whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.cpp

@@ -0,0 +1,158 @@
+#include "conv2d-dw.hpp"
+
+struct conv2d_dw_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct conv2d_dw_kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+static inline conv2d_dw_kernel_bounds dw_calculate_kernel_bounds(int out_x, int out_y,
+                                                                  const conv2d_dw_params & p) {
+    conv2d_dw_kernel_bounds bounds;
+    bounds.y_min = sycl::max(0, (p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.y_max = sycl::min(p.kernel_h,
+                             (p.in_h + p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.x_min = sycl::max(0, (p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    bounds.x_max = sycl::min(p.kernel_w,
+                             (p.in_w + p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    return bounds;
+}
+
+static inline int dw_calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+// whcn layout: input/output stored as [N, C, H, W]
+struct dw_whcn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + c * p.in_w * p.in_h + y * p.in_w + x;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return c * p.kernel_h * p.kernel_w + ky * p.kernel_w + kx;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + c * p.out_w * p.out_h + y * p.out_w + x;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        out_x  = global_idx % p.out_w;
+        out_y  = (global_idx / p.out_w) % p.out_h;
+        c      = (global_idx / (p.out_w * p.out_h)) % p.channels;
+        n      = global_idx / (p.out_w * p.out_h * p.channels);
+    }
+};
+
+// cwhn layout: input/output stored as [N, H, W, C]
+struct dw_cwhn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + (y * p.in_w + x) * p.channels + c;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return (ky * p.kernel_w + kx) * p.channels + c;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + y * (p.out_w * p.channels) + x * p.channels + c;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        c      = global_idx % p.channels;
+        out_x  = (global_idx / p.channels) % p.out_w;
+        out_y  = (global_idx / (p.channels * p.out_w)) % p.out_h;
+        n      = global_idx / (p.channels * p.out_w * p.out_h);
+    }
+};
+
+template <typename Layout>
+static void conv2d_dw_kernel(const float * input, const float * kernel, float * output,
+                             const conv2d_dw_params p, const sycl::nd_item<3> & item_ct1) {
+    const int global_idx     = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int total_elements = p.batches * p.channels * p.out_h * p.out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    int n, c, out_y, out_x;
+    Layout::unpack_indices(global_idx, p, n, c, out_y, out_x);
+
+    float acc = 0.0f;
+    const conv2d_dw_kernel_bounds bounds = dw_calculate_kernel_bounds(out_x, out_y, p);
+
+    for (int ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+        const int in_y = dw_calculate_input_coord(out_y, ky, p.stride_y, p.dilation_y, p.padding_y);
+        for (int kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+            const int in_x = dw_calculate_input_coord(out_x, kx, p.stride_x, p.dilation_x, p.padding_x);
+            acc += input[Layout::input_index(n, c, in_y, in_x, p)] *
+                   kernel[Layout::kernel_index(c, ky, kx, p)];
+        }
+    }
+
+    output[Layout::output_index(n, c, out_y, out_x, p)] = acc;
+}
+
+template <typename Layout>
+static void conv2d_dw_sycl(const float * x_d, const float * w_d, float * y_d,
+                            const conv2d_dw_params p, const queue_ptr & stream) {
+    const int total      = p.batches * p.channels * p.out_h * p.out_w;
+    const int num_blocks = (total + SYCL_CONV2D_DW_BLOCK_SIZE - 1) / SYCL_CONV2D_DW_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_DW_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_dw_kernel<Layout>(x_d, w_d, y_d, p, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv2d_dw(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    const conv2d_dw_params params = { in_w, in_h, out_w, out_h, kernel_w, kernel_h,
+                                      stride_x, stride_y, padding_x, padding_y,
+                                      dilation_x, dilation_y, channels, batches };
+
+    const queue_ptr stream = ctx.stream();
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_sycl<dw_whcn_layout>(x_d, w_d, y_d, params, stream);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_sycl<dw_cwhn_layout>(x_d, w_d, y_d, params, stream);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv2d_dw");
+    }
+}

