whispercpp 1.3.4 → 1.3.6

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

@@ -595,6 +595,25 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
+static void mul_mat_vec_mxfp4_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols, const int nrows,
+                                        dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_MXFP4 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+
+    {
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 mul_mat_vec_q<QK_MXFP4, QI_MXFP4, block_mxfp4, VDR_MXFP4_Q8_1_MMVQ, vec_dot_mxfp4_q8_1>(
+                                     vx, vy, dst, ncols, nrows, item_ct1);
+                             });
+        });
+    }
+}
+
+
 static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -1123,6 +1142,9 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
             case GGML_TYPE_IQ4_XS:
                 mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
+            case GGML_TYPE_MXFP4:
+                mul_mat_vec_mxfp4_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
             default:
                 GGML_ABORT("fatal error");
         }

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

@@ -202,47 +202,34 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const int6
     }
 }
 
-static void l2_norm_f32(const float* x, float* dst, const int ncols, const float eps,
-    const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
-    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
-        item_ct1.get_local_id(1);
-    const int tid = item_ct1.get_local_id(2);
-    const int nthreads = item_ct1.get_local_range(2);
-    const int nwarps = nthreads / WARP_SIZE;
+template<int warp_size>
+static void l2_norm_f32(const float * x, float * dst, const int ncols,
+    const int64_t stride_row, const int64_t stride_channel,
+    const int64_t stride_sample, const float eps,
+    const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {
+    const int nrows     = item_ct1.get_group_range(2);
+    const int nchannels = item_ct1.get_group_range(1);
+
+    const int row     = item_ct1.get_group(2);
+    const int channel = item_ct1.get_group(1);
+    const int sample  = item_ct1.get_group(0);
+    const int tid     = item_ct1.get_local_id(2);
+
+    x   += sample*stride_sample + channel*stride_channel + row*stride_row;
+    dst += ((sample*nchannels + channel)*nrows + row)*ncols;
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row * ncols + col];
+        const float xi = x[col];
         tmp += xi * xi;
     }
 
-    // sum up partial sums
-    tmp = warp_reduce_sum(tmp, item_ct1);
-    if (block_size > WARP_SIZE) {
-
-        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        /*
-        DPCT1118:3: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
-        item_ct1.barrier(sycl::access::fence_space::local_space);
-        size_t nreduce = nwarps / WARP_SIZE;
-        tmp = 0.f;
-        for (size_t i = 0; i < nreduce; i += 1)
-        {
-            tmp += s_sum[lane_id + i * WARP_SIZE];
-        }
-        tmp = warp_reduce_sum(tmp, item_ct1);
-    }
-
-    const float scale = sycl::rsqrt(sycl::max(tmp, eps * eps));
+    tmp = block_reduce<block_reduce_method::SUM, warp_size>(tmp, s_sum, block_size);
+    const float scale = sycl::rsqrt(sycl::fmax(tmp, eps * eps));
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row * ncols + col] = scale * x[row * ncols + col];
+        dst[col] = scale * x[col];
     }
 }
 
@@ -251,7 +238,6 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
         const float eps, queue_ptr stream, int device) {
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
@@ -334,7 +320,6 @@ static void group_norm_f32_sycl(const float* x, float* dst,
 
 static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
         const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -371,43 +356,50 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
     }
 }
 
