madd(Av2, Bv4, Cv42);
+            Cv52 = madd(Av2, Bv5, Cv52);
+
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+            Cv03 = madd(Av3, Bv0, Cv03);
+            Cv13 = madd(Av3, Bv1, Cv13);
+            Cv23 = madd(Av3, Bv2, Cv23);
+            Cv33 = madd(Av3, Bv3, Cv33);
+            Cv43 = madd(Av3, Bv4, Cv43);
+            Cv53 = madd(Av3, Bv5, Cv53);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+        C[ldc * (jj + 4) + (ii + 0)] = hsum(Cv40);
+        C[ldc * (jj + 4) + (ii + 1)] = hsum(Cv41);
+        C[ldc * (jj + 4) + (ii + 2)] = hsum(Cv42);
+        C[ldc * (jj + 4) + (ii + 3)] = hsum(Cv43);
+        C[ldc * (jj + 5) + (ii + 0)] = hsum(Cv50);
+        C[ldc * (jj + 5) + (ii + 1)] = hsum(Cv51);
+        C[ldc * (jj + 5) + (ii + 2)] = hsum(Cv52);
+        C[ldc * (jj + 5) + (ii + 3)] = hsum(Cv53);
+    }
+
+    inline void gemm_bloc_4x5(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+        D Cv40 = set_zero<D>();
+        D Cv41 = set_zero<D>();
+        D Cv42 = set_zero<D>();
+        D Cv43 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            V Bv4 = load<V>(B + ldb * (jj + 4) + l);
+
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv40 = madd(Av0, Bv4, Cv40);
+
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv41 = madd(Av1, Bv4, Cv41);
+
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv42 = madd(Av2, Bv4, Cv42);
+
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+            Cv03 = madd(Av3, Bv0, Cv03);
+            Cv13 = madd(Av3, Bv1, Cv13);
+            Cv23 = madd(Av3, Bv2, Cv23);
+            Cv33 = madd(Av3, Bv3, Cv33);
+            Cv43 = madd(Av3, Bv4, Cv43);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+        C[ldc * (jj + 4) + (ii + 0)] = hsum(Cv40);
+        C[ldc * (jj + 4) + (ii + 1)] = hsum(Cv41);
+        C[ldc * (jj + 4) + (ii + 2)] = hsum(Cv42);
+        C[ldc * (jj + 4) + (ii + 3)] = hsum(Cv43);
+    }
+
+    inline void gemm_bloc_4x4(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv23 = madd(Av3, Bv2, Cv23);
+
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv33 = madd(Av3, Bv3, Cv33);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+    }
+
+    inline void gemm_bloc_4x3(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv23 = madd(Av3, Bv2, Cv23);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+    }
+
+    inline void gemm_bloc_4x2(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+    }
+
+    inline void gemm_bloc_4x1(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+    }
+
+    inline void gemm_bloc_2x2(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+    }
+
+    inline void gemm_bloc_2x1(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+    }
+
+    template <int RM, int RN>
+    inline void gemm_bloc(int64_t ii, int64_t jj) {
+        if constexpr (RM == 4) {
+            if constexpr (RN == 6) { return gemm_bloc_4x6(ii, jj); }
+            if constexpr (RN == 5) { return gemm_bloc_4x5(ii, jj); }
+            if constexpr (RN == 4) { return gemm_bloc_4x4(ii, jj); }
+            if constexpr (RN == 3) { return gemm_bloc_4x3(ii, jj); }
+            if constexpr (RN == 2) { return gemm_bloc_4x2(ii, jj); }
+            if constexpr (RN == 1) { return gemm_bloc_4x1(ii, jj); }
+        } else if constexpr (RM == 2) {
+            if constexpr (RN == 2) { return gemm_bloc_2x2(ii, jj); }
+            if constexpr (RN == 1) { return gemm_bloc_2x1(ii, jj); }
+        }
+    }
+
+    template <int RM, int RN, int BM>
+    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
+        GGML_ASSERT(m % (RM * BM) == 0);
+        const int64_t ytiles = m / (RM * BM);
+        const int64_t xtiles = (n + RN -1) / RN;
+        const int64_t jj_RN = (xtiles - (xtiles * RN - n));
+
+        // "round" bloc_size to "nearest" BN
+        const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
+        const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
+        const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
+        const int64_t nb_job = ytiles * NB_BN;
+
+        if (params->ith == 0) {
+            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        int64_t job = params->ith;
+        while (job < nb_job) {
+            const int64_t ii = (job % ytiles) * RM * BM;
+            const int64_t jb =  job / ytiles;
+            const int64_t jr0 = BLOC_POS(jb  , jj_BN, SIZE_BN);
+            const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
+
+            const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
+            const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
+            const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
+
+            for (int64_t bi = 0; bi < BM * RM; bi += RM) {
+                int64_t jj = jj0;
+                for (; jj < jj1; jj += RN) {
+                    gemm_bloc<RM, RN>(ii + bi, jj);
+                }
+                if constexpr (RN > 1) {
+                    for (; jj < jj2; jj += RN - 1) {
+                        gemm_bloc<RM, RN-1>(ii + bi, jj);
+                    }
+                }
+                GGML_ASSERT(jj == jj2);
+            }
+
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
+
+        ggml_barrier(params->threadpool);
+        return;
+    }
+
+    const ggml_compute_params * params;
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+};
+#endif
+
 //////////////////////////////////////////////////////////////////////////////////////////
 // QUANT ZERO MATRIX MULTIPLICATION
 
