whispercpp 1.3.7 → 1.3.8

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

@@ -2,8 +2,10 @@
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/presets.hpp"
 
-static void 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, sycl::float2* s_sum, int block_size) {
+static void norm_f32(const float* x, float* dst, const int ncols,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
 
     const int nrows = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -16,16 +18,16 @@ static void norm_f32(const float* x, float* dst, const int ncols, const int64_t
     const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
 
-    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
-    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+    const auto src_offset = calculate_offset<3>({src_stride_sample, src_stride_channel, src_stride_row}, {sample, channel, row});
+    const auto dst_offset = calculate_offset<3>({dst_stride_sample, dst_stride_channel, dst_stride_row}, {sample, channel, row});
 
-    x += strided_offset;
-    dst += packed_offset;
+    x += src_offset;
+    dst += dst_offset;
 
     sycl::float2 mean_var = sycl::float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         mean_var.x() += xi;
         mean_var.y() += xi * xi;
     }
@@ -54,7 +56,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const int64_t
     const float inv_std = sycl::rsqrt(var + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = (x[col] - mean) * inv_std;
+        dst[col * dst_stride_col] = (x[col * src_stride_col] - mean) * inv_std;
     }
 }
 
@@ -145,8 +147,10 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     }
 }
 
-static void rms_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, int block_size) {
+static void rms_norm_f32(const float* x, float* dst, const int ncols,
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
 
     const int nrows = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -160,17 +164,17 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const int6
     const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
 
-    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
-    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+    const auto src_offset = calculate_offset<3>({src_stride_sample, src_stride_channel, src_stride_row}, {sample, channel, row});
+    const auto dst_offset = calculate_offset<3>({dst_stride_sample, dst_stride_channel, dst_stride_row}, {sample, channel, row});
 
-    x   += strided_offset;
-    dst += packed_offset;
+    x   += src_offset;
+    dst += dst_offset;
 
 
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         tmp += xi * xi;
     }
 
@@ -198,14 +202,15 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const int6
     const float scale = sycl::rsqrt(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        dst[col * dst_stride_col] = scale * x[col * src_stride_col];
     }
 }
 
 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 int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel,
+    const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,
+    const int64_t dst_stride_channel, const int64_t dst_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);
@@ -215,13 +220,13 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
     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;
+    x   += sample*src_stride_sample + channel*src_stride_channel + row*src_stride_row;
+    dst += sample*dst_stride_sample + channel*dst_stride_channel + row*dst_stride_row;
 
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[col];
+        const float xi = x[col * src_stride_col];
         tmp += xi * xi;
     }
 
@@ -229,12 +234,13 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
     const float scale = sycl::rsqrt(sycl::fmax(tmp, eps * eps));
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col] = scale * x[col];
+        dst[col * dst_stride_col] = scale * x[col * src_stride_col];
     }
 }
 
 static void 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 int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
         const float eps, queue_ptr stream, int device) {
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -245,7 +251,10 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                    norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -265,7 +274,10 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                    norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -319,7 +331,9 @@ 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) {
+    const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+    const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+    const float eps, queue_ptr stream, int device) {
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -330,7 +344,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                    rms_norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -350,7 +367,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
                 sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                    rms_norm_f32(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -363,9 +383,14 @@ static void l2_norm_f32_sycl(const float *   x,
                              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 int64_t   src_stride_col,
+                             const int64_t   src_stride_row,
+                             const int64_t   src_stride_channel,
+                             const int64_t   src_stride_sample,
+                             const int64_t   dst_stride_col,
+                             const int64_t   dst_stride_row,
+                             const int64_t   dst_stride_channel,
+                             const int64_t   dst_stride_sample,
                              const float     eps,
                              queue_ptr       stream,
                              int             device) {
@@ -379,7 +404,10 @@ static void l2_norm_f32_sycl(const float *   x,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(warp_size)]] {
-                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                    l2_norm_f32<warp_size>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1,
                         nullptr, warp_size);
                 });
             });
@@ -398,7 +426,9 @@ static void l2_norm_f32_sycl(const float *   x,
                     block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(warp_size)]] {
-                    l2_norm_f32<warp_size>(x, dst, ncols, stride_row, stride_channel, stride_sample,
+                    l2_norm_f32<warp_size>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
                         eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
@@ -421,12 +451,20 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     memcpy(&eps, dst->op_params, sizeof(float));
     GGML_ASSERT(eps >= 0.0f);
     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;
-
-    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
+
+    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03,
+        ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
@@ -465,11 +503,19 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
     GGML_TENSOR_UNARY_OP_LOCALS
     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;
-    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
+    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03,
+        ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -644,13 +690,21 @@ void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(eps >= 0.0f);
 