-static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
-    const int nrows, const float eps,
-    queue_ptr stream, int device) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
-    // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
+template<int warp_size>
+static void l2_norm_f32_sycl(const float *   x,
+                             float *         dst,
+                             const int       ncols,
+                             const int       nrows,
+                             const int       nchannels,
+                             const int       nsamples,
+                             const int64_t   stride_row,
+                             const int64_t   stride_channel,
+                             const int64_t   stride_sample,
+                             const float     eps,
+                             queue_ptr       stream,
+                             int             device) {
+    const dpct::dim3 blocks_num(nrows, nchannels, nsamples);
+
     if (ncols < 1024) {
-        const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+        const dpct::dim3 block_dims(warp_size, 1, 1);
         stream->submit([&](sycl::handler& cgh) {
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                sycl::nd_range<3>(blocks_num * block_dims,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
+                [[sycl::reqd_sub_group_size(warp_size)]] {
+                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                        nullptr, warp_size);
                 });
             });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
-        assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+        assert(work_group_size % (warp_size * warp_size) == 0);
         const sycl::range<3> block_dims(1, 1, work_group_size);
-        /*
-        DPCT1049:19: The work-group size passed to the SYCL kernel may exceed
-        the limit. To get the device limit, query
-        info::device::max_work_group_size. Adjust the work-group size if needed.
-        */
+        int lsm_size =  block_dims[2] > warp_size ? work_group_size / warp_size * sizeof(float): 0;
         stream->submit([&](sycl::handler& cgh) {
-            sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
+            sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(lsm_size),
                 cgh);
+
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                sycl::nd_range<3>(blocks_num * block_dims,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
+                [[sycl::reqd_sub_group_size(warp_size)]] {
+                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -480,22 +472,185 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
-void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); // dz
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32); // x
+    GGML_ASSERT(dst->type         == GGML_TYPE_F32);
+
+    float eps = 1e-5f;
+    std::memcpy(&eps, dst->op_params, sizeof(float));
+    if (!(eps > 0.0f) || !std::isfinite(eps)) eps = 1e-5f;
+
+    const float * g_base  = static_cast<const float *>(dst->src[0]->data); // dz
+    const float * x_base  = static_cast<const float *>(dst->src[1]->data); // x
+          float * dx_base = static_cast<      float *>(dst->data);
+
+    const int64_t D  = dst->ne[0];
+    const int64_t n1 = dst->ne[1], n2 = dst->ne[2], n3 = dst->ne[3]; (void) n3;
+    const int64_t N  = ggml_nrows(dst);
+    if (D == 0 || N == 0) return;
+
+    const ggml_tensor *G = dst->src[0];
+    const ggml_tensor *X = dst->src[1];
+    const int ts = (int) ggml_type_size(X->type);
+    GGML_ASSERT((size_t) X->nb[0]   == (size_t) ts);
+    GGML_ASSERT((size_t) G->nb[0]   == (size_t) ts);
+    GGML_ASSERT((size_t) dst->nb[0] == (size_t) ts);
+
+    const int64_t xs1 = X->nb[1] / ts, xs2 = X->nb[2] / ts, xs3 = X->nb[3] / ts;
+    const int64_t gs1 = G->nb[1] / ts, gs2 = G->nb[2] / ts, gs3 = G->nb[3] / ts;
+    const int64_t ds1 = dst->nb[1] / ts, ds2 = dst->nb[2] / ts, ds3 = dst->nb[3] / ts;
+
+    dpct::queue_ptr q = ctx.stream();
+
+    // work-group size: multiple of WARP_SIZE, capped by device and 256, and not larger than D
+    const int device_max_wg = ggml_sycl_info().max_work_group_sizes[ctx.device];
+    auto roundup = [](int v, int m) { return ((v + m - 1) / m) * m; };
+    int wg_cap = 256;
+    if (device_max_wg > 0) wg_cap = std::min(wg_cap, device_max_wg);
+    int WG = std::max(WARP_SIZE, std::min(roundup((int)std::min<int64_t>(D, wg_cap), WARP_SIZE), wg_cap));
+
+    // FP32 path: per-thread compensated accumulation + hierarchical reduction
+    q->submit([&](sycl::handler &cgh) {
+        const int nwarps_loc = std::max(1, WG / WARP_SIZE);
+        // store one partial value per warp (xx and xg) for cross-warp reduction
+        auto l_xx   = sycl::local_accessor<sycl::float2, 1>(sycl::range<1>(nwarps_loc), cgh);
+        auto l_xg   = sycl::local_accessor<sycl::float2, 1>(sycl::range<1>(nwarps_loc), cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, N) * sycl::range<3>(1, 1, WG),
+                              sycl::range<3>(1, 1, WG)),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                const int row = item_ct1.get_group(2);
+                const int tid = item_ct1.get_local_id(2);
+
+                const int64_t i1 = row % n1;
+                const int64_t i2 = (row / n1) % n2;
+                const int64_t i3 = row / (n1 * n2);
+
+                const float *__restrict x_row = x_base + i3 * xs3 + i2 * xs2 + i1 * xs1;
+                const float *__restrict g_row = g_base + i3 * gs3 + i2 * gs2 + i1 * gs1;
+                float *__restrict d_row       = dx_base + i3 * ds3 + i2 * ds2 + i1 * ds1;
+
+                // per-thread accumulation (compensated by default)
+                float sum_xx = 0.f, sum_xg = 0.f;
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                float c_xx = 0.f, c_xg = 0.f;
+#endif
+                for (int64_t col = tid; col < D; col += WG) {
+                    const float xv = x_row[col];
+                    const float gv = g_row[col];
+#ifdef GGML_SYCL_RMS_BACK_FAST
+                    sum_xx += xv * xv;
+                    sum_xg += xv * gv;
+#else
+                    float y1 = xv * xv - c_xx;
+                    float t1 = sum_xx + y1;
+                    c_xx = (t1 - sum_xx) - y1;
+                    sum_xx = t1;
+
+                    float y2 = xv * gv - c_xg;
+                    float t2 = sum_xg + y2;
+                    c_xg = (t2 - sum_xg) - y2;
+                    sum_xg = t2;
+#endif
+                }
+
+                // warp-level reduction
+                sycl::float2 xx = sycl::float2(sum_xx,
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                    c_xx
+#else
+                    0.f
+#endif
+                );
+                sycl::float2 xg = sycl::float2(sum_xg,
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                    c_xg
+#else
+                    0.f
+#endif
+                );
+                xx = warp_reduce_sum(xx, item_ct1);
+                xg = warp_reduce_sum(xg, item_ct1);
+
+                // cross-warp reduction using local memory (single barrier)
+                const auto sub_group = item_ct1.get_sub_group();
+                const auto sg_id     = sub_group.get_group_linear_id();
+                const auto wi_in_sg  = sub_group.get_local_linear_id();
+                const int nthreads   = item_ct1.get_local_range(2);
+                const int nwarps     = nthreads / WARP_SIZE;
+
+                sycl::float2 xx_total = xx;
+                sycl::float2 xg_total = xg;
+                if (nwarps > 1) {
+                    if (wi_in_sg == 0) {
+                        l_xx[sg_id] = xx;
+                        l_xg[sg_id] = xg;
+                    }
+                    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+                    if (sg_id == 0) {
+                        const unsigned wi_u = wi_in_sg;
+                        sycl::float2 xx_first = (wi_u < static_cast<unsigned>(nwarps)) ? l_xx[wi_u] : sycl::float2(0.f, 0.f);
+                        sycl::float2 xg_first = (wi_u < static_cast<unsigned>(nwarps)) ? l_xg[wi_u] : sycl::float2(0.f, 0.f);
+                        xx_total = warp_reduce_sum(xx_first, item_ct1);
+                        xg_total = warp_reduce_sum(xg_first, item_ct1);
+                    } else {
+                        // other subgroups keep their local totals; they'll be ignored
+                        xx_total = xx;
+                        xg_total = xg;
+                    }
+                    // ensure all threads see the first-subgroup result via broadcast below
+                }
+
+                // compute inv_r and coeff once per row and broadcast to the whole work-group
+                float inv_r = 0.f;
+                float coeff = 0.f;
+                if (tid == 0) {
+                    const float sum_xx_f  = xx_total.x() + xx_total.y();
+                    const float sum_xdz_f = xg_total.x() + xg_total.y();
+                    const float mean_eps  = sum_xx_f / (float) D + eps;
+                    const float sum_eps   = sum_xx_f + eps * (float) D;
+                    inv_r = sycl::rsqrt(mean_eps);
+                    coeff = -sum_xdz_f / sum_eps;
+                }
+                inv_r = sycl::group_broadcast(item_ct1.get_group(), inv_r);
+                coeff = sycl::group_broadcast(item_ct1.get_group(), coeff);
+
+                for (int64_t col = tid; col < D; col += WG) {
+                    d_row[col] = (g_row[col] + coeff * x_row[col]) * inv_r;
+                }
+            });
+    });
 