data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_DW_HPP
+#define GGML_SYCL_CONV2D_DW_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_DW_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d_dw(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_DW_HPP

data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.cpp

@@ -0,0 +1,125 @@
+#include "conv2d-transpose.hpp"
+#include "convert.hpp"
+
+template <typename kernel_t>
+static void conv2d_transpose_kernel(const float * input, const kernel_t * kernel, float * output,
+                                    const int in_w, const int in_h,
+                                    const int out_w, const int out_h,
+                                    const int kernel_w, const int kernel_h,
+                                    const int stride,
+                                    const int c_in, const int c_out, const int batches,
+                                    const sycl::nd_item<3> & item_ct1) {
+    const int global_idx     = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int total_elements = out_w * out_h * c_out * batches;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int out_x = global_idx % out_w;
+    const int out_y = (global_idx / out_w) % out_h;
+    const int c_idx = (global_idx / (out_w * out_h)) % c_out;
+    const int n_idx = global_idx / (out_w * out_h * c_out);
+
+    float acc = 0.0f;
+
+    for (int c_in_idx = 0; c_in_idx < c_in; ++c_in_idx) {
+        for (int kh = 0; kh < kernel_h; ++kh) {
+            int in_y = out_y - kh;
+            if (in_y < 0 || in_y % stride) {
+                continue;
+            }
+            in_y /= stride;
+            if (in_y >= in_h) {
+                continue;
+            }
+
+            for (int kw = 0; kw < kernel_w; ++kw) {
+                int in_x = out_x - kw;
+                if (in_x < 0 || in_x % stride) {
+                    continue;
+                }
+                in_x /= stride;
+                if (in_x >= in_w) {
+                    continue;
+                }
+
+                const int input_idx  = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + in_w * in_y + in_x;
+                const int kernel_idx = (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx +
+                                       kernel_w * kh + kw;
+
+                acc += input[input_idx] * ggml_sycl_cast<float>(kernel[kernel_idx]);
+            }
+        }
+    }
+
+    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + out_w * out_y + out_x] = acc;
+}
+
+template <typename kernel_t>
+static void conv2d_transpose_sycl(const float * input_d, const kernel_t * kernel_d, float * output_d,
+                                   const int in_w, const int in_h,
+                                   const int out_w, const int out_h,
+                                   const int kernel_w, const int kernel_h,
+                                   const int stride,
+                                   const int c_in, const int c_out, const int batches,
+                                   const queue_ptr & stream) {
+    const int total      = out_w * out_h * c_out * batches;
+    const int num_blocks = (total + SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_transpose_kernel<kernel_t>(input_d, kernel_d, output_d,
+                                             in_w, in_h, out_w, out_h, kernel_w, kernel_h,
+                                             stride, c_in, c_out, batches, item_ct1);
+        });
+}
+
+// input:  (W, H, C_in, N)
+// kernel: (W, H, C_out, C_in)
+// output: (W, H, C_out, N)
+void ggml_sycl_op_conv2d_transpose(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(input));
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const float * input_d  = (const float *) input->data;
+    float *       output_d = (float *) dst->data;
+    const void *  kernel_d = kernel->data;
+
+    const int input_w      = input->ne[0];
+    const int input_h      = input->ne[1];
+    const int channels_in  = input->ne[2];
+    const int batches      = input->ne[3];
+    const int output_w     = dst->ne[0];
+    const int output_h     = dst->ne[1];
+    const int channels_out = kernel->ne[2];
+    const int kernel_w     = kernel->ne[0];
+    const int kernel_h     = kernel->ne[1];
+    const int stride       = dst->op_params[0];
+
+    GGML_ASSERT(channels_in == kernel->ne[3]);
+    GGML_ASSERT(stride > 0);
+
+    const queue_ptr stream = ctx.stream();
+
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_transpose_sycl<sycl::half>(input_d, (const sycl::half *) kernel_d, output_d,
+                                          input_w, input_h, output_w, output_h, kernel_w, kernel_h,
+                                          stride, channels_in, channels_out, batches, stream);
+    } else {
+        conv2d_transpose_sycl<float>(input_d, (const float *) kernel_d, output_d,
+                                     input_w, input_h, output_w, output_h, kernel_w, kernel_h,
+                                     stride, channels_in, channels_out, batches, stream);
+    }
+}