     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;
+    const size_t tdst = ggml_type_size(dst->type);
+    GGML_ASSERT(nb00 % ts0 == 0 && nb01 % ts0 == 0 && nb02 % ts0 == 0 && nb03 % ts0 == 0);
+    GGML_ASSERT(nb0 % tdst == 0 && nb1 % tdst == 0 && nb2 % tdst == 0 && nb3 % tdst == 0);
+    const int64_t ss0 = nb00 / ts0;
+    const int64_t ss1 = nb01 / ts0;
+    const int64_t ss2 = nb02 / ts0;
+    const int64_t ss3 = nb03 / ts0;
+    const int64_t ds0 = nb0 / tdst;
+    const int64_t ds1 = nb1 / tdst;
+    const int64_t ds2 = nb2 / tdst;
+    const int64_t ds3 = nb3 / tdst;
 
     /*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);
+    l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03,
+            ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, stream, ctx.device);
 }

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

@@ -1,11 +1,12 @@
 #include "outprod.hpp"
+#include "convert.hpp"
 
 void ggml_sycl_op_out_prod(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(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_Q1_0);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
@@ -20,11 +21,31 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(ne01 == ne11);  // Inner dimensions must match
     GGML_ASSERT(ne0 == ne00);   // Output rows match src0 rows
     GGML_ASSERT(ne1 == ne10);   // Output cols match src1 cols
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+    GGML_ASSERT(ne2 % ne02 == 0);
+    GGML_ASSERT(ne3 % ne03 == 0);
 
     // Get data pointers
-    const float* src0_d = (const float*)src0->data;
-    const float* src1_d = (const float*)src1->data;
-    float* dst_d = (float*)dst->data;
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float * dst_d = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> src0_as_f32(ctx.pool());
+    int64_t src0_nb02 = nb02;
+    int64_t src0_nb03 = nb03;
+    if (src0->type == GGML_TYPE_Q1_0) {
+        scope_op_debug_print scope_dbg_print(__func__, "/to_fp32_sycl", dst, /*num_src=*/2,
+                                             " : converting src0 Q1_0 to fp32");
+        src0_d = src0_as_f32.alloc(ne00 * ne01 * ne02 * ne03);
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type, dst);
+        GGML_ASSERT(to_fp32_sycl != nullptr);
+        to_fp32_sycl(src0->data, const_cast<float *>(src0_d), ne00 * ne01 * ne02 * ne03, stream);
+
+        // Dequantized src0 buffer is contiguous fp32 [ne00, ne01, ne02, ne03].
+        src0_nb02 = ne00 * ne01 * (int64_t) sizeof(float);
+        src0_nb03 = ne00 * ne01 * ne02 * (int64_t) sizeof(float);
+    }
 
     // GEMM parameters
     const float alpha = 1.0f;
@@ -35,12 +56,27 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     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);
 
+    const int64_t r2 = ne2 / ne02;
+    const int64_t r3 = ne3 / ne03;
+
     try {
-        // 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) {
+        // OUT_PROD applies independently to each (i2, i3) destination plane.
+        for (int64_t i3 = 0; i3 < ne3; ++i3) {
+            for (int64_t i2 = 0; i2 < ne2; ++i2) {
+                const int64_t i03 = i3 / r3;
+                const int64_t i02 = i2 / r2;
+
+                const float * src0_plane = (const float *) ((const char *) src0_d + i02 * src0_nb02 + i03 * src0_nb03);
+                const float * src1_plane = (const float *) ((const char *) src1_d + i2 * nb12 + i3 * nb13);
+                float * dst_plane = (float *) ((char *) dst_d + i2 * nb2 + i3 * nb3);
+
+                // Perform matrix multiplication using oneMKL GEMM
+                oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op,
+                                                      ne0, ne1, ne01, alpha, src0_plane, ne00,
+                                                      src1_plane, ldb, beta, dst_plane, ne0);
+            }
+        }
+    } catch (sycl::exception const& exc) {
         std::cerr << exc.what() << std::endl;
         GGML_ASSERT(false);
     }

