whispercpp 1.3.6 → 1.3.8

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

@@ -48,6 +48,287 @@ inline void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
     }
 }
 
+inline void cpy_blck_f32_q1_0(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q1_0 *  dsti = (block_q1_0 *) cdsti;
+
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; ++j) {
+        sum_abs += sycl::fabs((float) xi[j]);
+    }
+
+    dsti->d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; ++j) {
+        dsti->qs[j] = 0;
+    }
+
+    for (int j = 0; j < QK1_0; ++j) {
+        if (xi[j] >= 0.0f) {
+            dsti->qs[j / 8] |= (1u << (j % 8));
+        }
+    }
+}
+
+inline int best_index_mxfp4(const float x, const float e) {
+    int best_index = 0;
+    float best_err = sycl::fabs((float) (kvalues_mxfp4[0] * e - x));
+    for (int i = 1; i < 16; ++i) {
+        const float err = sycl::fabs((float) (kvalues_mxfp4[i] * e - x));
+        if (err < best_err) {
+            best_index = i;
+            best_err = err;
+        }
+    }
+    return best_index;
+}
+
+inline int nearest_int_sycl(float x) {
+    const float val = x + 12582912.0f;
+    int i;
+    memcpy(&i, &val, sizeof(int));
+    return (i & 0x007fffff) - 0x00400000;
+}
+
+inline int nearest_int_ggml_sycl(float x) {
+    return (int) sycl::round((float) x);
+}
+
+inline uint8_t clamp_u8(const int x, const int lo, const int hi) {
+    return (uint8_t) dpct::max(lo, dpct::min(hi, x));
+}
+
+inline int8_t clamp_i8(const int x, const int lo, const int hi) {
+    return (int8_t) dpct::max(lo, dpct::min(hi, x));
+}
+
+constexpr float GROUP_MAX_EPS_SYCL = 1e-15f;
+
+inline float make_qx_quants_sycl(int n, int nmax, const float * x, int8_t * L, int rmse_type, const float * qw) {
+    float max = 0.0f;
+    float amax = 0.0f;
+    for (int i = 0; i < n; ++i) {
+        const float ax = sycl::fabs(x[i]);
+        if (ax > amax) {
+            amax = ax;
+            max = x[i];
+        }
+    }
+    if (amax < GROUP_MAX_EPS_SYCL) {
+        for (int i = 0; i < n; ++i) {
+            L[i] = 0;
+        }
+        return 0.0f;
+    }
+
+    float iscale = -nmax / max;
+    if (rmse_type == 0) {
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int_ggml_sycl(iscale * x[i]);
+            L[i] = (int8_t) (nmax + dpct::max(-nmax, dpct::min(nmax - 1, l)));
+        }
+        return 1.0f / iscale;
+    }
+
+    bool return_early = false;
+    if (rmse_type < 0) {
+        rmse_type = -rmse_type;
+        return_early = true;
+    }
+
+    float sumlx = 0.0f;
+    float suml2 = 0.0f;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int_ggml_sycl(iscale * x[i]);
+        l = dpct::max(-nmax, dpct::min(nmax - 1, l));
+        L[i] = (int8_t) (l + nmax);
+
+        const float w = qw ? qw[i] : (rmse_type == 1 ? x[i] * x[i] :
+            rmse_type == 2 ? 1.0f : rmse_type == 3 ? sycl::fabs(x[i]) : sycl::sqrt(sycl::fabs(x[i])));
+
+        sumlx += w * x[i] * l;
+        suml2 += w * l * l;
+    }
+
+    float scale = suml2 ? sumlx / suml2 : 0.0f;
+    if (return_early) {
+        return suml2 > 0.0f ? 0.5f * (scale + 1.0f / iscale) : 1.0f / iscale;
+    }
+
+    float best = scale * sumlx;
+    for (int is = -9; is <= 9; ++is) {
+        if (is == 0) {
+            continue;
+        }
+        iscale = -(nmax + 0.1f * is) / max;
+        sumlx = 0.0f;
+        suml2 = 0.0f;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int_ggml_sycl(iscale * x[i]);
+            l = dpct::max(-nmax, dpct::min(nmax - 1, l));
+            const float w = qw ? qw[i] : (rmse_type == 1 ? x[i] * x[i] :
+                rmse_type == 2 ? 1.0f : rmse_type == 3 ? sycl::fabs(x[i]) : sycl::sqrt(sycl::fabs(x[i])));
+            sumlx += w * x[i] * l;
+            suml2 += w * l * l;
+        }
+
+        if (suml2 > 0.0f && sumlx * sumlx > best * suml2) {
+            for (int i = 0; i < n; ++i) {
+                int l = nearest_int_ggml_sycl(iscale * x[i]);
+                L[i] = (int8_t) (nmax + dpct::max(-nmax, dpct::min(nmax - 1, l)));
+            }
+            scale = sumlx / suml2;
+            best = scale * sumlx;
+        }
+    }
+
+    return scale;
+}
+
+inline float make_q3_quants_sycl(int n, int nmax, const float * x, int8_t * L, bool do_rmse) {
+    float max = 0.0f;
+    float amax = 0.0f;
+    for (int i = 0; i < n; ++i) {
+        const float ax = sycl::fabs(x[i]);
+        if (ax > amax) {
+            amax = ax;
+            max = x[i];
