whispercpp 1.3.5 → 1.3.7

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

@@ -2,13 +2,6 @@
 #include "dequantize.hpp"
 #include "presets.hpp"
 
-#if defined(__INTEL_LLVM_COMPILER)
-    #if __has_include(<sycl/ext/oneapi/bfloat16.hpp>)
-        #include <sycl/ext/oneapi/bfloat16.hpp>
-        #define GGML_SYCL_HAS_BF16
-    #endif
-#endif
-
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k,
                              const sycl::nd_item<3> &item_ct1) {
@@ -114,6 +107,19 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_row_q3_K_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
+                                             dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+        [=](sycl::nd_item<3> item_ct1) {
+            dequantize_block_q3_K_reorder(vx, y, item_ct1, nb);
+        });
+}
+
 template <typename dst_t>
 static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -151,6 +157,25 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int
 
 }
 
+template <typename dst_t>
+static void dequantize_row_q8_0_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+
+    dpct::has_capability_or_fail(stream->get_device(),
+                                    {sycl::aspect::fp16});
+
+    int constexpr WARP_K = WARP_SIZE * QK8_0;
+    const int n_warp = (k + WARP_K - 1) / WARP_K;
+    GGML_ASSERT(k % QK8_0 == 0);
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) *
+        sycl::range<3>(1, 1, WARP_SIZE),
+        sycl::range<3>(1, 1, WARP_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
+            dequantize_block_q8_0_reorder(vx, y, k, item_ct1);
+        });
+
+}
+
 template <typename dst_t>
 static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -240,6 +265,23 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_row_q5_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    stream->submit([&](sycl::handler & cgh) {
+        sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(K_SCALE_SIZE), cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) {
+                dequantize_block_q5_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
+            });
+    });
+}
+
 template <typename dst_t>
 static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -482,6 +524,75 @@ static void dequantize_row_mxfp4_sycl(const void * vx, dst_t * y, const int64_t
         });
 }
 