data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_TRANSPOSE_HPP
+#define GGML_SYCL_CONV2D_TRANSPOSE_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d_transpose(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_TRANSPOSE_HPP

data/ext/sources/ggml/src/ggml-sycl/conv2d.cpp

@@ -0,0 +1,150 @@
+#include "conv2d.hpp"
+#include "convert.hpp"
+
+struct conv2d_params {
+    const int64_t IW, IH;
+    const int64_t OW, OH;
+    const int64_t KW, KH;
+    const int64_t ST_X, ST_Y;
+    const int64_t PD_X, PD_Y;
+    const int64_t DL_X, DL_Y;
+    const int64_t IC, OC;
+    const int64_t B;
+    const int64_t TOTAL;
+};
+
+struct conv2d_kernel_bounds {
+    int64_t y_min, y_max;
+    int64_t x_min, x_max;
+};
+
+static inline int64_t conv2d_max64(int64_t a, int64_t b) {
+    return (a > b) ? a : b;
+}
+
+static inline int64_t conv2d_min64(int64_t a, int64_t b) {
+    return (a < b) ? a : b;
+}
+
+static inline conv2d_kernel_bounds calculate_kernel_bounds(int64_t out_x, int64_t out_y, const conv2d_params & P) {
+    conv2d_kernel_bounds bounds;
+    bounds.y_min = conv2d_max64(0, (P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.y_max = conv2d_min64(P.KH, (P.IH + P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.x_min = conv2d_max64(0, (P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    bounds.x_max = conv2d_min64(P.KW, (P.IW + P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    return bounds;
+}
+
+static inline int calculate_input_coord(int64_t out_coord, int64_t kern_coord, int64_t stride,
+                                        int64_t dilation, int64_t padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+// whcn layout helpers (matching ggml tensor memory order)
+static inline int64_t whcn_input_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv2d_params & P) {
+    return n * (P.IC * P.IW * P.IH) + c * P.IW * P.IH + y * P.IW + x;
+}
+
+static inline int64_t whcn_kernel_index(int64_t c_out, int64_t c_in, int64_t ky, int64_t kx, const conv2d_params & P) {
+    return c_out * (P.IC * P.KH * P.KW) + c_in * (P.KH * P.KW) + ky * P.KW + kx;
+}
+
+static inline int64_t whcn_output_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv2d_params & P) {
+    return n * (P.OC * P.OW * P.OH) + c * P.OW * P.OH + y * P.OW + x;
+}
+
+template <typename T>
+static void conv2d_kernel(const float * input, const T * kernel, float * output,
+                          const conv2d_params P, const sycl::nd_item<3> & item_ct1) {
+    const int64_t global_idx = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+
+    if (global_idx >= P.TOTAL) {
+        return;
+    }
+
+    const int64_t out_x  = global_idx % P.OW;
+    const int64_t out_y  = (global_idx / P.OW) % P.OH;
+    const int64_t c_out  = (global_idx / (P.OW * P.OH)) % P.OC;
+    const int64_t n      = global_idx / (P.OW * P.OH * P.OC);
+
+    float acc = 0.0f;
+
+    const conv2d_kernel_bounds bounds = calculate_kernel_bounds(out_x, out_y, P);
+
+    for (int64_t c_in = 0; c_in < P.IC; ++c_in) {
+        for (int64_t ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+            const int64_t in_y = calculate_input_coord(out_y, ky, P.ST_Y, P.DL_Y, P.PD_Y);
+            for (int64_t kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+                const int64_t in_x = calculate_input_coord(out_x, kx, P.ST_X, P.DL_X, P.PD_X);
+                const float input_val  = input[whcn_input_index(n, c_in, in_y, in_x, P)];
+                const T     kernel_val = kernel[whcn_kernel_index(c_out, c_in, ky, kx, P)];
+                acc += input_val * ggml_sycl_cast<float>(kernel_val);
+            }
+        }
+    }
+
+    output[whcn_output_index(n, c_out, out_y, out_x, P)] = acc;
+}
+
+template <typename T>
+static void conv2d_sycl(const float * X_D, const T * K_D, float * Y_D,
+                        const conv2d_params P, const queue_ptr & stream) {
+    const int num_blocks = (P.TOTAL + SYCL_CONV2D_BLOCK_SIZE - 1) / SYCL_CONV2D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_kernel<T>(X_D, K_D, Y_D, P, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+    const float *       K_D    = (const float *) kernel->data;
+    const float *       X_D    = (const float *) input->data;
+    float *             Y_D    = (float *) dst->data;
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    // same number of input channels
+    GGML_ASSERT(input->ne[2] == kernel->ne[2]);
+
+    const queue_ptr stream = ctx.stream();
+
+    const int32_t * p    = (const int32_t *) dst->op_params;
+    const int       ST_X = p[0];
+    const int       ST_Y = p[1];
+    const int       PD_X = p[2];
+    const int       PD_Y = p[3];
+    const int       DL_X = p[4];
+    const int       DL_Y = p[5];
+
+    // no cwhn layout support
+    GGML_ASSERT(p[6] == 0);
+
+    const int IW = input->ne[0];
+    const int IH = input->ne[1];
+    const int OW = dst->ne[0];
+    const int OH = dst->ne[1];
+    const int KW = kernel->ne[0];
+    const int KH = kernel->ne[1];
+    const int IC = input->ne[2];
+    const int OC = kernel->ne[3];
+    const int B  = input->ne[3];
+
+    const int64_t     total  = (int64_t) B * OC * OH * OW;
+    const conv2d_params params = { IW, IH, OW, OH, KW, KH, ST_X, ST_Y, PD_X, PD_Y, DL_X, DL_Y, IC, OC, B, total };
+
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_sycl<sycl::half>(X_D, (const sycl::half *) K_D, Y_D, params, stream);
+    } else {
+        conv2d_sycl<float>(X_D, K_D, Y_D, params, stream);
+    }
+}