+        }
+    }
+
+    if (amax < GROUP_MAX_EPS_SYCL) {
+        for (int i = 0; i < n; ++i) {
+            L[i] = 0;
+        }
+        return 0.0f;
+    }
+
+    const float iscale = -nmax / max;
+    if (do_rmse) {
+        float sumlx = 0.0f;
+        float suml2 = 0.0f;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int_ggml_sycl(iscale * x[i]);
+            l = dpct::max(-nmax, dpct::min(nmax - 1, l));
+            L[i] = (int8_t) l;
+            const float w = x[i] * x[i];
+            sumlx += w * x[i] * l;
+            suml2 += w * l * l;
+        }
+
+        for (int itry = 0; itry < 5; ++itry) {
+            int n_changed = 0;
+            for (int i = 0; i < n; ++i) {
+                const float w = x[i] * x[i];
+                float slx = sumlx - w * x[i] * L[i];
+                if (slx > 0.0f) {
+                    float sl2 = suml2 - w * L[i] * L[i];
+                    int new_l = nearest_int_ggml_sycl(x[i] * sl2 / slx);
+                    new_l = dpct::max(-nmax, dpct::min(nmax - 1, new_l));
+                    if (new_l != L[i]) {
+                        slx += w * x[i] * new_l;
+                        sl2 += w * new_l * new_l;
+                        if (sl2 > 0.0f && slx * slx * suml2 > sumlx * sumlx * sl2) {
+                            L[i] = (int8_t) new_l;
+                            sumlx = slx;
+                            suml2 = sl2;
+                            ++n_changed;
+                        }
+                    }
+                }
+            }
+            if (!n_changed) {
+                break;
+            }
+        }
+
+        for (int i = 0; i < n; ++i) {
+            L[i] += nmax;
+        }
+        return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
+    }
+
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int_ggml_sycl(iscale * x[i]);
+        l = dpct::max(-nmax, dpct::min(nmax - 1, l));
+        L[i] = (int8_t) (l + nmax);
+    }
+
+    return 1.0f / iscale;
+}
+
+inline void set_scale_min_k4(int j, uint8_t * q, uint8_t d, uint8_t m) {
+    if (j < 4) {
+        q[j]     = (q[j] & 0xC0) | (d & 0x3F);
+        q[j + 4] = (q[j + 4] & 0xC0) | (m & 0x3F);
+    } else {
+        q[j + 4] = (d & 0x0F) | ((m & 0x0F) << 4);
+        q[j - 4] = (q[j - 4] & 0x3F) | ((d >> 4) << 6);
+        q[j - 0] = (q[j - 0] & 0x3F) | ((m >> 4) << 6);
+    }
+}
+
+inline void get_scale_min_k4_local(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63;
+        m = q[j + 4] & 63;
+    } else {
+        d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
+        m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
+    }
+}
+
+inline void cpy_blck_f32_mxfp4(const char * cxi, char * cdsti) {
+    const float *   xi   = (const float *) cxi;
+    block_mxfp4 *   dsti = (block_mxfp4 *) cdsti;
+
+    float amax = 0.0f;
+    for (int j = 0; j < QK_MXFP4; ++j) {
+        amax = sycl::fmax(amax, sycl::fabs((float) xi[j]));
+    }
+
+    const uint8_t e = amax > 0.0f ? (uint8_t) (sycl::floor(sycl::log2(amax)) - 2 + 127) : 0;
+    const float d = GGML_E8M0_TO_FP32_HALF(e);
+
+    dsti->e = e;
+
+    for (int j = 0; j < QK_MXFP4 / 2; ++j) {
+        const uint8_t x0 = best_index_mxfp4(xi[0 + j], d);
+        const uint8_t x1 = best_index_mxfp4(xi[QK_MXFP4 / 2 + j], d);
+
+        dsti->qs[j]  = x0;
+        dsti->qs[j] |= x1 << 4;
+    }
+}
+
+inline void cpy_blck_f32_nvfp4(const char * cxi, char * cdsti) {
+    const float *   xi   = (const float *) cxi;
+    block_nvfp4 *   dsti = (block_nvfp4 *) cdsti;
+
+    constexpr int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+    for (int s = 0; s < n_sub; ++s) {
+        const float * xb = xi + s * QK_NVFP4_SUB;
+
+        float amax = 0.0f;
+        for (int j = 0; j < QK_NVFP4_SUB; ++j) {
+            amax = sycl::fmax(amax, sycl::fabs((float) xb[j]));
+        }
+
+        const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
+        dsti->d[s] = ue;
+        const float d = ggml_sycl_ue4m3_to_fp32(ue);
+
+        for (int j = 0; j < QK_NVFP4_SUB / 2; ++j) {
+            const uint8_t x0 = best_index_mxfp4(xb[0 + j], d);
+            const uint8_t x1 = best_index_mxfp4(xb[QK_NVFP4_SUB / 2 + j], d);
+
+            dsti->qs[s * (QK_NVFP4_SUB / 2) + j] = x0 | (x1 << 4);
+        }
+    }
+}
+
+
 inline void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
     const float * xi   = (const float *) cxi;
     block_q4_0 *  dsti = (block_q4_0 *) cdsti;