+template <typename dst_t>
+static void dequantize_row_nvfp4_sycl(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    GGML_ASSERT(k % QK_NVFP4 == 0);
+    const int nb = k / QK_NVFP4;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+        [=](sycl::nd_item<3> /*item_ct1*/) {
+            dequantize_block_nvfp4(vx, y, k);
+        });
+}
+
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_nc(const void * __restrict__ vx, dst_t * __restrict__ y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    auto          item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t i00 = 2 * (int64_t(item_ct1.get_local_range(2)) * item_ct1.get_group(2) + item_ct1.get_local_id(2));
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int64_t i01 = item_ct1.get_group(1);
+    const int64_t i02 = item_ct1.get_group(0) % ne02;
+    const int64_t i03 = item_ct1.get_group(0) / ne02;
+
+    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+
+    const int64_t ib = ibx0 + i00/qk; // block index
+    const int64_t iqs = (i00%qk)/qr; // quant index
+    const int64_t iybs = i00 - i00%qk; // y block start index
+    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    #ifdef GGML_SYCL_F16
+        sycl::half2 v;
+    #else
+        sycl::float2 v;
+    #endif
+
+    dequantize_kernel(vx, ib, iqs, v);
+
+    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
+    y[iy0 + 0]        = ggml_sycl_cast<dst_t>(v.x());
+    y[iy0 + y_offset] = ggml_sycl_cast<dst_t>(v.y());
+}
+
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_nc_sycl(const void *    vx,
+                                  dst_t *         y,
+                                  const int64_t   ne00,
+                                  const int64_t   ne01,
+                                  const int64_t   ne02,
+                                  const int64_t   ne03,
+                                  const int64_t   s01,
+                                  const int64_t   s02,
+                                  const int64_t   s03,
+                                  dpct::queue_ptr stream) {
+    const dpct::dim3 num_blocks((ne00 + 2 * SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2 * SYCL_DEQUANTIZE_BLOCK_SIZE), ne01,
+                                ne02 * ne03);
+    stream->parallel_for(sycl::nd_range<3>(num_blocks * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                                           sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             GGML_UNUSED(item_ct1);
+                             dequantize_block_nc<qk, qr, dequantize_kernel>(vx, y, ne00, ne01, ne02, s01, s02, s03);
+                         });
+}
 template <typename src_t, typename dst_t>
 static void convert_unary_nc(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
                           const int64_t ne02, const int64_t s01, const int64_t s02, const int64_t s03,
@@ -545,11 +656,20 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q8_0_sycl_reorder;
+            } else {
+                return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            }
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_sycl;
         case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q3_K_sycl_reorder;
+            } else {
+                return dequantize_row_q3_K_sycl;
+            }
         case GGML_TYPE_Q4_K:
             if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
                 return dequantize_row_q4_K_sycl_reorder;
@@ -557,7 +677,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
                 return dequantize_row_q4_K_sycl;
             }
         case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q5_K_sycl_reorder;
+            } else {
+                return dequantize_row_q5_K_sycl;
+            }
         case GGML_TYPE_Q6_K:
             if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
                 return dequantize_row_q6_K_sycl_reorder;
@@ -584,6 +708,8 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
             return dequantize_row_iq4_nl_sycl;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_sycl;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_sycl;
         case GGML_TYPE_F32:
             return convert_unary_sycl<float>;
 #ifdef GGML_SYCL_HAS_BF16
@@ -591,6 +717,7 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
             return convert_unary_sycl<sycl::ext::oneapi::bfloat16>;
 #endif
         default:
+            GGML_ABORT("fatal error: unsupport data type=%s\n", ggml_type_name(type));
             return nullptr;
     }
 }
@@ -611,11 +738,20 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q8_0_sycl_reorder;
+            } else {
+                return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            }
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_sycl;
         case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q3_K_sycl_reorder;
+            } else {
+                return dequantize_row_q3_K_sycl;
+            }
         case GGML_TYPE_Q4_K:
             if (dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
@@ -624,7 +760,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
                 return dequantize_row_q4_K_sycl;
             }
         case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q5_K_sycl_reorder;
+            } else {
+                return dequantize_row_q5_K_sycl;
+            }
         case GGML_TYPE_Q6_K:
             if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
                 return dequantize_row_q6_K_sycl_reorder;
@@ -651,6 +791,8 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
             return dequantize_row_iq4_nl_sycl;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_sycl;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_sycl;
         case GGML_TYPE_F16:
             return convert_unary_sycl<sycl::half>;
 #ifdef GGML_SYCL_HAS_BF16
@@ -658,11 +800,29 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
             return convert_unary_sycl<sycl::ext::oneapi::bfloat16>;
 #endif
         default:
+            GGML_ABORT("fatal error: unsupport data type=%s\n", ggml_type_name(type));
             return nullptr;
     }
 }
 
-to_fp16_nc_sycl_t get_to_fp16_nc_sycl(ggml_type type) {
+
+#ifdef GGML_SYCL_HAS_BF16
+to_bf16_sycl_t ggml_get_to_bf16_sycl(ggml_type type, ggml_tensor * /*dst*/) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_sycl<float>;
+        case GGML_TYPE_F16:
+            return convert_unary_sycl<sycl::half>;
+        case GGML_TYPE_BF16:
+            return convert_unary_sycl<sycl::ext::oneapi::bfloat16>;
+        default:
+            GGML_ABORT("fatal error: unsupport data type=%s\n", ggml_type_name(type));
+            return nullptr;
+    }
+}
+#endif
+
+to_fp16_nc_sycl_t ggml_get_to_fp16_nc_sycl(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_nc_sycl<float>;
@@ -670,6 +830,16 @@ to_fp16_nc_sycl_t get_to_fp16_nc_sycl(ggml_type type) {
         case GGML_TYPE_BF16:
             return convert_unary_nc_sycl<sycl::ext::oneapi::bfloat16>;
 #endif
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_nc_sycl<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_nc_sycl<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_nc_sycl<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_nc_sycl<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_nc_sycl<QK8_0, QR8_0, dequantize_q8_0>;
         default:
             return nullptr;
     }