@@ -1086,10 +1798,27 @@ class tinyBLAS_Q0_AVX {
       } \
    } \
 
+template<typename T>
+struct mma_instr;
+
+template<>
+struct mma_instr<ggml_bf16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvbf16ger2pp(acc, a, b);
+    }
+};
+
+template<>
+struct mma_instr<ggml_fp16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvf16ger2pp(acc, a, b);
+    }
+};
+
 template <typename TA, typename TB, typename TC>
-class tinyBLAS_BF16_PPC {
+class tinyBLAS_HP16_PPC {
   public:
-    tinyBLAS_BF16_PPC(int64_t k,
+    tinyBLAS_HP16_PPC(int64_t k,
                 const TA *A, int64_t lda,
                 const TB *B, int64_t ldb,
                 TC *C, int64_t ldc,
@@ -1407,8 +2136,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -1424,8 +2153,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x+4], vec_B[x]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -1444,10 +2173,10 @@ class tinyBLAS_BF16_PPC {
             packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
             packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
-                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_2, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_3, vec_A[x+4], vec_B[x+4]);
             }
         }
 
@@ -1478,7 +2207,7 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
                 for (int x = 0; x<2; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -1513,8 +2242,8 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
                 for (int x = 0; x<4; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -1577,44 +2306,91 @@ template <typename TA>
 class tinyBLAS_Q0_PPC {
   public:
     tinyBLAS_Q0_PPC(int64_t k,
-                const TA *A, int64_t lda,
-                const block_q8_0 *B, int64_t ldb,
-                float *C, int64_t ldc,
-                int ith, int nth)
+             const TA * A, int64_t lda,
+             const block_q8_0 * B, int64_t ldb,
+             float * C, int64_t ldc,
+             int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
     }
 
     void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
+        const int64_t mc = 64;
+        const int64_t kc = 64;
+        int64_t nc = 64;
+        int64_t n_aligned = 0;
+        if (n % 64 == 0) {
+            n_aligned = n;
+        } else if (n == 4) {
+            n_aligned = 4;
+        } else if (n < 64) {
+            n_aligned = (n / 8) * 8;
+        } else {
+            n_aligned = (n / 64) * 64;
+        }
+
+        if (n_aligned > 0) {
+            if (n_aligned % 64 == 0)      nc = 64;
+            else if (n_aligned == n)      nc = n;
+            else if (n_aligned % 32 == 0) nc = 32;
+            else if (n_aligned % 24 == 0) nc = 24;
+            else if (n_aligned % 16 == 0) nc = 16;
+            else                          nc = 8;
+        }
+        bool can_use_tiled = n_aligned > 0 && (m % mc == 0) && (k % kc == 0);
+        if (can_use_tiled) {
+            matmul_tiled(m, n_aligned, mc, nc, kc);
+            if (n > n_aligned) {
+                mnpack(0, m, n_aligned, n);
+            }
+        } else {
+            mnpack(0, m, 0, n);
+        }
     }
 
   private:
+    inline void save_res(int ii, int jj, int idx, vector float * fin_res, int RM = 4, int RN = 4) {
+        for (int I = 0; I < RM; I++) {
+            for (int J = 0; J < RN; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&fin_res[idx + I] + J);
+            }
+        }
+    }
 
