whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp

@@ -1,12 +1,19 @@
 #include "common.hpp"
+#include "ggml-sycl/presets.hpp"
 #include "ggml.h"
 #include "element_wise.hpp"
 
+#define SYCL_GLOBAL_ID_LOOP(K, ITEM) \
+    for (auto i = ITEM.get_global_id(0); i < (size_t)K; i += ITEM.get_global_range(0))
+
+#define SYCL_LOCAL_ID_CALC(ITEM, IDX) \
+    (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))
+
+
 static void acc_f32(const float * x, const float * y, float * dst, const int ne,
     const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset, const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+    const int nb1, const int nb2, int offset, const sycl::nd_item<1> &item_ct1) {
+    const int i = SYCL_LOCAL_ID_CALC(item_ct1, 0);
     if (i >= ne) {
         return;
     }
@@ -21,239 +28,280 @@ static void acc_f32(const float * x, const float * y, float * dst, const int ne,
     }
 }
 
+/* Unary OP funcs */
 template<typename T>
-static void sgn(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = x[i] > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x[i] < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
-    }
+static __dpct_inline__ T op_sgn(T x) {
+    return x > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
 }
 
 template<typename T>
-static void abs_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = sycl::fabs(x[i]);
-    }
+static __dpct_inline__ T op_abs(T x) {
+    return sycl::fabs(x);
 }
 
 template<typename T>
-static void elu_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = (x[i] > static_cast<T>(0.f)) ? x[i] : sycl::expm1(x[i]);
-    }
+static __dpct_inline__ T op_elu(T x) {
+    return (x > static_cast<T>(0.f)) ? x : sycl::expm1(x);
 }
 
 template<typename T>
-static void gelu(const T * x, T * dst, const int k,
-                     const sycl::nd_item<3> &item_ct1) {
+static __dpct_inline__ T op_gelu(T x) {
     const T GELU_COEF_A    = static_cast<T>(0.044715f);
     const T SQRT_2_OVER_PI = static_cast<T>(0.79788456080286535587989211986876f);
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+    return static_cast<T>(0.5f) * x *
+           (static_cast<T>(1.0f) +
+            sycl::tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+}
 
-    if (i >= k) {
-        return;
-    }
+template<typename T>
+static __dpct_inline__ T op_silu(T x) {
+    return x / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+}
 
-    float xi = x[i];
-    dst[i] = static_cast<T>(0.5f) * xi *
-             (static_cast<T>(1.0f) +
-              sycl::tanh(SQRT_2_OVER_PI * xi * (static_cast<T>(1.0f) + GELU_COEF_A * xi * xi)));
+template<typename T>
+static __dpct_inline__ T op_gelu_quick(T x) {
+    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
+    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF_LOCAL * x)));
 }
 
 template<typename T>
-static void silu(const T * x, T * dst, const int k,
-                     const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static __dpct_inline__ T op_gelu_erf(T x) {
+    const T SQRT_2_INV = static_cast<T>(0.70710678118654752440084436210484f);
+    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + sycl::erf(x * SQRT_2_INV));
+}
 
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] / (static_cast<T>(1.0f) + sycl::native::exp(-x[i]));
+template<typename T>
+static __dpct_inline__ T op_tanh(T x) {
+    return sycl::tanh(x);
 }
 
 template<typename T>
-static void gelu_quick(const T *x, T *dst, int k,
-                           const sycl::nd_item<3> &item_ct1) {
-    const float GELU_QUICK_COEF = -1.702f;
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF * x[i])));
+static __dpct_inline__ T op_relu(T x) {
+    return sycl::fmax(x, static_cast<T>(0));
 }
 
 template<typename T>
-static void tanh(const T *x, T *dst, int k,
-                     const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::tanh((x[i]));
+static __dpct_inline__ T op_sigmoid(T x) {
+    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(-x));
 }
 
 template<typename T>
-static void relu(const T * x, T * dst, const int k,
-                     const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static __dpct_inline__ T op_sqrt(T x) {
+    return sycl::sqrt(x);
+}
 