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

@@ -0,0 +1,255 @@
+#include "cross_entropy_loss.hpp"
+
+#include <cstdint>
+#include <cmath>
+
+template <bool has_shared>
+static __dpct_inline__ void cross_entropy_loss_f32_kernel(
+        const float * __restrict__ logits,
+        const float * __restrict__ labels,
+        float * __restrict__ row_loss,
+        const int nclasses,
+        const int nrows,
+        float * __restrict__ smem,
+        const sycl::nd_item<3> & item) {
+
+    const int row = item.get_group(2);
+    const int tid = item.get_local_id(2);
+
+    logits += (int64_t) row * nclasses;
+    labels += (int64_t) row * nclasses;
+
+    float max_logit = -INFINITY;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = logits[i];
+        max_logit = sycl::fmax(max_logit, v);
+        if (has_shared) {
+            smem[i] = v;
+        }
+    }
+    max_logit = warp_reduce_max<WARP_SIZE>(max_logit);
+
+    float sum_exp = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = has_shared ? smem[i] : logits[i];
+        sum_exp += sycl::exp(v - max_logit);
+    }
+    sum_exp = warp_reduce_sum<WARP_SIZE>(sum_exp);
+    const float log_sum = sycl::log(sum_exp);
+
+    float loss = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = has_shared ? smem[i] : logits[i];
+        loss += (v - max_logit - log_sum) * labels[i];
+    }
+    loss = -warp_reduce_sum<WARP_SIZE>(loss) / (float) nrows;
+
+    if (tid == 0) {
+        row_loss[row] = loss;
+    }
+}
+
+template <bool has_shared>
+static __dpct_inline__ void cross_entropy_loss_back_f32_kernel(
+        const float * __restrict__ grad,
+        const float * __restrict__ logits,
+        const float * __restrict__ labels,
+        float * __restrict__ dst,
+        const int nclasses,
+        const int nrows,
+        float * __restrict__ smem,
+        const sycl::nd_item<3> & item) {
+
+    const int row = item.get_group(2);
+    const int tid = item.get_local_id(2);
+
+    logits += (int64_t) row * nclasses;
+    labels += (int64_t) row * nclasses;
+    dst    += (int64_t) row * nclasses;
+
+    float max_logit = -INFINITY;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = logits[i];
+        max_logit = sycl::fmax(max_logit, v);
+        if (has_shared) {
+            smem[i] = v;
+        }
+    }
+    max_logit = warp_reduce_max<WARP_SIZE>(max_logit);
+
+    float sum_exp = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = sycl::exp((has_shared ? smem[i] : logits[i]) - max_logit);
+        sum_exp += v;
+        if (has_shared) {
+            smem[i] = v;
+        } else {
+            dst[i] = v;
+        }
+    }
+    sum_exp = warp_reduce_sum<WARP_SIZE>(sum_exp);
+    const float inv_sum = 1.0f / sum_exp;
+
+    const float d_by_nrows = grad[0] / (float) nrows;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float sm_num = has_shared ? smem[i] : dst[i];
+        dst[i] = (sm_num * inv_sum - labels[i]) * d_by_nrows;
+    }
+}
+
+static void cross_entropy_reduce_rows(
+        ggml_backend_sycl_context & ctx,
+        const float * row_loss,
+        float * dst,
+        const int64_t nrows) {
+    if (nrows == 1) {
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            ctx.stream()->memcpy(dst, row_loss, sizeof(float))));
+        return;
+    }
+
+    ggml_sycl_pool_alloc<float> tmp_alloc(ctx.pool(), nrows);
+    float * tmp = tmp_alloc.get();
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        ctx.stream()->memcpy(tmp, row_loss, nrows * sizeof(float))));
+
+    int64_t cur = nrows;
+    while (cur > 1) {
+        const int64_t out = (cur + WARP_SIZE - 1) / WARP_SIZE;
+        const sycl::range<3> block(1, 1, WARP_SIZE);
+        const sycl::range<3> grid(1, 1, out);
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                const int row = item.get_group(2);
+                const int tid = item.get_local_id(2);
+                const int64_t i = (int64_t) row * WARP_SIZE + tid;
+                float v = i < cur ? tmp[i] : 0.0f;
+                v = warp_reduce_sum<WARP_SIZE>(v);
+                if (tid == 0) {
+                    tmp[row] = v;
+                }
+            });
+        cur = out;
