whispercpp 1.3.6 → 1.3.8

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

@@ -9,9 +9,12 @@
 #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 int64_t ne,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
+static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+    const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+    const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
+    const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
     auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
     const int64_t i = SYCL_LOCAL_ID_CALC(item_ct1, 2);
 
@@ -30,9 +33,18 @@ static void acc_f32(const float * x, const float * y, float * dst, const int64_t
     tmp -= i11 * s11;
     const int64_t i10 = tmp;
 
-    float val = x[i];
+    int64_t tmp_dst = i;
+    const int64_t i3 = tmp_dst / (ne2*ne1*ne0);
+    tmp_dst -= i3 * (ne2*ne1*ne0);
+    const int64_t i2 = tmp_dst / (ne1*ne0);
+    tmp_dst -= i2 * (ne1*ne0);
+    const int64_t i1 = tmp_dst / ne0;
+    tmp_dst -= i1 * ne0;
+    const int64_t i0 = tmp_dst;
+
+    float val = *(const float *) (x + i0*nb00 + i1*nb01 + i2*nb02 + i3*nb03);
     if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
-        val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
+        val += *(const float *) (y + i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13);
     }
     dst[i] = val;
 }
@@ -43,14 +55,44 @@ 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 __dpct_inline__ T op_abs(T x) {
-    return sycl::fabs(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fabs(x);  // or experimental namespace if needed
+    } else {
+        return sycl::fabs(x);
+    }
+}
+
+template<typename T>
+static __dpct_inline__ T op_expm1(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::expm1(static_cast<float>(x))
+        );
+    } else {
+        return sycl::expm1(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_elu(T x) {
-    return (x > static_cast<T>(0.f)) ? x : sycl::expm1(x);
+    return (x > static_cast<T>(0.f)) ? x : op_expm1(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_tanh(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        constexpr int ver = __INTEL_LLVM_COMPILER;
+#if defined(__INTEL_LLVM_COMPILER) && (__INTEL_LLVM_COMPILER >= 20260000)
+            return sycl::ext::oneapi::experimental::tanh(x);
+#else
+            return static_cast<T>(sycl::tanh(static_cast<float>(x)));
+#endif
+    } else {
+        return sycl::tanh(x);
+    }
 }
 
 template<typename T>
@@ -59,69 +101,106 @@ static __dpct_inline__ T op_gelu(T x) {
     const T SQRT_2_OVER_PI = static_cast<T>(0.79788456080286535587989211986876f);
     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)));
+            op_tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_exp(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::exp(x);
+    } else {
+        return sycl::exp(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_silu(T x) {
-    return x / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return x / (static_cast<T>(1.0f) + op_exp(-x));
 }
 
 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)));
+static __dpct_inline__ T op_erf(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::erf(static_cast<float>(x))
+        );
+    } else {
+        return sycl::erf(x);
+    }
 }
 
 template<typename T>
 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));
+    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + op_erf(x * SQRT_2_INV));
 }
 
 template<typename T>
-static __dpct_inline__ T op_tanh(T x) {
-    return sycl::tanh(x);
+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) + op_exp(GELU_QUICK_COEF_LOCAL * x)));
 }
 
 template<typename T>
 static __dpct_inline__ T op_relu(T x) {
-    return sycl::fmax(x, static_cast<T>(0));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0));
+    } else {
+        return sycl::fmax(x, static_cast<T>(0));
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_sigmoid(T x) {
-    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(-x));
 }
 
 template<typename T>
 static __dpct_inline__ T op_sqrt(T x) {
-    return sycl::sqrt(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sqrt(x);
+    } else {
+        return sycl::sqrt(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_sin(T x) {
-    return sycl::sin(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sin(x);
+    } else {
+        return sycl::sin(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_cos(T x) {
-    return sycl::cos(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::cos(x);
+    } else {
+        return sycl::cos(x);
+    }
 }
 
