vendor : update cpp-httplib to 0.44.0 (#22919)

* add im2col_3d

* format code

* update the ops.md

docs/ops.md

@@ -61,13 +61,14 @@ Legend:
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4159,6 +4159,11 @@ static void ggml_sycl_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     ggml_sycl_op_im2col(ctx, dst);
 }
 
+static void ggml_sycl_im2col_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_im2col_3d(ctx, dst);
+}
+
 static void ggml_sycl_sum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
@@ -4456,6 +4461,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_IM2COL:
             ggml_sycl_im2col(ctx, dst);
             break;
+        case GGML_OP_IM2COL_3D:
+            ggml_sycl_im2col_3d(ctx, dst);
+            break;
         case GGML_OP_POOL_2D:
             ggml_sycl_pool2d(ctx, dst);
             break;
@@ -5175,6 +5183,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
         case GGML_OP_IM2COL:
+        case GGML_OP_IM2COL_3D:
         case GGML_OP_UPSCALE:
             return true;
         case GGML_OP_SUM:

ggml/src/ggml-sycl/im2col.cpp

@@ -1,6 +1,6 @@
 //
 // MIT license
-// Copyright (C) 2024 Intel Corporation
+// Copyright (C) 2026 Intel Corporation
 // SPDX-License-Identifier: MIT
 //
 
@@ -12,125 +12,389 @@
 
 #include "im2col.hpp"
 
-#include <sycl/sycl.hpp>
-#include <type_traits>  // For std::is_same_v
-
-#include "ggml.h"
+#define MAX_GRIDDIM_Z 65535
 
 template <typename T>
-static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_t offset_delta, int64_t IC, int64_t IW,
-                          int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
-                          int s0, int s1, int p0, int p1, int d0, int d1, const sycl::nd_item<3> & item_ct1) {
-    const int64_t work_group_size = item_ct1.get_local_range(2);
-    const int64_t global_id       = item_ct1.get_local_id(2) + (work_group_size * item_ct1.get_group(2));
+static  void im2col_kernel(
+        const float * x, T * dst,
+        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
+        int64_t IC_IH_IW, int64_t IH_IW, int64_t N_OH, int64_t KH_KW, int64_t IC_KH_KW,
+        int s0, int s1, int p0, int p1, int d0, int d1) {
+    auto          item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t i        = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (i >= IC_KH_KW) {
+        return;
+    }
 
-    // make each work-item deal with more elements since sycl global range can not exceed max int
-    for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
-        const int64_t ksize = OW * KH;
-        const int64_t kx    = i / ksize;
-        const int64_t kd    = kx * ksize;
-        const int64_t ky    = (i - kd) / OW;
-        const int64_t ix    = i % OW;
+    const int64_t iic = i / (KH_KW);
+    const int64_t rem = i - iic * KH_KW;
+    const int64_t ikh = rem / KW;
+    const int64_t ikw = rem - ikh * KW;
 
-        const int64_t oh    = item_ct1.get_group(1);
-        const int64_t batch = item_ct1.get_group(0) / IC;
-        const int64_t ic    = item_ct1.get_group(0) % IC;
+    const int64_t iow = item_ct1.get_group(1);
+    for (int64_t iz = item_ct1.get_group(0); iz < N_OH; iz += MAX_GRIDDIM_Z) {
+        const int64_t  in = iz / OH;
+        const int64_t  ioh = iz - in * OH;
 
-        const int64_t iiw = (ix * s0) + (kx * d0) - p0;
-        const int64_t iih = (oh * s1) + (ky * d1) - p1;
+        const int64_t iiw = iow * s0 + ikw * d0 - p0;
+        const int64_t iih = ioh * s1 + ikh * d1 - p1;
 
-        const int64_t offset_dst = (((batch * OH + oh) * OW + ix) * CHW) + (ic * (KW * KH) + ky * KW + kx);
+        const int64_t offset_dst =
+            ((in * OH + ioh) * OW + iow) * IC_KH_KW + iic * KH_KW + ikh * KW + ikw;
 
-        const int64_t offset_src_base = (ic * offset_delta) + (batch * batch_offset);
-        const int64_t offset_src      = offset_src_base + (iih * IW) + iiw;
-
-        const bool  out_of_bounds = (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW);
-        const float src_val       = out_of_bounds ? 0.0f : x[offset_src];
-
-        if constexpr (std::is_same_v<T, sycl::half>) {
-            dst[offset_dst] = sycl::half(src_val);
-        } else if constexpr (std::is_same_v<T, float>) {
-            dst[offset_dst] = src_val;
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+            dst[offset_dst] = 0.0f;
+        } else {
+            const int64_t offset_src = iic * IC_IH_IW + in * IH_IW;
+            dst[offset_dst] = x[offset_src + iih * IW + iiw];
         }
     }
+
+    GGML_UNUSED(IC);
+    GGML_UNUSED(KH);
 }
 