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

@@ -23,12 +23,42 @@ typedef to_t_sycl_t<sycl::half> to_fp16_sycl_t;
 to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst);
 to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor * dst);
 
+#ifdef GGML_SYCL_HAS_BF16
+typedef to_t_sycl_t<sycl::ext::oneapi::bfloat16> to_bf16_sycl_t;
+to_bf16_sycl_t ggml_get_to_bf16_sycl(ggml_type type, ggml_tensor * dst);
+#endif
+
 // Nc = Non-contiguous
 template <typename T>
 using to_t_nc_sycl_t = void (*)(const void * x, T * y, int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
                                    int64_t s01, int64_t s02, int64_t s03, dpct::queue_ptr queue);
 
 typedef to_t_nc_sycl_t<sycl::half> to_fp16_nc_sycl_t;
-to_fp16_nc_sycl_t get_to_fp16_nc_sycl(ggml_type type);
+to_fp16_nc_sycl_t ggml_get_to_fp16_nc_sycl(ggml_type type);
+
+template<typename dst_t, typename src_t>
+ inline dst_t ggml_sycl_cast(src_t x) {
+    if constexpr (std::is_same_v<dst_t, src_t>) {
+        return x;
+#ifdef GGML_SYCL_HAS_BF16
+    } else if constexpr (std::is_same_v<dst_t, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::bfloat16(float(x));
+    } else if constexpr (std::is_same_v<src_t, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<float>(x);
+#endif
+    } else if constexpr (std::is_same_v<src_t, sycl::float2> && std::is_same_v<dst_t, sycl::half2>) {
+        return x.template convert<sycl::half, sycl::rounding_mode::rte>();
+#ifdef GGML_SYCL_HAS_BF16
+    } else if constexpr (std::is_same_v<src_t, sycl::float2> &&
+                         std::is_same_v<dst_t, sycl::vec<sycl::ext::oneapi::bfloat16, 2>>) {
+        return {x.x, x.y};
+#endif
+    } else if constexpr(std::is_same_v<dst_t, int32_t>) {
+        return int32_t(x);
+    } else {
+        return float(x);
+    }
+}
+
 
 #endif  // GGML_SYCL_CONVERT_HPP

data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp

@@ -18,7 +18,7 @@ static void count_equal(const T *__restrict__ x, const T *__restrict__ y,
         nequal += xi == yi;
     }
 
-    nequal = warp_reduce_sum(nequal);
+    nequal = warp_reduce_sum<WARP_SIZE>(nequal);
 
     if (item_ct1.get_local_id(2) != 0) {
         return;