 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)));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmin(
+            static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(
+                                      static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        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);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return x * sycl::ext::oneapi::experimental::fmin(static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        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>
@@ -129,13 +208,17 @@ static __dpct_inline__ T op_log(T x) {
     if (x <= static_cast<T>(0)) {
         return neg_infinity<T>();
     }
-    return sycl::log(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::log(x);
+    } else {
+        return sycl::log(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_softplus(T x) {
     const float xf = (float) x;
-    const float ax = sycl::fabs(xf);
+    const float ax = op_abs(xf);
     const float m  = sycl::fmax(xf, 0.0f);
     const float y  = m + sycl::log1p(sycl::exp(-ax));
     return (T) y;
@@ -154,8 +237,14 @@ static __dpct_inline__ T op_step(T x) {
 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)) +
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0)) +
+           sycl::ext::oneapi::experimental::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+
+    } else {
+        return sycl::fmax(x, static_cast<T>(0)) +
            sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+    }
 }
 
 template<typename T>
@@ -170,22 +259,40 @@ static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {
 
 template<typename T>
 static __dpct_inline__ T op_floor(T x) {
-    return sycl::floor(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::floor(x);
+    } else {
+        return sycl::floor(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_ceil(T x) {
-    return sycl::ceil(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::ceil(x);
+    } else {
+        return sycl::ceil(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_round(T x) {
-    return sycl::round(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::round(static_cast<float>(x))
+        );
+    } else {
+        return sycl::round(x);
+    }
 }
 
 template<typename T>
 static __dpct_inline__ T op_trunc(T x) {
-    return sycl::trunc(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::trunc(x);
+    } else {
+        return sycl::trunc(x);
+    }
 }
 
 template<typename T, typename F>
@@ -266,13 +373,6 @@ static void unary_op_clamp_kernel(const T * x, T * dst, const int k, const sycl:
     }
 }
 
-template<typename T>
-static void unary_op_floor_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_floor(x[i]);
-    }
-}
-
 template<typename T>
 static void unary_op_ceil_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
     SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
@@ -280,44 +380,6 @@ static void unary_op_ceil_kernel(const T * x, T * dst, const int k, const sycl::
     }
 }
 
-template<typename T>
-static void unary_op_round_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_round(x[i]);
-    }
-}
-
-template<typename T>
-static void unary_op_trunc_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_trunc(x[i]);
-    }
-}
-
-template<typename  T>
-static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
-                        const int nb02, const int nb03, const int ne10, const int ne11,
-                        const int ne12, const int ne13, const float sf0, const float sf1,
-                        const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) {
-    int index = item_ct1.get_local_id(0) +
-               item_ct1.get_group(0) * item_ct1.get_local_range(0);
-    if (index >= ne10 * ne11 * ne12 * ne13) {
-        return;
-    }
-    // operation
-    int i10 = index % ne10;
-    int i11 = (index / ne10) % ne11;
-    int i12 = (index / (ne10 * ne11)) % ne12;
-    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
-
-    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);
-}
-
 template<typename T>
 static void clamp(const T * x, T * dst, const float min, const float max, const int k,
                       const sycl::nd_item<1> &item_ct1) {
@@ -372,15 +434,24 @@ static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const
 }
 
 namespace ggml_sycl_detail {
-static void acc_f32_sycl(const float *x, const float *y, float *dst,
-                         const int64_t n_elements, const int64_t ne10, const int64_t ne11,
-                         const int64_t ne12, const int64_t ne13, const int64_t s1, const int64_t s2, const int64_t s3,
+static void acc_f32_sycl(const char *x, const char *y, float *dst,
+                         const int64_t n_elements,
+                         const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+                         const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+                         const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+                         const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
+                         const int64_t s1, const int64_t s2, const int64_t s3,
                          const int64_t offset, queue_ptr stream) {
     const 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, ne13, s1, s2, s3, offset);
+                         [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                            acc_f32(x, y, dst, n_elements,
+                                ne0, ne1, ne2, ne3,
+                                nb00, nb01, nb02, nb03,
+                                ne10, ne11, ne12, ne13,
+                                nb10, nb11, nb12, nb13,
+                                s1, s2, s3, offset);
                          });
 }
 
@@ -392,24 +463,10 @@ static void arange_kernel(T * dst, const int k, T start, T step,
     }
 }
 
-template<typename T>
-static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
-                             const int nb02, const int nb03, const int ne10, const int ne11,
-                             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 = 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) {
-            upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
-        });
-}
-
 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) {
-    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);
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16 || dst->src[0]->type == GGML_TYPE_BF16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_BF16);
     GGML_ASSERT(dst->src[0]->type == dst->type);
 
     dpct::queue_ptr main_stream = ctx.stream();
@@ -421,6 +478,14 @@ static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx,
                 kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
+#ifdef GGML_SYCL_HAS_BF16
+        case GGML_TYPE_BF16:
+            {
+                auto data_pts = cast_data<sycl::ext::oneapi::bfloat16>(dst);
+                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);
@@ -505,42 +570,6 @@ static inline void dispatch_ggml_sycl_op_fused_glu(ggml_backend_sycl_context & c
     }
 }
 