+    }
+
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        ctx.stream()->memcpy(dst, tmp, sizeof(float))));
+}
+
+void ggml_sycl_cross_entropy_loss(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_scalar(dst));
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int64_t nclasses = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * logits_d = (const float *) src0->data;
+    const float * labels_d = (const float *) src1->data;
+    float * dst_d = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> row_loss_alloc(ctx.pool(), nrows);
+    float * row_loss = row_loss_alloc.get();
+
+    const sycl::range<3> block(1, 1, WARP_SIZE);
+    const sycl::range<3> grid(1, 1, nrows);
+    const size_t nbytes_shared = (size_t) nclasses * sizeof(float);
+    const size_t smpbo = ggml_sycl_info().devices[ctx.device].smpbo;
+
+    if (nbytes_shared <= smpbo) {
+        ctx.stream()->submit([&](sycl::handler & cgh) {
+            sycl::local_accessor<float, 1> smem(sycl::range<1>(nclasses), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid * block, block),
+                [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    cross_entropy_loss_f32_kernel<true>(
+                        logits_d, labels_d, row_loss,
+                        (int) nclasses, (int) nrows,
+                        get_pointer(smem), item);
+                });
+        });
+    } else {
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                cross_entropy_loss_f32_kernel<false>(
+                    logits_d, labels_d, row_loss,
+                    (int) nclasses, (int) nrows,
+                    nullptr, item);
+            });
+    }
+
+    cross_entropy_reduce_rows(ctx, row_loss, dst_d, nrows);
+}
+
+void ggml_sycl_cross_entropy_loss_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/3);
+
+    const ggml_tensor * grad  = dst->src[0];
+    const ggml_tensor * src0f = dst->src[1];
+    const ggml_tensor * src1f = dst->src[2];
+
+    GGML_ASSERT(grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0f->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1f->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_scalar(grad));
+    GGML_ASSERT(ggml_is_contiguous(grad));
+    GGML_ASSERT(ggml_is_contiguous(src0f));
+    GGML_ASSERT(ggml_is_contiguous(src1f));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0f, src1f));
+    GGML_ASSERT(ggml_are_same_shape(src0f, dst));
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int64_t nclasses = src0f->ne[0];
+    const int64_t nrows = ggml_nrows(src0f);
+
+    const float * grad_d  = (const float *) grad->data;
+    const float * logits_d = (const float *) src0f->data;
+    const float * labels_d = (const float *) src1f->data;
+    float * dst_d = (float *) dst->data;
+
+    const sycl::range<3> block(1, 1, WARP_SIZE);
+    const sycl::range<3> grid(1, 1, nrows);
+    const size_t nbytes_shared = (size_t) nclasses * sizeof(float);
+    const size_t smpbo = ggml_sycl_info().devices[ctx.device].smpbo;
+
+    if (nbytes_shared <= smpbo) {
+        ctx.stream()->submit([&](sycl::handler & cgh) {
+            sycl::local_accessor<float, 1> smem(sycl::range<1>(nclasses), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid * block, block),
+                [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    cross_entropy_loss_back_f32_kernel<true>(
+                        grad_d, logits_d, labels_d, dst_d,
+                        (int) nclasses, (int) nrows,
+                        get_pointer(smem), item);
+                });
+        });
+    } else {
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                cross_entropy_loss_back_f32_kernel<false>(
+                    grad_d, logits_d, labels_d, dst_d,
+                    (int) nclasses, (int) nrows,
+                    nullptr, item);
+            });
+    }
+}

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

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.hpp"
+
+void ggml_sycl_cross_entropy_loss(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_cross_entropy_loss_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

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);