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::fmax((x[i]), static_cast<T>(0));
+template<typename T>
+static __dpct_inline__ T op_sin(T x) {
+    return sycl::sin(x);
 }
 
 template<typename T>
-static void sigmoid(const T * x, T * dst, const int k,
-                            const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static __dpct_inline__ T op_cos(T x) {
+    return sycl::cos(x);
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static __dpct_inline__ T op_hardsigmoid(T x) {
+    return sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_hardswish(T x) {
+    return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_exp(T x) {
+    return sycl::exp(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_log(T x) {
+    if (x <= static_cast<T>(0)) {
+        return neg_infinity<T>();
     }
-    dst[i] = 1.0f / (static_cast<T>(1.0f) + sycl::native::exp(-x[i]));
+    return sycl::log(x);
 }
 
 template<typename T>
-static void sqrt(const T * x, T * dst, const int k,
-                            const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static __dpct_inline__ T op_neg(T x) {
+    return -x;
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static __dpct_inline__ T op_step(T x) {
+    return (x > static_cast<T>(0.0f)) ? static_cast<T>(1.0f) : static_cast<T>(0.0f);
+}
+
+template<typename T>
+static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {
+    T neg_slope_T = static_cast<T>(negative_slope);
+    return sycl::fmax(x, static_cast<T>(0)) +
+           sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+}
+
+template<typename T>
+static __dpct_inline__ T op_sqr(T x) {
+    return x * x;
+}
+
+template<typename T>
+static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {
+    return x < static_cast<T>(min_val) ? static_cast<T>(min_val) : (x > static_cast<T>(max_val) ? static_cast<T>(max_val) : x);
+}
+
+template<typename T>
+static void unary_op_sgn_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sgn(x[i]);
     }
-    dst[i] = sycl::sqrt(x[i]);
 }
 
 template<typename T>
-static void sin(const T * x, T * dst, const int k,
-                            const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_abs_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_abs(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_elu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_elu(x[i]);
     }
-    dst[i] = sycl::sin(x[i]);
 }
 
 template<typename T>
-static void cos(const T * x, T * dst, const int k,
-                            const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_gelu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_silu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_silu(x[i]);
     }
-    dst[i] = sycl::cos(x[i]);
 }
 
 template<typename T>
-static void hardsigmoid(const T * x, T * dst, const int k,
-                            const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_gelu_quick_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu_quick(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_gelu_erf_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu_erf(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_tanh_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_tanh(x[i]);
     }
-    dst[i] = sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x[i] + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
 }
 
 template<typename T>
-static void hardswish(const T * x, T * dst, const int k,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_relu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_relu(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_sigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sigmoid(x[i]);
     }
-    dst[i] = x[i] * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x[i] + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
 }
 
 template<typename T>
-static void exp(const T * x, T * dst, const int k,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_sqrt_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sqrt(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_sin_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sin(x[i]);
     }
-    dst[i] = sycl::exp(x[i]);
 }
 
 template<typename T>
-static void log(const T * x, T * dst, const int k,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_cos_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_cos(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_hardsigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_hardsigmoid(x[i]);
     }
-    T xi = x[i];
-    if (xi <= 0) {
-        dst[i] = neg_infinity<T>();
-    } else {
-        dst[i] = sycl::log(xi);
+}
+
+template<typename T>
+static void unary_op_hardswish_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_hardswish(x[i]);
     }
 }
 
 template<typename T>
-static void neg(const T * x, T * dst, const int k,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_exp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_exp(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_log_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_log(x[i]);
     }
-    dst[i] = -x[i];
 }
 
 template<typename T>
-static void step(const T * x, T * dst, const int k,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_neg_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_neg(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_step_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_step(x[i]);
     }
-    dst[i] = x[i] > static_cast<T>(0.0f);
 }
 
 template<typename T>
-static void leaky_relu(const T *x, T *dst, const int k, const float negative_slope,
-                           const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-    if (i >= k) {
-        return;
+static void unary_op_leaky_relu_kernel(const T * x, T * dst, const int k, float negative_slope, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_leaky_relu(x[i], negative_slope);
     }
-    dst[i] = sycl::fmax((x[i]), static_cast<T>(0)) +
-             sycl::fmin((x[i]), static_cast<T>(0.0f)) * negative_slope;
 }
 
 template<typename T>
-static void sqr(const T * x, T * dst, const int k,
-                    const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+static void unary_op_sqr_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sqr(x[i]);
+    }
+}
 
-    if (i >= k) {
-        return;
+template<typename T>
+static void unary_op_clamp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1, float min_val, float max_val) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_clamp(x[i], min_val, max_val);
     }
-    dst[i] = x[i] * x[i];
 }
 
 template<typename  T>
@@ -272,10 +320,10 @@ static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
     int i12 = (index / (ne10 * ne11)) % ne12;
     int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
 