-    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-       for (int I = 0; I < RM; I++) {
-          for (int J = 0; J < RN; J++) {
-             *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
-          }
-       }
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&vec_C[I] + J);
+            }
+        }
     }
 
-    template<int size>
-    inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
-       vector signed int vec_C[4];
-       vector float CA[4] = {0};
-       vector float res[4] = {0};
-       __builtin_mma_disassemble_acc(vec_C, ACC);
-       for (int i = 0; i < 4; i++) {
-          CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
-          res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
-          fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
-       }
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C + ii+ ((jj + J) * ldc) + I);
+                *c_ptr += *((float *)&vec_C[I] + J);
+            }
+        }
     }
-    /* This function processes quantized data from block_q4_0 elements.
-     * First the we try to extract the two int4 values stored in single int8_t into two signed int8.
-     * And then we subtract each of the resultant element with 8, to convert signed int8 to unsigned int8.
-     * Also compute the rowsum which is required to compensate the above conversion. */
-    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
+
+    template<typename ArrayType>
+    inline void compute(acc_t * ACC, int c_idx, int s_idx, ArrayType & comparray, vector float * vs, vector float * fin_res) {
+        vector signed int vec_C[4];
+        vector float CA[4] = {0};
+        vector float res[4] = {0};
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int i = 0; i < 4; i++) {
+            CA[i] = vec_splats((float)(((double)comparray[c_idx + i]) * -128.0));
+            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
+            fin_res[s_idx + i] = vec_madd(res[i], vs[s_idx + i], fin_res[s_idx + i]);
+        }
+    }
+
+    inline void process_q4_elements(vector signed char (&c)[2], int * ca) {
         const vector signed char lowMask = vec_splats((signed char)0xF);
         const vector unsigned char v4 = vec_splats((unsigned char)0x4);
         const vector signed char v8 = vec_splats((signed char)0x8);
@@ -1631,7 +2407,7 @@ class tinyBLAS_Q0_PPC {
     }
 