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

@@ -0,0 +1,148 @@
+#include "cumsum.hpp"
+#include "common.hpp"
+
+#include <algorithm>
+
+#define SYCL_CUMSUM_BLOCK_SIZE 256
+
+static __dpct_inline__ float warp_prefix_inclusive_sum_f32(float x, const sycl::nd_item<3> & item) {
+    return sycl::inclusive_scan_over_group(item.get_sub_group(), x, sycl::plus<float>());
+}
+
+static void cumsum_f32_kernel(
+        const float * __restrict__ src, float * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t  d1, const int64_t  d2, const int64_t  d3,
+        const sycl::nd_item<3> & item, float * smem) {
+
+    const int tid = item.get_local_id(2);
+    const int block_size = item.get_local_range(2);
+    const int lane = tid % WARP_SIZE;
+    const int warp = tid / WARP_SIZE;
+    const int warps_per_block = block_size / WARP_SIZE;
+
+    float * s_vals      = smem;
+    float * s_warp_sums = smem + block_size;
+    float * s_carry     = smem + block_size + warps_per_block;
+
+    if (tid == 0) {
+        s_carry[0] = 0.0f;
+    }
+    item.barrier(sycl::access::fence_space::local_space);
+
+    const int64_t i3 = item.get_group(0);
+    const int64_t i2 = item.get_group(1);
+    const int64_t i1 = item.get_group(2);
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    const float * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
+    float       * dst_row = dst + i1 * d1  + i2 * d2  + i3 * d3;
+
+    constexpr int num_unroll = 4;
+    float temp[num_unroll];
+
+    for (int64_t i = 0; i < ne00; i += num_unroll * block_size) {
+        int64_t idx = i + tid * num_unroll;
+
+        temp[0] = (idx < ne00 ? src_row[idx] : 0.0f);
+#pragma unroll
+        for (int j = 1; j < num_unroll; j++) {
+            temp[j] = temp[j - 1];
+            if (idx + j < ne00) {
+                temp[j] += src_row[idx + j];
+            }
+        }
+
+        float val = (idx < ne00) ? temp[num_unroll - 1] : 0.0f;
+
+        val = warp_prefix_inclusive_sum_f32(val, item);
+        s_vals[tid] = val;
+
+        if (lane == WARP_SIZE - 1) {
+            s_warp_sums[warp] = val;
+        }
+        item.barrier(sycl::access::fence_space::local_space);
+
+        if (warp == 0) {
+            float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f;
+            float inc = warp_prefix_inclusive_sum_f32(w, item);
+            if (tid < warps_per_block) {
+                s_warp_sums[tid] = inc - w;
+            }
+            if (tid == warps_per_block - 1) {
+                s_carry[1] = inc;
+            }
+        }
+        item.barrier(sycl::access::fence_space::local_space);
+
+        float carry = s_carry[0];
+        float final_offset = s_vals[tid] + s_warp_sums[warp] + carry - temp[num_unroll - 1];
+
+#pragma unroll
+        for (int j = 0; j < num_unroll; j++) {
+            if (idx + j < ne00) {
+                dst_row[idx + j] = temp[j] + final_offset;
+            }
+        }
+
+        item.barrier(sycl::access::fence_space::local_space);
+
+        if (tid == 0) {
+            s_carry[0] += s_carry[1];
+        }
+    }
+}
+
+inline void ggml_sycl_op_cumsum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    dpct::queue_ptr stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float * src_d = static_cast<const float *>(src0->data);
+    float       * dst_d = static_cast<float *>(dst->data);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const size_t ts = sizeof(float);
+    const int64_t s01 = src0->nb[1] / ts;
+    const int64_t s02 = src0->nb[2] / ts;
+    const int64_t s03 = src0->nb[3] / ts;
+    const int64_t d1  = dst->nb[1] / ts;
+    const int64_t d2  = dst->nb[2] / ts;
+    const int64_t d3  = dst->nb[3] / ts;
+
+    const int num_warps = (ne00 + WARP_SIZE - 1) / WARP_SIZE;
+    int block_size = num_warps * WARP_SIZE;
+    block_size = std::min(block_size, SYCL_CUMSUM_BLOCK_SIZE);
+    const int warps_per_block = block_size / WARP_SIZE;
+    const int smem_size = block_size + warps_per_block + 2;
+
+    const sycl::range<3> grid(ne03, ne02, ne01);
+    const sycl::range<3> block(1, 1, block_size);
+
+    stream->submit([&](sycl::handler & cgh) {
+        sycl::local_accessor<float, 1> smem_acc(sycl::range<1>(smem_size), cgh);
+        cgh.parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                cumsum_f32_kernel(src_d, dst_d, ne00, ne01, ne02, ne03,
+                                  s01, s02, s03, d1, d2, d3,
+                                  item, get_pointer(smem_acc));
+            });
+    });
+}
+
+void ggml_sycl_cumsum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_cumsum(ctx, dst);
+}

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

@@ -0,0 +1,5 @@
+#pragma once
+
+#include "common.hpp"
+
+void ggml_sycl_cumsum(ggml_backend_sycl_context & ctx, ggml_tensor * dst);