-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) {
-    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);
-
-    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) {
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                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;
-            }
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                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:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
 template<typename F>
 static inline void ggml_sycl_op_unary(
         ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
@@ -570,7 +599,7 @@ static inline void ggml_sycl_op_unary(
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_generic_kernel(
                         src, dst_ptr, k_elements,
                         ne0, ne1, ne2, ne3,
@@ -598,7 +627,7 @@ static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_ten
     stream->parallel_for(
         sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE),
                           sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             arange_kernel(dst_ptr, k, start, step, item_ct1);
         });
 }
@@ -679,6 +708,12 @@ static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor
     });
 }
 
+static inline void ggml_sycl_op_expm1(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_expm1(x);
+    });
+}
+
 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) {
@@ -686,7 +721,7 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -724,7 +759,7 @@ static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tenso
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -737,7 +772,7 @@ static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -750,7 +785,7 @@ static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
                 });
         });
@@ -765,7 +800,7 @@ static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
                 });
         }, negative_slope);
@@ -778,21 +813,12 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     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_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     float min_val;
     float max_val;
@@ -804,23 +830,16 @@ static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tens
             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) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                     clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
                 });
         }, min_val, max_val);
 }
 
 static inline void ggml_sycl_op_floor(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_floor_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_floor(x);
+    });
 }
 
 static inline void ggml_sycl_op_ceil(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -830,37 +849,23 @@ static inline void ggml_sycl_op_ceil(ggml_backend_sycl_context & ctx, ggml_tenso
 }
 
 static inline void ggml_sycl_op_round(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_round_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_round(x);
+    });
 }
 
 static inline void ggml_sycl_op_trunc(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_trunc_kernel(src, dst_ptr, k_elements, item_ct1);
-                });
-        });
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_trunc(x);
+    });
 }
 
 static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const float * src0_d = (const float *) src0->data;
-    const float * src1_d = (const float *) src1->data;
+    const char  * src0_d = (const char  *) src0->data;
+    const char  * src1_d = (const char  *) src1->data;
     float       * dst_d  = (float       *)  dst->data;
 
     dpct::queue_ptr stream = ctx.stream();
@@ -869,17 +874,20 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(ggml_is_contiguous(src1));
     GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
     GGML_ASSERT(ggml_is_contiguously_allocated(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    const int64_t s1     = dst->op_params[0] / sizeof(float);
-    const int64_t s2     = dst->op_params[1] / sizeof(float);
-    const int64_t s3     = dst->op_params[2] / sizeof(float);
-    const int64_t offset = dst->op_params[3] / sizeof(float);
+    const int64_t s1     = (int64_t) ((const int32_t *) dst->op_params)[0] / (int64_t) sizeof(float);
+    const int64_t s2     = (int64_t) ((const int32_t *) dst->op_params)[1] / (int64_t) sizeof(float);
+    const int64_t s3     = (int64_t) ((const int32_t *) dst->op_params)[2] / (int64_t) sizeof(float);
+    const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);
 
     ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+        src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
         s1, s2, s3, offset, stream);
 }
 
@@ -888,7 +896,8 @@ static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tens
         [](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) {
+                    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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -899,7 +908,8 @@ static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tens
         [](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) {
+                    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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -910,7 +920,8 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
         [](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) {
+                    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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -925,7 +936,6 @@ __dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alp
     return out_glu;
 }
 
-
 template <typename T>
 static void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k,
                               const int64_t n, const int64_t o0, const int64_t o1,
@@ -959,7 +969,7 @@ static void swiglu_oai_sycl(const T *       x,
     const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
                                            sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                              swiglu_oai_kernel(x, g, dst, k, n, o0, o1, alpha, limit, item_ct1);
                          });
 }
@@ -1013,7 +1023,8 @@ static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_
         [](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) {
+                    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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -1024,7 +1035,8 @@ static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggm
         [](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) {
+                    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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
             });
         });
@@ -1101,6 +1113,11 @@ void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_exp(ctx, dst);
 }
 
+void ggml_sycl_expm1(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_expm1(ctx, dst);
+}
+
 void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_log(ctx, dst);
@@ -1131,12 +1148,6 @@ void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_sqr(ctx, dst);
 }
 
-void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
-    ggml_sycl_op_upscale(ctx, dst);
-}
-
-
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_clamp(ctx, dst);