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
 template <typename T>
-static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
-                                 int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
-                                 int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
-    const int64_t parallel_elements = OW * KW * KH;
-    const int64_t num_blocks        = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
-
-    // decrease global range when it exceeds the max int
-    int64_t local_size = downsample_sycl_global_range(batch * IC * OH * num_blocks, SYCL_IM2COL_BLOCK_SIZE);
-
-    sycl::range<3> block_nums(batch * IC, OH, num_blocks);
-    sycl::range<3> local_range(1, 1, local_size);
-
-    const int64_t CHW = IC * KH * KW;
-
-    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
-        im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
-                         p0, p1, d0, d1, item_ct1);
-    });
+static void im2col_sycl(const float *   x,
+                        T *             dst,
+                        int64_t         IW,
+                        int64_t         IH,
+                        int64_t         OW,
+                        int64_t         OH,
+                        int64_t         KW,
+                        int64_t         KH,
+                        int64_t         IC,
+                        int64_t         N,
+                        int64_t         IC_IH_IW,
+                        int64_t         IH_IW,
+                        int             s0,
+                        int             s1,
+                        int             p0,
+                        int             p1,
+                        int             d0,
+                        int             d1,
+                        dpct::queue_ptr stream) {
+    const int64_t IC_KH_KW = IC * KH * KW;
+    const int64_t num_blocks = (IC_KH_KW + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+    const int64_t N_OH = N * OH;
+    const int64_t KH_KW = KW*KH;
+    dpct::dim3    block_nums(num_blocks, OW, MIN(N_OH, MAX_GRIDDIM_Z));
+    /*
+    DPCT1049:73: 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.
+    */
+    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
+                                           sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             im2col_kernel(x, dst, IC, IW, IH, OH, OW, KW, KH, IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
+                                           s0, s1, p0, p1, d0, d1);
+                         });
 }
 
-static void im2col_sycl_f16(const float * x, sycl::half * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH,
-                            int64_t KW, int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset,
-                            int64_t offset_delta, int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
-    if (!stream->get_device().has(sycl::aspect::fp16)) {
-        throw sycl::exception(sycl::make_error_code(sycl::errc::kernel_not_supported),
-                              "Device does not support half precision (fp16) operations!");
-    }
-    im2col_sycl_internal<sycl::half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0,
-                                     p1, d0, d1, stream);
+static void im2col_sycl_f16(const float *   x,
+                            sycl::half *    dst,
+                            int64_t         IW,
+                            int64_t         IH,
+                            int64_t         OW,
+                            int64_t         OH,
+                            int64_t         KW,
+                            int64_t         KH,
+                            int64_t         IC,
+                            int64_t         N,
+                            int64_t         IC_IH_IW,
+                            int64_t         IH_IW,
+                            int             s0,
+                            int             s1,
+                            int             p0,
+                            int             p1,
+                            int             d0,
+                            int             d1,
+                            dpct::queue_ptr stream) {
+    im2col_sycl<sycl::half>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
-static void im2col_sycl_f32(const float * x, float * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
-                            int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta, int s0,
-                            int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
-    im2col_sycl_internal<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1,
-                                d0, d1, stream);
+static void im2col_sycl_f32(const float *   x,
+                            float *         dst,
+                            int64_t         IW,
+                            int64_t         IH,
+                            int64_t         OW,
+                            int64_t         OH,
+                            int64_t         KW,
+                            int64_t         KH,
+                            int64_t         IC,
+                            int64_t         N,
+                            int64_t         IC_IH_IW,
+                            int64_t         IH_IW,
+                            int             s0,
+                            int             s1,
+                            int             p0,
+                            int             p1,
+                            int             d0,
+                            int             d1,
+                            dpct::queue_ptr stream) {
+    im2col_sycl<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
 void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr     stream = ctx.stream();
 
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 
-    const int32_t s0 = ((const int32_t *) (dst->op_params))[0];
-    const int32_t s1 = ((const int32_t *) (dst->op_params))[1];
-    const int32_t p0 = ((const int32_t *) (dst->op_params))[2];
-    const int32_t p1 = ((const int32_t *) (dst->op_params))[3];
-    const int32_t d0 = ((const int32_t *) (dst->op_params))[4];
-    const int32_t d1 = ((const int32_t *) (dst->op_params))[5];
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
 