     template <typename V1, typename V2>
-    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
+    inline void vector_permute_store(V2 & s1, V2 & s2, V2 & s3, V2 & s4, V1 * vecOffset, bool flip) {
         vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
         vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
         vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
@@ -1655,21 +2431,170 @@ class tinyBLAS_Q0_PPC {
             t8 = vec_xor(t8, xor_vector);
         }
         vec_xst(t5, 0, vecOffset);
-        vec_xst(t6, 0, vecOffset+16);
-        vec_xst(t7, 0, vecOffset+32);
-        vec_xst(t8, 0, vecOffset+48);
+        vec_xst(t6, 0, vecOffset + 16);
+        vec_xst(t7, 0, vecOffset + 32);
+        vec_xst(t8, 0, vecOffset + 48);
+    }
+
+    inline void unpack_q4_to_q8(vector signed char packed, vector signed char & lo, vector signed char & hi) {
+        const vector signed char lowMask = vec_splats((signed char)0x0F);
+        const vector signed char v8      = vec_splats((signed char)0x08);
+        const vector unsigned char v4    = vec_splats((unsigned char)4);
+        lo = vec_and(packed, lowMask);
+        hi = vec_sr(packed, v4);
+        lo = vec_sub(lo, v8);
+        hi = vec_sub(hi, v8);
+    }
+
+    inline void vector_permute_store_fp16(vec_t * c, unsigned char * vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        for (int i = 0; i < 4; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        for (int i = 4; i < 8; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        s[0] = vec_perm(t[0], t[2], swiz3);
+        s[1] = vec_perm(t[0], t[2], swiz4);
+        s[2] = vec_perm(t[1], t[3], swiz3);
+        s[3] = vec_perm(t[1], t[3], swiz4);
+        s[4] = vec_perm(t[4], t[6], swiz3);
+        s[5] = vec_perm(t[4], t[6], swiz4);
+        s[6] = vec_perm(t[5], t[7], swiz3);
+        s[7] = vec_perm(t[5], t[7], swiz4);
+        for (int i = 0; i < 8; ++i) {
+            vec_xst(s[i], 0, (vec_t *)(vecOffset + i * 16));
+        }
+    }
+
+    static inline void convert_and_scale_q8(vector signed char raw, vector float v_scale, vector unsigned short & out_hi, vector unsigned short & out_lo) {
+        vector signed short i16_hi = vec_unpackh(raw);
+        vector signed short i16_lo = vec_unpackl(raw);
+
+        vector float f_hi_h = vec_ctf(vec_unpackh(i16_hi), 0);
+        vector float f_hi_l = vec_ctf(vec_unpackl(i16_hi), 0);
+        vector float f_lo_h = vec_ctf(vec_unpackh(i16_lo), 0);
+        vector float f_lo_l = vec_ctf(vec_unpackl(i16_lo), 0);
+        out_hi = vec_pack_to_short_fp32(vec_mul(f_hi_h, v_scale), vec_mul(f_hi_l, v_scale));
+        out_lo = vec_pack_to_short_fp32(vec_mul(f_lo_h, v_scale), vec_mul(f_lo_l, v_scale));
+    }
+
+    void packNormal_q4_fp16(const block_q4_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        for (int i = 0; i < rows; i += 8) {
+            const block_q4_0 * rows_base[8];
+            for (int r = 0; r < 8; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[8][4];
+                for (int r = 0; r < 8; r++) {
+                    const block_q4_0 * current_blk = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(current_blk->d));
+                    vector signed char v_qs = vec_xl(0, (const vector signed char *)current_blk->qs);
+                    vector signed char c1, c2;
+                    unpack_q4_to_q8(v_qs, c1, c2);
+                    convert_and_scale_q8(c1, v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c2, v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int c = 0; c < 4; c++) {
+                    vector unsigned char c_arr[8];
+                    for (int r = 0; r < 8; r++) {
+                        c_arr[r] = (vector unsigned char)hp_res[r][c];
+                    }
+                    vector_permute_store_fp16((vec_t *)c_arr, vecOffset);
+                    vecOffset += 128;
+                }
+            }
+        }
+    }
+
+    template <int chunk_size>
+    static inline void pack_q8_block(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        const vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        const vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        const vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        const vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        for (int i = 0; i < rows; i += chunk_size) {
+            const block_q8_0 * rows_base[chunk_size];
+            for (int r = 0; r < chunk_size; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[chunk_size][4];
+                for (int r = 0; r < chunk_size; r++) {
+                    const block_q8_0 * b = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(b->d));
+                    vector signed char c[2];
+                    __vector_pair pair = __builtin_vsx_lxvp(0, (__vector_pair *)b->qs);
+                    __builtin_vsx_disassemble_pair(c, & pair);
+                    convert_and_scale_q8(c[0], v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c[1], v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int col = 0; col < 4; col++) {
+                    if constexpr (chunk_size == 8) {
+                        vec_t t[8];
+                        t[0] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        t[1] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        t[2] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        t[3] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+                        t[4] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz1);
+                        t[5] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz2);
+                        t[6] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz1);
+                        t[7] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz2);
+
+                        vec_xst(vec_perm(t[0], t[2], swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t[0], t[2], swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t[1], t[3], swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t[1], t[3], swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vec_xst(vec_perm(t[4], t[6], swiz3), 0, (vec_t *)(vecOffset + 64));
+                        vec_xst(vec_perm(t[4], t[6], swiz4), 0, (vec_t *)(vecOffset + 80));
+                        vec_xst(vec_perm(t[5], t[7], swiz3), 0, (vec_t *)(vecOffset + 96));
+                        vec_xst(vec_perm(t[5], t[7], swiz4), 0, (vec_t *)(vecOffset + 112));
+                        vecOffset += 128;
+                    } else {
+                        vec_t t0 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        vec_t t1 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        vec_t t2 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        vec_t t3 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+
+                        vec_xst(vec_perm(t0, t2, swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t0, t2, swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t1, t3, swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t1, t3, swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vecOffset += 64;
+                    }
+                }
+            }
+        }
+    }
+
+    void packNormal_q8_fp16(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        if (rows == 4) {
+            pack_q8_block<4>(a, lda, rows, blocks, vec);
+        } else {
+            pack_q8_block<8>(a, lda, rows, blocks, vec);
+        }
     }
 