-    int i00 = i10 / sf0;
-    int i01 = i11 / sf1;
-    int i02 = i12 / sf2;
-    int i03 = i13 / sf3;
+    int i00 = static_cast<int>(i10 / sf0);
+    int i01 = static_cast<int>(i11 / sf1);
+    int i02 = static_cast<int>(i12 / sf2);
+    int i03 = static_cast<int>(i13 / sf3);
 
     dst[index] = *(const T *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
@@ -283,8 +331,7 @@ static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
 template <typename T>
 static void pad(const T  *x, T *dst, const int ne0, const int ne00, const int ne01, const int ne02,
                     const sycl::nd_item<3> &item_ct1) {
-    int nidx = item_ct1.get_local_id(2) +
-               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    int nidx = SYCL_LOCAL_ID_CALC(item_ct1, 2);
     if (nidx >= ne0) {
         return;
     }
@@ -301,285 +348,72 @@ static void pad(const T  *x, T *dst, const int ne0, const int ne00, const int ne
     }
 }
 
-
 template<typename T>
 static void clamp(const T * x, T * dst, const float min, const float max, const int k,
-                      const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-
-    if (i >= k) {
-        return;
+                      const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
     }
-
-    dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
-}
-
-static void acc_f32_sycl(const float *x, const float *y, float *dst,
-                         const int n_elements, const int ne10, const int ne11,
-                         const int ne12, const int nb1, const int nb2,
-                         const int offset, queue_ptr stream) {
-    int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
-                    item_ct1);
-        });
-}
-
-template<typename T>
-static void gelu_sycl(const T *x, T *dst, const int k,
-                          queue_ptr stream) {
-    const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            gelu(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void silu_sycl(const T *x, T *dst, const int k,
-                          queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            silu(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            sgn(x, dst, k, item_ct1);
-            });
-}
-
-template<typename T>
-static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            abs_op(x, dst, k, item_ct1);
-            });
-}
-
-
-template<typename T>
-static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            elu_op(x, dst, k, item_ct1);
-            });
-}
-
-template<typename T>
-static void gelu_quick_sycl(const T *x, T *dst, const int k,
-                                queue_ptr stream) {
-    const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            gelu_quick(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void tanh_sycl(const T *x, T *dst, const int k,
-                          queue_ptr stream) {
-    const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            tanh(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void relu_sycl(const T *x, T *dst, const int k,
-                          queue_ptr stream) {
-    const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            relu(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void hardsigmoid_sycl(const T *x, T *dst, const int k,
-                                 queue_ptr stream) {
-    const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            hardsigmoid(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void hardswish_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            hardswish(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void exp_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            exp(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void log_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            log(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void neg_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            neg(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void step_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            step(x, dst, k, item_ct1);
-        });
 }
 
 template<typename T>
-static void sigmoid_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sigmoid(x, dst, k, item_ct1);
-        });
+static void gated_op_fused_geglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu(x[j0]) * g[j1];
+    }
 }
 
 template<typename T>
-static void sqrt_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sqrt(x, dst, k, item_ct1);
-        });
+static void gated_op_fused_reglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_relu(x[j0]) * g[j1];
+    }
 }
 
 template<typename T>
-static void sin_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sin(x, dst, k, item_ct1);
-        });
+static void gated_op_fused_swiglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1)  {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_silu(x[j0]) * g[j1];
+    }
 }
 
 template<typename T>
-static void cos_sycl(const T *x, T *dst, const int k,
-                               queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            cos(x, dst, k, item_ct1);
-        });
+static void gated_op_fused_geglu_erf(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu_erf(x[j0]) * g[j1];
+    }
 }
 
 template<typename T>
-static void leaky_relu_sycl(const T *x, T *dst, const int k,
-                                const float negative_slope,
-                                queue_ptr stream) {
-    const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            leaky_relu(x, dst, k, negative_slope, item_ct1);
-        });
+static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu_quick(x[j0]) * g[j1];
+    }
 }
 
-template<typename T>
-static void sqr_sycl(const T *x, T *dst, const int k,
-                         queue_ptr stream) {
-    const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE;
+namespace ggml_sycl_detail {
+static void acc_f32_sycl(const float *x, const float *y, float *dst,
+                         const int n_elements, const int ne10, const int ne11,
+                         const int ne12, const int nb1, const int nb2,
+                         const int offset, queue_ptr stream) {
+    int num_blocks = ceil_div(n_elements, SYCL_ACC_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sqr(x, dst, k, item_ct1);
+        sycl::nd_range<1>(sycl::range<1>(num_blocks) *
+                              sycl::range<1>(SYCL_ACC_BLOCK_SIZE),
+                          sycl::range<1>(SYCL_ACC_BLOCK_SIZE)),
+        [=](sycl::nd_item<1> item_ct1) {
+            acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
+                    item_ct1);
         });
 }
 