-    const bool is_2D = ((const int32_t *) (dst->op_params))[6] == 1;
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
 
     const int64_t IC = src1->ne[is_2D ? 2 : 1];
     const int64_t IH = is_2D ? src1->ne[1] : 1;
-    const int64_t IW = src1->ne[0];
+    const int64_t IW =         src1->ne[0];
 
     const int64_t KH = is_2D ? src0->ne[1] : 1;
-    const int64_t KW = src0->ne[0];
+    const int64_t KW =         src0->ne[0];
 
     const int64_t OH = is_2D ? dst->ne[2] : 1;
-    const int64_t OW = dst->ne[1];
+    const int64_t OW =         dst->ne[1];
 
-    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / sizeof(float);
-    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
-    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / sizeof(float);
+    const int64_t IC_IH_IW = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t N        = src1->ne[is_2D ? 3 : 2];
+    const int64_t IH_IW    = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
 
-    queue_ptr stream = ctx.stream();
-
-    if (dst->type == GGML_TYPE_F16) {
-        im2col_sycl_f16((const float *) src1->data, (sycl::half *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
-                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    if(dst->type == GGML_TYPE_F16) {
+        im2col_sycl_f16(src1_d, (sycl::half *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1,
+                        d0, d1, stream);
     } else {
-        im2col_sycl_f32((const float *) src1->data, (float *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
-                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+        im2col_sycl_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
+    }
+}
+
+// [N*IC, ID, IH, IW] => [N*OD, OH, OW, IC * KD * KH * KW]
+template <typename T>
+static  void im2col_3d_kernel(
+        const float * src, T * dst,
+        int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW, int64_t OC,
+        int64_t KD, int64_t KH, int64_t KW, int64_t OD, int64_t OH, int64_t OW,
+        int64_t OH_OW, int64_t KD_KH_KW, int64_t ID_IH_IW, int64_t KH_KW, int64_t IH_IW, int64_t IC_ID_IH_IW,
+        int64_t IC_KD_KH_KW, int64_t OW_KD_KH_KW, int64_t OD_OH_OW_IC_KD_KH_KW, int64_t OH_OW_IC_KD_KH_KW,
+        int64_t OW_IC_KD_KH_KW, int64_t N_OD_OH, int64_t OD_OH,
+        int64_t stride_q, int64_t stride_z, int64_t stride_y, int64_t stride_x,
+        int s0, int s1, int s2, int p0, int p1, int p2, int d0, int d1, int d2) {
+    auto          item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t i        = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (i >= IC_KD_KH_KW) {
+        return;
+    }
+    GGML_UNUSED(N); GGML_UNUSED(OC); GGML_UNUSED(OH_OW); GGML_UNUSED(OD); GGML_UNUSED(OW); GGML_UNUSED(KD); GGML_UNUSED(KH);
+    GGML_UNUSED(ID_IH_IW); GGML_UNUSED(IH_IW); GGML_UNUSED(IC_ID_IH_IW); GGML_UNUSED(OW_KD_KH_KW);
+
+    const int64_t iic = i / KD_KH_KW;
+    const int64_t ikd = (i - iic * KD_KH_KW) / KH_KW;
+    const int64_t ikh = (i - iic * KD_KH_KW - ikd * KH_KW) / KW;
+    const int64_t ikw = i % KW;
+
+    const int64_t iow = item_ct1.get_group(1);
+    for (int64_t iz = item_ct1.get_group(0); iz < N_OD_OH; iz += MAX_GRIDDIM_Z) {
+        const int64_t in  = iz / OD_OH;
+        const int64_t iod = (iz - in*OD_OH) / OH;
+        const int64_t ioh = iz % OH;
+
+        const int64_t iiw = iow * s0 + ikw * d0 - p0;
+        const int64_t iih = ioh * s1 + ikh * d1 - p1;
+        const int64_t iid = iod * s2 + ikd * d2 - p2;
+
+        const int64_t offset_dst = in*OD_OH_OW_IC_KD_KH_KW + iod*OH_OW_IC_KD_KH_KW + ioh*OW_IC_KD_KH_KW + iow*IC_KD_KH_KW + iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw;
+
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
+            dst[offset_dst] = 0.0f;
+        } else {
+            const int64_t offset_src = ((in * IC + iic) * stride_q) + (iid * stride_z) + (iih * stride_y) + (iiw * stride_x);
+            dst[offset_dst] = src[offset_src];
+        }
+    }
+}
+
+// [N*IC, ID, IH, IW] => [N*OD, OH, OW, IC * KD * KH * KW]
+template <typename T>
+static void im2col_3d_sycl(const float *   src,
+                           T *             dst,
+                           int64_t         N,
+                           int64_t         IC,
+                           int64_t         ID,
+                           int64_t         IH,
+                           int64_t         IW,
+                           int64_t         OC,
+                           int64_t         KD,
+                           int64_t         KH,
+                           int64_t         KW,
+                           int64_t         OD,
+                           int64_t         OH,
+                           int64_t         OW,
+                           int64_t         stride_q,
+                           int64_t         stride_z,
+                           int64_t         stride_y,
+                           int64_t         stride_x,
+                           int             s0,
+                           int             s1,
+                           int             s2,
+                           int             p0,
+                           int             p1,
+                           int             p2,
+                           int             d0,
+                           int             d1,
+                           int             d2,
+                           dpct::queue_ptr stream) {
+    const int64_t OH_OW = OH*OW;
+    const int64_t KD_KH_KW = KD*KH*KW;
+    const int64_t ID_IH_IW = ID*IH*IW;
+    const int64_t KH_KW = KH*KW;
+    const int64_t IH_IW = IH*IW;
+    const int64_t IC_KD_KH_KW = IC*KD*KH*KW;
+    const int64_t OW_KD_KH_KW = OW*KD*KH*KW;
+    const int64_t N_OD_OH = N*OD*OH;
+    const int64_t OD_OH = OD*OH;
+    const int64_t IC_ID_IH_IW = IC*ID*IH*IW;
+    const int64_t OD_OH_OW_IC_KD_KH_KW = OD*OH*OW*IC*KD*KH*KW;
+    const int64_t OH_OW_IC_KD_KH_KW = OH*OW*IC*KD*KH*KW;
+    const int64_t OW_IC_KD_KH_KW = OW*IC*KD*KH*KW;
+    const int64_t num_blocks = (IC_KD_KH_KW + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+    dpct::dim3    block_nums(num_blocks, OW, MIN(N_OD_OH, MAX_GRIDDIM_Z));