     template<int size>
-    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
+    void packNormalInt4(const TA * a, int64_t lda, int rows, int cols, int8_t * vec, std::array<int, size> & comparray) {
         int64_t i, j;
-        TA *aoffset = NULL;
-        int8_t *vecOffset = NULL;
-        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
-        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        TA * aoffset = NULL;
+        int8_t * vecOffset = NULL;
+        TA * aoffset1 = NULL, * aoffset2 = NULL, * aoffset3 = NULL, * aoffset4 = NULL;
+        TA * aoffset5 = NULL, * aoffset6 = NULL, * aoffset7 = NULL, * aoffset8 = NULL;
         vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
         vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
-        aoffset = const_cast<TA*>(a);
+        aoffset = const_cast<TA *>(a);
         vecOffset = vec;
         j = (rows >> 3);
         if (j > 0) {
@@ -1686,27 +2611,27 @@ class tinyBLAS_Q0_PPC {
                 i = (cols >> 2);
                 if (i > 0) {
                     do {
-                        c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
-                        c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                        c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                        c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));
-                        c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset5->qs));
-                        c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset6->qs));
-                        c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset7->qs));
-                        c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset8->qs));
-
-                        process_q4_elements(c1, &comparray[0]);
-                        process_q4_elements(c2, &comparray[1]);
-                        process_q4_elements(c3, &comparray[2]);
-                        process_q4_elements(c4, &comparray[3]);
-                        process_q4_elements(c5, &comparray[4]);
-                        process_q4_elements(c6, &comparray[5]);
-                        process_q4_elements(c7, &comparray[6]);
-                        process_q4_elements(c8, &comparray[7]);
+                        c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
+                        c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                        c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                        c4[1] = vec_xl(0, (const vector signed char *)aoffset4->qs);
+                        c5[1] = vec_xl(0, (const vector signed char *)aoffset5->qs);
+                        c6[1] = vec_xl(0, (const vector signed char *)aoffset6->qs);
+                        c7[1] = vec_xl(0, (const vector signed char *)aoffset7->qs);
+                        c8[1] = vec_xl(0, (const vector signed char *)aoffset8->qs);
+
+                        process_q4_elements(c1, & comparray[0]);
+                        process_q4_elements(c2, & comparray[1]);
+                        process_q4_elements(c3, & comparray[2]);
+                        process_q4_elements(c4, & comparray[3]);
+                        process_q4_elements(c5, & comparray[4]);
+                        process_q4_elements(c6, & comparray[5]);
+                        process_q4_elements(c7, & comparray[6]);
+                        process_q4_elements(c8, & comparray[7]);
                         vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
+                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset + 128, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset + 192, false);
                         aoffset1 += lda;
                         aoffset2 += lda;
                         aoffset3 += lda;
@@ -1732,17 +2657,17 @@ class tinyBLAS_Q0_PPC {
             i = (cols >> 2);
             if (i > 0) {
                 do {
-                    c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
-                    c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));
-
-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
+                    c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                    c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                    c4[1] = vec_xl(0, (const vector signed char *)aoffset4->qs);
+
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                     vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                     aoffset1 += lda;
                     aoffset2 += lda;
                     aoffset3 += lda;
@@ -1761,17 +2686,17 @@ class tinyBLAS_Q0_PPC {
             if (i > 0) {
                 do {
                     switch(rows) {
-                        case 3: c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
-                        case 2: c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset2->qs));
-                        case 1: c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
+                        case 3: c3[1] = vec_xl(0, (const vector signed char *)aoffset3->qs);
+                        case 2: c2[1] = vec_xl(0, (const vector signed char *)aoffset2->qs);
+                        case 1: c1[1] = vec_xl(0, (const vector signed char *)aoffset1->qs);
                             break;
                     }
-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                     vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                     aoffset1 += lda;
                     aoffset2 += lda;
                     aoffset3 += lda;
@@ -1781,38 +2706,39 @@ class tinyBLAS_Q0_PPC {
             }
         }
     }
+
     template<typename VA, typename VB>
-    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    void packNormal(const block_q8_0 * a, int64_t lda, int rows, int cols, VA * vec, bool flip) {
         int64_t i, j;
-        block_q8_0 *aoffset = NULL;
-        VA *vecOffset = NULL;
-        block_q8_0* aoffsets[8];
+        block_q8_0 * aoffset = NULL;
+        VA * vecOffset = NULL;
+        block_q8_0 * aoffsets[8];
         __vector_pair arr[8];
         VB c[8][2] = {0};
         VB c1[8] = {0}; VB c2[8] = {0};
-        aoffset = const_cast<block_q8_0*>(a);
+        aoffset = const_cast<block_q8_0 *>(a);
         vecOffset = vec;
         j = (rows >> 3);
         if (j > 0) {
             do {
                 aoffsets[0] = aoffset;
                 for (int it = 1; it < 8; it++)
-                    aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffsets[it] = aoffsets[it - 1] + lda;
                 aoffset += 8 * lda;
 