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+}
 
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *) src0->data;
+    float * dst_d = (float *) dst->data;
+    dpct::queue_ptr     stream = ctx.stream();
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows = ggml_nrows(dst->src[0]);
-    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
-    float * dst_dd = static_cast<float *>(dst->data);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
 
-    l2_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
 
+    /*support both WARP_SIZE or WARP_32_SIZE in code
+      choose by hardware for better performance
+    */
+    l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream, ctx.device);
 }

data/ext/sources/ggml/src/ggml-sycl/norm.hpp

@@ -19,6 +19,8 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
+void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
+
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);

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

@@ -32,12 +32,12 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Handle transposition of src1
     const bool src1_T = ggml_is_transposed(src1);
-    const oneapi::math::transpose src1_op = src1_T ? oneapi::math::transpose::nontrans : oneapi::math::transpose::trans;
+    const oneapi::mkl::transpose src1_op = src1_T ? oneapi::mkl::transpose::nontrans : oneapi::mkl::transpose::trans;
     const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float);
 
     try {
-        // Perform matrix multiplication using oneMath GEMM
-        oneapi::math::blas::column_major::gemm(get_onemath_backend(*stream), oneapi::math::transpose::nontrans, src1_op,
+        // Perform matrix multiplication using oneMKL GEMM
+        oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op,
                                                ne0, ne1, ne01, alpha, src0_d, ne00, src1_d, ldb, beta, dst_d, ne0);
     }
     catch (sycl::exception const& exc) {