+    /*
+    DPCT1049:74: 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.
+    */
+    stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
+                                           sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             im2col_3d_kernel(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, OH_OW, KD_KH_KW,
+                                              ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW, IC_KD_KH_KW, OW_KD_KH_KW,
+                                              OD_OH_OW_IC_KD_KH_KW, OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,
+                                              stride_q, stride_z, stride_y, stride_x, s0, s1, s2, p0, p1, p2, d0, d1,
+                                              d2);
+                         });
+}
+
+static void im2col_3d_sycl_f16(const float *   src,
+                               sycl::half *    dst,
+                               int64_t         N,
+                               int64_t         IC,
+                               int64_t         ID,
+                               int64_t         IH,
+                               int64_t         IW,
+                               int64_t         OC,
+                               int64_t         KD,
+                               int64_t         KH,
+                               int64_t         KW,
+                               int64_t         OD,
+                               int64_t         OH,
+                               int64_t         OW,
+                               int64_t         stride_q,
+                               int64_t         stride_z,
+                               int64_t         stride_y,
+                               int64_t         stride_x,
+                               int             s0,
+                               int             s1,
+                               int             s2,
+                               int             p0,
+                               int             p1,
+                               int             p2,
+                               int             d0,
+                               int             d1,
+                               int             d2,
+                               dpct::queue_ptr stream) {
+    im2col_3d_sycl<sycl::half>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, stride_q, stride_z, stride_y,
+                               stride_x, s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+}
+
+static void im2col_3d_sycl_f32(const float *   src,
+                               float *         dst,
+                               int64_t         N,
+                               int64_t         IC,
+                               int64_t         ID,
+                               int64_t         IH,
+                               int64_t         IW,
+                               int64_t         OC,
+                               int64_t         KD,
+                               int64_t         KH,
+                               int64_t         KW,
+                               int64_t         OD,
+                               int64_t         OH,
+                               int64_t         OW,
+                               int64_t         stride_q,
+                               int64_t         stride_z,
+                               int64_t         stride_y,
+                               int64_t         stride_x,
+                               int             s0,
+                               int             s1,
+                               int             s2,
+                               int             p0,
+                               int             p1,
+                               int             p2,
+                               int             d0,
+                               int             d1,
+                               int             d2,
+                               dpct::queue_ptr stream) {
+    im2col_3d_sycl<float>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                          stride_q, stride_z, stride_y, stride_x,
+                          s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+}
+
+void ggml_sycl_op_im2col_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr     stream = ctx.stream();
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+    const int32_t s2 = ((const int32_t *)(dst->op_params))[2];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[3];
+    const int32_t p1 = ((const int32_t *)(dst->op_params))[4];
+    const int32_t p2 = ((const int32_t *)(dst->op_params))[5];
+    const int32_t d0 = ((const int32_t *)(dst->op_params))[6];
+    const int32_t d1 = ((const int32_t *)(dst->op_params))[7];
+    const int32_t d2 = ((const int32_t *)(dst->op_params))[8];
+    const int32_t IC = ((const int32_t *)(dst->op_params))[9];
+
+    const int64_t N  = ne13 / IC;
+    const int64_t ID = ne12;
+    const int64_t IH = ne11;
+    const int64_t IW = ne10;
+
+    const int64_t OC = ne03 / IC;
+    const int64_t KD = ne02;
+    const int64_t KH = ne01;
+    const int64_t KW = ne00;
+
+    const int64_t OD = ne3 / N;
+    const int64_t OH = ne2;
+    const int64_t OW = ne1;
+
+    const size_t  es       = ggml_element_size(src1);
+    const int64_t stride_x = src1->nb[0] / es;
+    const int64_t stride_y = src1->nb[1] / es;
+    const int64_t stride_z = src1->nb[2] / es;
+    const int64_t stride_q = src1->nb[3] / es;
+
+    if(dst->type == GGML_TYPE_F16) {
+        im2col_3d_sycl_f16(src1_d, (sycl::half *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                           stride_q, stride_z, stride_y, stride_x,
+                           s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+    } else {
+        im2col_3d_sycl_f32(src1_d, (float *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                           stride_q, stride_z, stride_y, stride_x,
+                           s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
     }
 }