                 i = (cols >> 3);
                 if (i > 0) {
                 do {
                     for (int it = 0; it < 8; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                         c1[it] = c[it][0];
                         c2[it] = c[it][1];
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
-                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
-                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
+                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset + 128, flip);
+                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset + 192, flip);
                     for (int it = 0; it < 8; it++)
                         aoffsets[it] += lda;
                     vecOffset += 256;
@@ -1822,7 +2748,6 @@ class tinyBLAS_Q0_PPC {
             j--;
         } while(j > 0);
     }
-
     if (rows & 4) {
             aoffsets[0]  = aoffset;
             for (int it = 1; it < 4; it++ )
@@ -1832,13 +2757,13 @@ class tinyBLAS_Q0_PPC {
             if (i > 0) {
                do {
                     for (int it = 0; it < 4; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                         c1[it] = c[it][0];
                         c2[it] = c[it][1];
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                     for (int it = 0; it < 4; it++) {
                         aoffsets[it] += lda;
                     }
@@ -1851,24 +2776,24 @@ class tinyBLAS_Q0_PPC {
         if (rows & 3) {
             aoffsets[0]  = aoffset;
             for (int it = 1; it < 3; it++ )
-                aoffsets[it] = aoffsets[it-1] + lda;
+                aoffsets[it] = aoffsets[it - 1] + lda;
             i = (cols >> 3);
             if (i > 0) {
                 do {
                     switch(rows) {
-                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[2]->qs);
-                                __builtin_vsx_disassemble_pair(c[2], &arr[2]);
+                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[2]->qs);
+                                __builtin_vsx_disassemble_pair(c[2], & arr[2]);
                                 c1[2] = c[2][0]; c2[2] = c[2][1];
-                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[1]->qs);
-                                __builtin_vsx_disassemble_pair(c[1], &arr[1]);
+                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[1]->qs);
+                                __builtin_vsx_disassemble_pair(c[1], & arr[1]);
                                 c1[1] = c[1][0]; c2[1] = c[1][1];
-                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[0]->qs);
-                                __builtin_vsx_disassemble_pair(c[0], &arr[0]);
+                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[0]->qs);
+                                __builtin_vsx_disassemble_pair(c[0], & arr[0]);
                                 c1[0] = c[0][0]; c2[0] = c[0][1];
                                 break;
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                     for (int it = 0; it < 3; it++)
                          aoffsets[it] += lda;
                     vecOffset += 128;
@@ -1923,26 +2848,26 @@ class tinyBLAS_Q0_PPC {
         vector float vs[8] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<4>((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<4>((A + (ii * lda) + l), lda, 4, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 4, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x], vec_B[x+8]);
             }
             for (int I = 0; I<4; I++) {
                 for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 4] + J) = (unhalf((A +((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 4; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -1953,11 +2878,11 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 0, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii, jj+4, 4, fin_res);
+        save_res(ii, jj + 4, 4, fin_res);
     }
 
     void KERNEL_8x4(int64_t ii, int64_t jj) {
@@ -1968,25 +2893,25 @@ class tinyBLAS_Q0_PPC {
         vector float vs[8] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 4, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
             }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+            for (int I = 0; I < 8; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 8; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -1997,45 +2922,46 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
     }
 