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

@@ -0,0 +1,185 @@
+//
+// MIT license
+// Copyright (C) 2026 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 "pool.hpp"
+#include <float.h>
+
+template <typename Ti, typename To>
+static void pool2d_nchw_kernel(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const Ti* src, To* dst, const enum ggml_op_pool op,
+        const sycl::nd_item<3> &item_ct1) {
+        int idx = item_ct1.get_local_id(2) +
+                  item_ct1.get_group(2) * item_ct1.get_local_range(2);
+        if (idx >= parallel_elements) {
+            return;
+        }
+
+        const int I_HW = ih * iw;
+        const int O_HW = oh * ow;
+        const int nc = idx / O_HW;
+        const int cur_oh = idx % O_HW / ow;
+        const int cur_ow = idx % O_HW % ow;
+        const Ti* i_ptr = src + nc * I_HW;
+        To* o_ptr = dst + nc * O_HW;
+        const int start_h = cur_oh * sh - ph;
+        const int bh = sycl::max(0, start_h);
+        const int eh = sycl::min(ih, start_h + kh);
+        const int start_w = cur_ow * sw - pw;
+        const int bw = sycl::max(0, start_w);
+        const int ew = sycl::min(iw, start_w + kw);
+
+        To res = 0;
+
+        switch (op) {
+            case GGML_OP_POOL_AVG: res = 0; break;
+            case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+            default:
+                res      = (To) sycl::nan(uint32_t(0));
+                break;
+        }
+
+        for (int i = bh; i < eh; i += 1) {
+            for (int j = bw; j < ew; j += 1) {
+                Ti cur = i_ptr[i * iw + j];
+                switch (op) {
+                    case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
+                    case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
+                    default:
+                        res = (To) sycl::nan(uint32_t(0));
+                        break;
+                }
+            }
+        }
+        o_ptr[cur_oh * ow + cur_ow] = res;
+}
+
+template <typename Ti, typename To>
+static void pool1d_ncw_kernel(
+        const int iw, const int ow,
+        const int k, const int s,
+        const int p, const int parallel_elements,
+        const Ti * src, To * dst, const enum ggml_op_pool op,
+        const sycl::nd_item<3> & item_ct1) {
+    int idx = item_ct1.get_local_id(2) +
+              item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int nc     = idx / ow;
+    const int cur_ow = idx % ow;
+    const Ti * i_ptr = src + nc * iw;
+    To *       o_ptr = dst + nc * ow;
+    const int  start = cur_ow * s - p;
+    const int  b     = sycl::max(0, start);
+    const int  e     = sycl::min(iw, start + k);
+
+    To res = 0;
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0;        break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default:
+            res = (To) sycl::nan(uint32_t(0));
+            break;
+    }
+
+    for (int j = b; j < e; j += 1) {
+        Ti cur = i_ptr[j];
+        switch (op) {
+            case GGML_OP_POOL_AVG: res += cur; break;
+            case GGML_OP_POOL_MAX: res = sycl::max(res, (To) cur); break;
+            default:
+                res = (To) sycl::nan(uint32_t(0));
+                break;
+        }
+    }
+
+    const int count = e - b;
+    if (op == GGML_OP_POOL_AVG) {
+        res = (count > 0) ? (res / count) : (To) 0;
+    }
+    o_ptr[cur_ow] = res;
+}
+
+void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    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));
+    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
+    float *       dst_dd  = static_cast<float *>(dst->data);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    const int64_t IH = dst->src[0]->ne[1];
+    const int64_t IW = dst->src[0]->ne[0];
+
+    const int64_t N = dst->ne[3];
+    const int64_t OC = dst->ne[2];
+    const int64_t OH = dst->ne[1];
+    const int64_t OW = dst->ne[0];
+
+    const int parallel_elements = N * OC * OH * OW;
+    const int num_blocks = (parallel_elements + SYCL_POOL2D_BLOCK_SIZE - 1) / SYCL_POOL2D_BLOCK_SIZE;
+    sycl::range<3> block_nums(1, 1, num_blocks);
+    main_stream->parallel_for(
+        sycl::nd_range<3>(block_nums *
+                              sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pool2d_nchw_kernel(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0,
+                               parallel_elements, src0_dd, dst_dd, op,
+                               item_ct1);
+        });
+}
+
+void ggml_sycl_op_pool1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    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));
+    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
+    float *       dst_dd  = static_cast<float *>(dst->data);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int s0 = opts[2];
+    const int p0 = opts[3];
+
+    const int64_t IW = dst->src[0]->ne[0];
+    const int64_t OW = dst->ne[0];
+    const int64_t NC = dst->ne[3] * dst->ne[2] * dst->ne[1];
+
+    const int parallel_elements = NC * OW;
+    const int num_blocks = (parallel_elements + SYCL_POOL1D_BLOCK_SIZE - 1) / SYCL_POOL1D_BLOCK_SIZE;
+    sycl::range<3> block_nums(1, 1, num_blocks);
+    main_stream->parallel_for(
+        sycl::nd_range<3>(block_nums *
+                              sycl::range<3>(1, 1, SYCL_POOL1D_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_POOL1D_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pool1d_ncw_kernel(IW, OW, k0, s0, p0,
+                              parallel_elements, src0_dd, dst_dd, op,
+                              item_ct1);
+        });
+}