ggml/src/ggml-sycl/im2col.hpp

@@ -1,6 +1,6 @@
 //
 // MIT license
-// Copyright (C) 2024 Intel Corporation
+// Copyright (C) 2026 Intel Corporation
 // SPDX-License-Identifier: MIT
 //
 
@@ -15,7 +15,9 @@
 
 #include "common.hpp"
 
-void ggml_sycl_op_im2col(
-        ggml_backend_sycl_context & ctx, ggml_tensor *dst);
+#define SYCL_IM2COL_BLOCK_SIZE 256
+
+void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_op_im2col_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 #endif // GGML_SYCL_IM2COL_HPP

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.43.4"
+HTTPLIB_VERSION = "refs/tags/v0.44.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

vendor/cpp-httplib/httplib.cpp

@@ -1161,12 +1161,11 @@ bool parse_header(const char *beg, const char *end, T fn) {
 
     if (!detail::fields::is_field_value(val)) { return false; }
 
-    if (case_ignore::equal(key, "Location") ||
-        case_ignore::equal(key, "Referer")) {
-      fn(key, val);
-    } else {
-      fn(key, decode_path_component(val));
-    }
+    // RFC 9110 §5.5: header field values are opaque octets and MUST NOT be
+    // percent-decoded by the recipient. Applications that need to interpret a
+    // value as a URI component should call httplib::decode_uri_component()
+    // (or decode_path_component()) explicitly.
+    fn(key, val);
 
     return true;
   }
@@ -6192,9 +6191,29 @@ ThreadPool::ThreadPool(size_t n, size_t max_n, size_t mqr)
 #endif
   max_thread_count_ = max_n == 0 ? n : max_n;
   threads_.reserve(base_thread_count_);
-  for (size_t i = 0; i < base_thread_count_; i++) {
-    threads_.emplace_back(std::thread([this]() { worker(false); }));
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+  try {
+#endif
+    for (size_t i = 0; i < base_thread_count_; i++) {
+      threads_.emplace_back(std::thread([this]() { worker(false); }));
+    }
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+  } catch (...) {
+    // If thread creation fails partway (e.g., pthread_create returns EAGAIN),
+    // signal the workers we already spawned to exit and join them so the
+    // vector destructor does not see joinable threads (which would call
+    // std::terminate). Then rethrow so the caller learns of the failure.
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      shutdown_ = true;
+    }
+    cond_.notify_all();
+    for (auto &t : threads_) {
+      if (t.joinable()) { t.join(); }
+    }
+    throw;
   }
+#endif
 }
 
 bool ThreadPool::enqueue(std::function<void()> fn) {

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.43.4"
-#define CPPHTTPLIB_VERSION_NUM "0x002b04"
+#define CPPHTTPLIB_VERSION "0.44.0"
+#define CPPHTTPLIB_VERSION_NUM "0x002c00"
 
 #ifdef _WIN32
 #if defined(_WIN32_WINNT) && _WIN32_WINNT < 0x0A00