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

@@ -0,0 +1,97 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+//#include "common.hpp"
+#include "pad.hpp"
+
+static void pad_f32(const float * src, float * dst,
+                    const int lp0, const int rp0, const int lp1, const int rp1,
+                    const int lp2, const int rp2, const int lp3, const int rp3,
+                    const int ne0, const int ne1, const int ne2, const int ne3,
+                    sycl::nd_item<3> item_ct1) {
+    int i0 = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    int i1 = item_ct1.get_group(1);
+    int i2 = item_ct1.get_group(0) % ne2;
+    int i3 = item_ct1.get_group(0) / ne2;
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    // operation
+    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+    if ((i0 >= lp0 && i0 < ne0 - rp0) &&
+        (i1 >= lp1 && i1 < ne1 - rp1) &&
+        (i2 >= lp2 && i2 < ne2 - rp2) &&
+        (i3 >= lp3 && i3 < ne3 - rp3)) {
+        const int64_t i00 = i0 - lp0;
+        const int64_t i01 = i1 - lp1;
+        const int64_t i02 = i2 - lp2;
+        const int64_t i03 = i3 - lp3;
+        const int64_t ne02 = ne2 - lp2 - rp2;
+        const int64_t ne01 = ne1 - lp1 - rp1;
+        const int64_t ne00 = ne0 - lp0 - rp0;
+
+        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) +
+                                i02 * (ne00 * ne01) + i01 * ne00 + i00;
+
+        dst[dst_idx] = src[src_idx];
+    } else {
+        dst[dst_idx] = 0.0f;
+    }
+}
+
+static void pad_f32_sycl(const float *src, float *dst, const int lp0,
+                         const int rp0, const int lp1, const int rp1,
+                         const int lp2, const int rp2, const int lp3,
+                         const int rp3, const int ne0, const int ne1,
+                         const int ne2, const int ne3,
+                         dpct::queue_ptr stream) {
+    int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
+    dpct::dim3 gridDim(num_blocks, ne1, ne2 * ne3);
+    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_f32(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1,
+                    ne2, ne3, item_ct1);
+        });
+}
+
+void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr     stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int32_t lp0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t rp0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t lp1 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t rp1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t lp2 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t rp2 = ((const int32_t*)(dst->op_params))[5];
+    const int32_t lp3 = ((const int32_t*)(dst->op_params))[6];
+    const int32_t rp3 = ((const int32_t*)(dst->op_params))[7];
+
+    pad_f32_sycl(src0_d, dst_d,
+                 lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
+                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
+
+void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_pad(ctx, dst);
+}

data/ext/sources/ggml/src/ggml-sycl/pad.hpp

@@ -0,0 +1,24 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_PAD_HPP
+#define GGML_SYCL_PAD_HPP
+
+#include "common.hpp"
+
+#define SYCL_PAD_BLOCK_SIZE 256
+
+void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_PAD_HPP

