whispercpp 1.3.6 → 1.3.8

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

@@ -24,4 +24,37 @@ void ggml_sycl_op_mul_mat_vec_q(
     const int64_t src1_ncols, const int64_t src1_padded_row_size,
     const dpct::queue_ptr &stream);
 
+// Requires standard (non-reorder) block layout for src0.
+// Returns false if src0_type isn't handled; caller should fall back.
+bool ggml_sycl_mul_mat_vec_q_id(
+    enum ggml_type     src0_type,
+    const void *       vx_base,             // start of stacked expert weights
+    const void *       vy,                  // pre-quantized src1 (Q8_1)
+    const int32_t *    ids_dev,             // device-side int32, length n_experts_used
+    float *            dst_base,
+    int                ncols,
+    int                nrows,
+    int                n_experts_used,
+    size_t             expert_weight_stride, // bytes between experts in vx_base
+    size_t             dst_row_stride,       // bytes between dst rows
+    size_t             src1_row_stride,      // 0 = shared src1, else per-expert stride in bytes
+    dpct::queue_ptr    stream);
+
+// Reorder (SoA) variant of the fused MoE expert GEMV.
+// vx_base: each expert slice (stride expert_weight_stride == src0->nb[2]) is a self-contained reorder/SoA layout.
+// vy: src1 quantized with quantize_and_reorder_q8_1_soa (per-row SoA). Returns false if src0_type isn't handled.
+bool ggml_sycl_mul_mat_vec_q_id_reorder(
+    enum ggml_type     src0_type,
+    const void *       vx_base,
+    const void *       vy,
+    const int32_t *    ids_dev,
+    float *            dst_base,
+    int                ncols,
+    int                nrows,
+    int                n_experts_used,
+    size_t             expert_weight_stride,
+    size_t             dst_row_stride,
+    size_t             src1_row_stride,
+    dpct::queue_ptr    stream);
+
 #endif // GGML_SYCL_MMVQ_HPP

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/pad.cpp

@@ -13,7 +13,8 @@
 //#include "common.hpp"
 #include "pad.hpp"
 
-static void pad_f32(const float * src, float * dst,
+static void pad_f32(const float * src, size_t s00, size_t s01, size_t s02, size_t s03,
+                    float * dst,
                     const int lp0, const int rp0, const int lp1, const int rp1,
                     const int lp2, const int rp2, const int lp3, const int rp3,
                     const int ne0, const int ne1, const int ne2, const int ne3,
@@ -27,7 +28,6 @@ static void pad_f32(const float * src, float * dst,
         return;
     }
 
-    // operation
     const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
     if ((i0 >= lp0 && i0 < ne0 - rp0) &&
         (i1 >= lp1 && i1 < ne1 - rp1) &&
@@ -37,12 +37,8 @@ static void pad_f32(const float * src, float * dst,
         const int64_t i01 = i1 - lp1;
         const int64_t i02 = i2 - lp2;
         const int64_t i03 = i3 - lp3;
-        const int64_t ne02 = ne2 - lp2 - rp2;
-        const int64_t ne01 = ne1 - lp1 - rp1;
-        const int64_t ne00 = ne0 - lp0 - rp0;
 
-        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) +
-                                i02 * (ne00 * ne01) + i01 * ne00 + i00;
+        const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
 
         dst[dst_idx] = src[src_idx];
     } else {
@@ -50,20 +46,19 @@ static void pad_f32(const float * src, float * dst,
     }
 }
 
-static void pad_f32_sycl(const float *src, float *dst, const int lp0,
-                         const int rp0, const int lp1, const int rp1,
-                         const int lp2, const int rp2, const int lp3,
-                         const int rp3, const int ne0, const int ne1,
-                         const int ne2, const int ne3,
+static void pad_f32_sycl(const float * src, size_t s00, size_t s01, size_t s02, size_t s03,
+                         float * dst, const int lp0, const int rp0, const int lp1, const int rp1,
+                         const int lp2, const int rp2, const int lp3, const int rp3,
+                         const int ne0, const int ne1, const int ne2, const int ne3,
                          dpct::queue_ptr stream) {
     int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
-    dpct::dim3 gridDim(num_blocks, ne1, ne2 * ne3);
+    sycl::range<3> grid(ne2 * ne3, ne1, num_blocks);
     stream->parallel_for(
-        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+        sycl::nd_range<3>(grid * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
         [=](sycl::nd_item<3> item_ct1) {
-            pad_f32(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1,
-                    ne2, ne3, item_ct1);
+            pad_f32(src, s00, s01, s02, s03, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
+                    ne0, ne1, ne2, ne3, item_ct1);
         });
 }
 
@@ -71,22 +66,27 @@ void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
     float * dst_d = (float *)dst->data;
-    dpct::queue_ptr     stream = ctx.stream();
+    dpct::queue_ptr stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
 
-    const int32_t lp0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t rp0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t lp1 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t rp1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t lp2 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t rp2 = ((const int32_t*)(dst->op_params))[5];
-    const int32_t lp3 = ((const int32_t*)(dst->op_params))[6];
-    const int32_t rp3 = ((const int32_t*)(dst->op_params))[7];
+    const size_t ts = ggml_type_size(src0->type);
+    const size_t s00 = src0->nb[0] / ts;
+    const size_t s01 = src0->nb[1] / ts;
+    const size_t s02 = src0->nb[2] / ts;
+    const size_t s03 = src0->nb[3] / ts;
 
-    pad_f32_sycl(src0_d, dst_d,
+    const int32_t lp0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t rp0 = ((const int32_t *)(dst->op_params))[1];
+    const int32_t lp1 = ((const int32_t *)(dst->op_params))[2];
+    const int32_t rp1 = ((const int32_t *)(dst->op_params))[3];
+    const int32_t lp2 = ((const int32_t *)(dst->op_params))[4];
+    const int32_t rp2 = ((const int32_t *)(dst->op_params))[5];
+    const int32_t lp3 = ((const int32_t *)(dst->op_params))[6];
+    const int32_t rp3 = ((const int32_t *)(dst->op_params))[7];
+
+    pad_f32_sycl(src0_d, s00, s01, s02, s03, dst_d,
                  lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
                  dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
 }