     void KERNEL_8x8(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[16] = {0};
         acc_t acc_0, acc_1, acc_2, acc_3;
+        acc_t acc_4, acc_5, acc_6, acc_7;
         std::array<int, 8> comparray {};
         vector float fin_res[16] = {0};
         vector float vs[16] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
-            __builtin_mma_xxsetaccz(&acc_2);
-            __builtin_mma_xxsetaccz(&acc_3);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
+            __builtin_mma_xxsetaccz(& acc_2);
+            __builtin_mma_xxsetaccz(& acc_3);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]);
-                __builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_2, vec_A[x], vec_B[x + 8]);
+                __builtin_mma_xvi8ger4pp(& acc_3, vec_A[x + 8], vec_B[x + 8]);
             }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+            for (int I = 0; I < 8 ; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 8] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 8; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -2046,15 +2972,96 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
-            compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
-            compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_2, 0, 8, comparray, vs, fin_res);
+            compute(& acc_3, 4, 12, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
-        save_res(ii, jj+4, 8, fin_res);
-        save_res(ii+4, jj+4, 12, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
+        save_res(ii, jj + 4, 8, fin_res);
+        save_res(ii + 4, jj + 4, 12, fin_res);
+    }
+
+    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t * vec_A, vec_t * vec_B) {
+        acc_t acc[8];
+        for (int i = 0; i < mc ; i += 16) {
+            for (int j = 0; j < nc; j += 8) {
+                int A0_base = (i / 16) * (2 * 32 * kc);
+                int B0_base = (j / 8) * (32 * kc);
+                for (int x = 0; x < 8; x++) {
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                }
+                for (int64_t kk = 0; kk < kc; kk++) {
+                    int A0_block_idx = A0_base + kk * 32;
+                    int B0_block_idx = B0_base + kk * 32;
+                    int A1_block_idx = A0_block_idx + 32 * kc;
+                    int B1_block_idx = B0_block_idx + 32 * kc;
+                    vec_t * A0_block = & vec_A[A0_block_idx];
+                    vec_t * B0_block = & vec_B[B0_block_idx];
+                    vec_t * A1_block = & vec_A[A1_block_idx];
+                    for (int it = 0; it < 4; it++) {
+                        for (int x = 0; x < 4; x++) {
+                            __builtin_mma_xvf16ger2pp(& acc[0], A0_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[1], A0_block[8 * it + x], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[2], A0_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[3], A0_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[4], A1_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[5], A1_block[8 * it + x], B0_block[8 * it+ x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[6], A1_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[7], A1_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                        }
+                    }
+                }
+                if (l == 0) {
+                    save_acc(& acc[0], ii + i, jj + j);
+                    save_acc(& acc[1], ii + i, jj + j + 4);
+                    save_acc(& acc[2], ii + i + 4, jj + j);
+                    save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    save_acc(& acc[4], ii + i + 8, jj + j);
+                    save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    save_acc(& acc[6], ii + i + 12, jj + j);
+                    save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                } else {
+                    add_save_acc(& acc[0], ii + i, jj + j);
+                    add_save_acc(& acc[1], ii + i, jj + j + 4);
+                    add_save_acc(& acc[2], ii + i + 4, jj + j);
+                    add_save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    add_save_acc(& acc[4], ii + i + 8, jj + j);
+                    add_save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    add_save_acc(& acc[6], ii + i + 12, jj + j);
+                    add_save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                }
+            }
+        }
+    }
+
+    void matmul_tiled(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        vec_t A_pack[mc * kc * 4];
+        vec_t B_pack[nc * kc * 4];
+        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles  = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                if constexpr(is_Ablock_q4) {
+                    packNormal_q4_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                } else {
+                    packNormal_q8_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                }
+                packNormal_q8_fp16(B + jj * ldb + kk, ldb, nc, kc, (uint8_t *)B_pack);
+                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack);
+            }
+        }
     }
 