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

@@ -0,0 +1,22 @@
+//
+// MIT license
+// Copyright (C) 2026 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_POOL_HPP
+#define GGML_SYCL_POOL_HPP
+
+#include "common.hpp"
+#include "presets.hpp"
+
+void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_op_pool1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_POOL_HPP

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

@@ -19,6 +19,7 @@
 #define WARP_SIZE GGML_SYCL_WARP_SIZE
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
+#define SYCL_COL2IM_1D_BLOCK_SIZE 256
 #define SYCL_GELU_BLOCK_SIZE 256
 #define SYCL_SILU_BLOCK_SIZE 256
 #define SYCL_TANH_BLOCK_SIZE 256
@@ -46,6 +47,7 @@
 #define SYCL_PAD_BLOCK_SIZE 256
 #define SYCL_ACC_BLOCK_SIZE 256
 #define SYCL_IM2COL_BLOCK_SIZE 256
+#define SYCL_POOL1D_BLOCK_SIZE 256
 #define SYCL_POOL2D_BLOCK_SIZE 256
 #define SYCL_ARGMAX_BLOCK_SIZE 256
 #define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
@@ -61,7 +63,7 @@
 #endif
 
 #ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
+#define K_QUANTS_PER_ITERATION 1
 #else
 static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
 #endif

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

@@ -135,7 +135,7 @@ static void set_rows_sycl(
 
     stream->parallel_for(
         sycl::nd_range<1>(grid_size * block_size, block_size),
-        [=](sycl::nd_item<1> item_ct1) {
+        [=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
             k_set_rows<TIn, TIdx, TOut>(
                 src0_d, src1_d, dst_d,
                 ne00, ne01, ne02,
@@ -202,6 +202,9 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
         case GGML_TYPE_Q8_0:
             set_rows_sycl_q<TIdx, block_q8_0, QK8_0, cpy_blck_f32_q8_0>(src0_d, src1_d, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q1_0:
+            set_rows_sycl_q<TIdx, block_q1_0, QK1_0, cpy_blck_f32_q1_0>(src0_d, src1_d, (block_q1_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_Q5_1:
             set_rows_sycl_q<TIdx, block_q5_1, QK5_1, cpy_blck_f32_q5_1>(src0_d, src1_d, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
             break;
@@ -217,7 +220,12 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
         case GGML_TYPE_IQ4_NL:
             set_rows_sycl_q<TIdx, block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>(src0_d, src1_d, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
             break;
-
+        case GGML_TYPE_MXFP4:
+            set_rows_sycl_q<TIdx, block_mxfp4, QK_MXFP4, cpy_blck_f32_mxfp4>(src0_d, src1_d, (block_mxfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_NVFP4:
+            set_rows_sycl_q<TIdx, block_nvfp4, QK_NVFP4, cpy_blck_f32_nvfp4>(src0_d, src1_d, (block_nvfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
         default:
             GGML_ABORT("Unsupported tensor type!");
             break;