data/ext/sources/ggml/src/ggml-sycl/pad_reflect_1d.cpp

@@ -0,0 +1,100 @@
+#include "pad_reflect_1d.hpp"
+
+static void pad_reflect_1d_kernel_f32(
+    const void *__restrict__ src0, void *__restrict__ dst, const int64_t ne0,
+    const int64_t ne00, const sycl::uint3 ne01, const int64_t ne02,
+    const int64_t ne03, const int64_t nb00, const int64_t nb01,
+    const int64_t nb02, const int64_t nb03, const int64_t nb0,
+    const int64_t nb1, const int64_t nb2, const int64_t nb3, const int p0,
+    const int p1, sycl::nd_item<3> item_ct1) {
+
+    const int64_t i3 = item_ct1.get_group(0);
+    const int64_t i2 = item_ct1.get_group(1);
+
+    const sycl::uint2 div_mod_packed =
+        fast_div_modulo(item_ct1.get_group(2), ne01);
+    const int64_t tile1 = div_mod_packed.y();
+    const int64_t tile0 = div_mod_packed.x();
+    const int64_t i1 = tile1;
+    const int64_t i0 =
+        item_ct1.get_local_id(2) + tile0 * item_ct1.get_local_range(2);
+
+    if (i0 >= ne0 || i1 >= ne01.z() || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
+
+    const char *src0_ptr =
+        (const char *)src0 + i3 * nb03 + i2 * nb02 + i1 * nb01;
+    char *dst_ptr = (char *)dst + i3 * nb3 + i2 * nb2 + i1 * nb1;
+
+    const int64_t rel_i0 = i0 - p0; // relative i0 in src0
+    int64_t src_idx;
+
+    if (rel_i0 < 0) {
+        // Left padding - reflect
+        src_idx = -rel_i0;
+    } else if (rel_i0 < ne00) {
+        // Middle - copy
+        src_idx = rel_i0;
+    } else {
+        // Right padding - reflect
+        src_idx = 2 * ne00 - 2 - rel_i0;
+    }
+    const float value = *(const float *)(src0_ptr + src_idx * nb00);
+    *(float *)(dst_ptr + i0 * nb0) = value;
+
+    GGML_UNUSED(p1);
+}
+
+void ggml_sycl_op_pad_reflect_1d(ggml_backend_sycl_context &ctx,
+                                 ggml_tensor *dst) {
+
+    const ggml_tensor *src0 = dst->src[0];
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int32_t *opts = (const int32_t *)dst->op_params;
+    const int p0 = opts[0];
+    const int p1 = opts[1];
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const sycl::uint3 ne01_packed = init_fastdiv_values(ne01);
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne0 = dst->ne[0];
+
+    GGML_ASSERT(ne0 == ne00 + p0 + p1);
+
+    constexpr int64_t bx = SYCL_PAD_REFLECT_1D_BLOCK_SIZE;
+    const int64_t tiles0 = (ne0 + bx - 1) / bx;
+    const dpct::dim3 grid_dims((unsigned)(ne01 * tiles0), (unsigned)ne02,
+                               (unsigned)ne03);
+    const dpct::dim3 block_dims((unsigned)bx, 1, 1);
+
+    stream->submit([&](sycl::handler &cgh) {
+        auto src0_data_ct0 = src0->data;
+        auto dst_data_ct1 = dst->data;
+        auto src0_nb_ct7 = src0->nb[0];
+        auto src0_nb_ct8 = src0->nb[1];
+        auto src0_nb_ct9 = src0->nb[2];
+        auto src0_nb_ct10 = src0->nb[3];
+        auto dst_nb_ct11 = dst->nb[0];
+        auto dst_nb_ct12 = dst->nb[1];
+        auto dst_nb_ct13 = dst->nb[2];
+        auto dst_nb_ct14 = dst->nb[3];
+
+        cgh.parallel_for(sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             pad_reflect_1d_kernel_f32(
+                                 src0_data_ct0, dst_data_ct1, ne0, ne00,
+                                 ne01_packed, ne02, ne03, src0_nb_ct7,
+                                 src0_nb_ct8, src0_nb_ct9, src0_nb_ct10,
+                                 dst_nb_ct11, dst_nb_ct12, dst_nb_ct13,
+                                 dst_nb_ct14, p0, p1, item_ct1);
+                         });
+    });
+}

data/ext/sources/ggml/src/ggml-sycl/pad_reflect_1d.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_PAD_REFLECT_1D_HPP
+#define GGML_SYCL_PAD_REFLECT_1D_HPP
+
+#include "common.hpp"
+
+#define SYCL_PAD_REFLECT_1D_BLOCK_SIZE 256
+
+void ggml_sycl_op_pad_reflect_1d(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
+
+#endif // GGML_SYCL_PAD_REFLECT_1D_HPP

data/ext/sources/ggml/src/ggml-sycl/presets.hpp

@@ -31,6 +31,7 @@
 #define SYCL_SQRT_BLOCK_SIZE 256
 #define SYCL_SIN_BLOCK_SIZE 256
 #define SYCL_SQR_BLOCK_SIZE 256
+#define SYCL_SET_BLOCK_SIZE 256
 #define SYCL_CPY_BLOCK_SIZE 32
 #define SYCL_SCALE_BLOCK_SIZE 256
 #define SYCL_CLAMP_BLOCK_SIZE 256
@@ -49,6 +50,7 @@
 #define SYCL_ARGMAX_BLOCK_SIZE 256
 #define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
 #define SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
+#define SYCL_ARANGE_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_SYCL_DMMV_X
@@ -71,4 +73,7 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA
 #define MUL_MAT_SRC1_COL_STRIDE 128
 
 #define QK_WARP_SIZE 32
+#define WARP_32_SIZE 32
+#define WARP_16_SIZE 16
+
 #endif // GGML_SYCL_PRESETS_HPP