     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
@@ -2079,32 +3086,32 @@ class tinyBLAS_Q0_PPC {
             vector float fin_res[4] = {0};
             vector float vs[4] = {0};
             vector float CA[4] = {0};
-            __builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value
-            __builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((A + (ii * lda) + 0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((B + (jj * ldb) + 0)->qs, 0, 1); // prefetch first value
             for (int l = 0; l < k; l++) {
-                __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_mma_xxsetaccz(&acc_0);
+                __builtin_prefetch((A + (ii * lda) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_prefetch((B + (jj * ldb) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_mma_xxsetaccz(& acc_0);
                 if (isAblock_q4) {
-                   packNormalInt4<4>((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray);
+                    packNormalInt4<4>((A + (ii * lda) + l), lda, RM, 4, (int8_t *)vec_A, comparray);
                 } else {
-                   packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                    packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, RM, 8, (int8_t *)vec_A, false);
                 }
-                packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
-                for(int x = 0; x < 8; x+=4) {
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]);
+                packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, RN, 8, (uint8_t *)vec_B, true);
+                for (int x = 0; x < 8; x += 4) {
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 1], vec_B[x + 1]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 2], vec_B[x + 2]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 3], vec_B[x + 3]);
                 }
-                for (int I = 0; I<RM; I++) {
-                    for (int J = 0; J<RN; J++) {
-                        *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+                for (int I = 0; I < RM; I++) {
+                    for (int J = 0; J < RN; J++) {
+                        *((float*)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                     }
                 }
-                __builtin_mma_disassemble_acc(vec_C, &acc_0);
+                __builtin_mma_disassemble_acc(vec_C, & acc_0);
                 if (!isAblock_q4) {
-                    auto aoffset = A+(ii*lda)+l;
+                    auto aoffset = A + (ii * lda) + l;
                     for (int i = 0; i < RM; i++) {
                         comparray[i] = 0;
                         int ca = 0;
@@ -2127,15 +3134,15 @@ class tinyBLAS_Q0_PPC {
 
     template<int RM, int RN>
     inline void kernel(int64_t ii, int64_t jj) {
-       if constexpr(RM == 4 && RN == 8) {
-          KERNEL_4x8(ii,jj);
-       } else if constexpr(RM == 8 && RN == 4) {
-          KERNEL_8x4(ii,jj);
-       } else if constexpr(RM == 8 && RN == 8) {
-          KERNEL_8x8(ii,jj);
-       } else {
-          assert(false && "RN/RM values not supported");
-       }
+        if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii,jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii,jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii,jj);
+        } else {
+            assert(false && "RN/RM values not supported");
+        }
     }
 
     template <int RM, int RN>
@@ -2154,11 +3161,11 @@ class tinyBLAS_Q0_PPC {
             kernel<RM, RN>(ii, jj);
         }
     }
-
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *C;
+    const TA * const A;
+    const block_q8_0 * const B;
+    float * C;
     const int64_t k;
+    int64_t kc;
     const int64_t lda;
     const int64_t ldb;
     const int64_t ldc;
@@ -2731,6 +3738,24 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
+#elif defined(__riscv_zvfh)
+    #if LMUL == 1
+        tinyBLAS_RVV<vfloat32m1_t, vfloat32m1_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #elif LMUL == 2
+        tinyBLAS_RVV<vfloat32m2_t, vfloat32m2_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #else // LMUL = 4
+        tinyBLAS_RVV<vfloat32m4_t, vfloat32m4_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #endif
+        return tb.matmul(m, n);
 #else
         return false;
 #endif
@@ -2762,17 +3787,38 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             return tb.matmul(m, n);
         }
 #elif defined(__MMA__)
-        if ((k % 8))
-                return false;
-        if(Btype == GGML_TYPE_BF16) {
-           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
-            (const ggml_bf16_t *)A, lda,
-            (const ggml_bf16_t *)B, ldb,
-            (float *)C, ldc,
-            params->ith, params->nth};
-        tb.matmul(m, n);
-        return true;
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS_HP16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
         }
+#elif defined(__riscv_zvfbfwma)
+        #if LMUL == 1
+            tinyBLAS_RVV<vfloat32m1_t, vbfloat16mf2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #elif LMUL == 2
+            tinyBLAS_RVV<vfloat32m2_t, vbfloat16m1_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #else // LMUL = 4
+            tinyBLAS_RVV<vfloat32m4_t, vbfloat16m2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #endif
+            return tb.matmul(m, n);
 #endif
         return false;
     }
@@ -2822,6 +3868,41 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__riscv_zvfh)
+        if (Btype == GGML_TYPE_F16) {
+        #if LMUL == 1
+            tinyBLAS_RVV<vfloat32m1_t, vfloat16mf2_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #elif LMUL == 2
+            tinyBLAS_RVV<vfloat32m2_t, vfloat16m1_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #else // LMUL = 4
+            tinyBLAS_RVV<vfloat32m4_t, vfloat16m2_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #endif
+            return tb.matmul(m, n);
+        }
+#elif defined(__MMA__)
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS_HP16_PPC<ggml_fp16_t, ggml_fp16_t, float> tb{ k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
+        }
 #endif
         return false;
     }