@@ -589,11 +423,10 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
                              const int ne12, const int ne13, const float sf0, const float sf1,
                              const float sf2, const float sf3, queue_ptr stream) {
     int dst_size = ne10 * ne11 * ne12 * ne13;
-    int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+    int num_blocks = ceil_div(dst_size, SYCL_UPSCALE_BLOCK_SIZE);
     sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
             upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
         });
 }
@@ -602,35 +435,19 @@ template<typename T>
 static void pad_sycl(const T *x, T *dst, const int ne00,
                          const int ne01, const int ne02, const int ne0,
                          const int ne1, const int ne2, queue_ptr stream) {
-    int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
+    int num_blocks = ceil_div(ne0, SYCL_PAD_BLOCK_SIZE);
     sycl::range<3> gridDim(ne2, ne1, num_blocks);
     stream->parallel_for(
-        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            pad(x, dst, ne0, ne00, ne01, ne02, item_ct1);
-        });
-}
-
-template<typename T>
-static void clamp_sycl(const T *x, T *dst, const float min,
-                           const float max, const int k,
-                           queue_ptr stream) {
-    const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            clamp(x, dst, min, max, k, item_ct1);
-        });
+                      sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
 }
 
-inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #if defined (GGML_SYCL_F16)
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-
 #else
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -643,14 +460,14 @@ inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(dst);
-                sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(dst);
-                sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -658,11 +475,11 @@ inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     }
 }
 
-inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_fused_glu(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #if defined (GGML_SYCL_F16)
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-
 #else
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -670,19 +487,66 @@ inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     GGML_ASSERT(dst->src[0]->type == dst->type);
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;;
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_is_contiguous_1(dst->src[0]));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
     switch (dst->type) {
 #if defined (GGML_SYCL_F16)
         case GGML_TYPE_F16:
             {
-                auto data_pts = cast_data<sycl::half>(dst);
-                abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                sycl::half * src0_p = (sycl::half *) src0_d;
+                sycl::half * src1_p = (sycl::half *) src1_d;
+
+                    if (!src1) {
+                        src0_p += swapped ? nc : 0;
+                        src1_p += swapped ? 0 : nc;
+                    }
+                kernel_invoker(src0_p,
+                               src1_p,
+                               (sycl::half *) dst_d,
+                               ggml_nelements(dst),
+                               nc,
+                               src0_o / sizeof(sycl::half),
+                               src1_o / sizeof(sycl::half),
+                               main_stream,
+                               std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
-                auto data_pts = cast_data<float>(dst);
-                abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                float * src0_p = (float *) src0_d;
+                float * src1_p = (float *) src1_d;
+
+                    if (!src1) {
+                        src0_p += swapped ? nc : 0;
+                        src1_p += swapped ? 0 : nc;
+                    }
+
+                kernel_invoker(src0_p,
+                               src1_p,
+                               (float *) dst_d,
+                               ggml_nelements(dst),
+                               nc,
+                               src0_o / sizeof(float),
+                               src1_o / sizeof(float),
+                               main_stream,
+                               std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -690,32 +554,41 @@ inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     }
 }
 
-
-inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #if defined (GGML_SYCL_F16)
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-
 #else
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 #endif
     GGML_ASSERT(dst->src[0]->type == dst->type);
+
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float sf0 = (float) dst->ne[0] / dst->src[0]->ne[0];
+    const float sf1 = (float) dst->ne[1] / dst->src[0]->ne[1];
+    const float sf2 = (float) dst->ne[2] / dst->src[0]->ne[2];
+    const float sf3 = (float) dst->ne[3] / dst->src[0]->ne[3];
     switch (dst->type) {
 #if defined (GGML_SYCL_F16)
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(dst);
-                elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
+                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
+                               main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(dst);
-                elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
+                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
+                               main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -723,7 +596,8 @@ inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     }
 }
 
-inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #if defined (GGML_SYCL_F16)
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
@@ -732,6 +606,7 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 #endif
     GGML_ASSERT(dst->src[0]->type == dst->type);
+    GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     switch (dst->type) {
@@ -739,14 +614,16 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(dst);
-                silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
+                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(dst);
-                silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
+                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -754,623 +631,320 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
     }
 }
 
-inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+} // namespace ggml_sycl_detail
+
+
+
+static inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sgn_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                gelu_quick_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                gelu_quick_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                tanh_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                tanh_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                hardsigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                hardsigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                hardswish_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                hardswish_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                exp_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                exp_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                log_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                log_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                sigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                sigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_abs_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                sqrt_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                sqrt_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_elu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                sin_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                sin_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SILU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SILU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_silu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                cos_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                cos_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                step_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                step_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_quick_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                neg_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                neg_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_erf_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
+static inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_TANH_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_TANH_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_TANH_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_tanh_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    float negative_slope;
-    memcpy(&negative_slope, dst->op_params, sizeof(float));
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                leaky_relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), negative_slope, main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                leaky_relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), negative_slope, main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_relu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
- #if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                sqr_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                sqr_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_HARDSIGMOID_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_hardsigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
+static inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_HARDSWISH_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_hardswish_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_exp_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
-    const float sf0 = (float) dst->ne[0] / dst->src[0]->ne[0];
-    const float sf1 = (float) dst->ne[1] / dst->src[0]->ne[1];
-    const float sf2 = (float) dst->ne[2] / dst->src[0]->ne[2];
-    const float sf3 = (float) dst->ne[3] / dst->src[0]->ne[3];
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                upscale_sycl(data_pts.src, data_pts.dst, dst->src[0]->nb[0], dst->src[0]->nb[1], dst->src[0]->nb[2],
-                        dst->src[0]->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
-                        main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                upscale_sycl(data_pts.src, data_pts.dst, dst->src[0]->nb[0], dst->src[0]->nb[1], dst->src[0]->nb[2],
-                        dst->src[0]->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
-                        main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE); // Using EXP block size
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1);  // just 3D tensors
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                pad_sycl(data_pts.src, data_pts.dst, dst->src[0]->ne[0], dst->src[0]->ne[1], dst->src[0]->ne[2], dst->ne[0],
-                        dst->ne[1], dst->ne[2], main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                pad_sycl(data_pts.src, data_pts.dst, dst->src[0]->ne[0], dst->src[0]->ne[1], dst->src[0]->ne[2], dst->ne[0],
-                        dst->ne[1], dst->ne[2], main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_neg_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-#if defined(GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
+static inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE); // Using NEG block size
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_step_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    float min;
-    float max;
-    memcpy(&min, dst->op_params, sizeof(float));
-    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+static inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIGMOID_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
-    switch (dst->type) {
-#if defined(GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                clamp_sycl(data_pts.src, data_pts.dst, min, max, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                clamp_sycl(data_pts.src, data_pts.dst, min, max, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
+static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SQRT_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE); // Using SIN block size
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float slope) {
+            const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
+                });
+        }, negative_slope);
+}
+
+static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SQR_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sqr_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_upscale(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int nb00, int nb01, int nb02, int nb03,
+           int ne10, int ne11, int ne12, int ne13, float sf0, float sf1, float sf2, float sf3,
+           queue_ptr stream) {
+            ggml_sycl_detail::upscale_sycl(src, dst_ptr, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, stream);
+        });
+}
+
+static inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_pad(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2,
+           queue_ptr stream) {
+            ggml_sycl_detail::pad_sycl(src, dst_ptr, ne00, ne01, ne02, ne0, ne1, ne2, stream);
+        });
+}
 
+static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    float min_val;
+    float max_val;
+    memcpy(&min_val, dst->op_params, sizeof(float));
+    memcpy(&max_val, (float *) dst->op_params + 1, sizeof(float));
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float min_arg, float max_arg) {
+            const int num_blocks = ceil_div(k_elements, SYCL_CLAMP_BLOCK_SIZE);
+            stream->parallel_for(
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
+                });
+        }, min_val, max_val);
+}
+
+static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -1386,7 +960,62 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
     // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
     int offset = dst->op_params[3] / 4; // offset in bytes
 
-    acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), dst->src[1]->ne[0], dst->src[1]->ne[1], dst->src[1]->ne[2], nb1, nb2, offset, main_stream);
+    ggml_sycl_detail::acc_f32_sycl(src0_dd, src1_dd, dst_dd, (int)ggml_nelements(dst), (int)dst->src[1]->ne[0], (int)dst->src[1]->ne[1], (int)dst->src[1]->ne[2], nb1, nb2, offset, main_stream);
+}
+
+static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            main_stream->parallel_for(
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
+            main_stream->parallel_for(
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
+            main_stream->parallel_for(
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            main_stream->parallel_for(
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            main_stream->parallel_for(
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
 }
 
 
@@ -1425,6 +1054,11 @@ void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_gelu_quick(ctx, dst);
 }
 
+void ggml_sycl_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_gelu_erf(ctx, dst);
+}
+
 void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_tanh(ctx, dst);
@@ -1509,3 +1143,28 @@ void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_elu(ctx, dst);
 }
+
+void ggml_sycl_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu(ctx, dst);
+}
+
+void ggml_sycl_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_reglu(ctx, dst);
+}
+
+void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_swiglu(ctx, dst);
+}
+
+void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu_erf(ctx, dst);
+}
+
+void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu_quick(ctx, dst);
+}