data/ext/sources/ggml/src/ggml-sycl/conv2d.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_HPP
+#define GGML_SYCL_CONV2D_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_HPP

data/ext/sources/ggml/src/ggml-sycl/conv3d.cpp

@@ -0,0 +1,224 @@
+#include "conv3d.hpp"
+
+static inline int64_t ggml_sycl_conv3d_calc_patch_total(const ggml_tensor * dst, int32_t n) {
+    return (int64_t) n * dst->ne[0] * dst->ne[1] * dst->ne[2];
+}
+
+static inline int64_t ggml_sycl_conv3d_calc_knl_n_total(const ggml_tensor * src0, int32_t c) {
+    return (int64_t) src0->ne[0] * src0->ne[1] * src0->ne[2] * c;
+}
+
+static inline void ggml_sycl_conv3d_write_output(
+        const ggml_tensor * dst,
+        const float * src, float * dst_data,
+        int64_t patch_total, int64_t oc,
+        int64_t dst_w, int64_t dst_h, int64_t dst_d,
+        dpct::queue_ptr stream) {
+    const int64_t dst_nb0 = dst->nb[0];
+    const int64_t dst_nb1 = dst->nb[1];
+    const int64_t dst_nb2 = dst->nb[2];
+    const int64_t dst_nb3 = dst->nb[3];
+    const int64_t total = patch_total * oc;
+    const int64_t block_size = 256;
+    const int64_t num_work_items = ((total + block_size - 1) / block_size) * block_size;
+
+    stream->parallel_for(sycl::range<1>(num_work_items), [=](sycl::id<1> id) {
+        const int64_t i = id[0];
+        if (i >= total) {
+            return;
+        }
+
+        const int64_t patch_idx = i / oc;
+        const int64_t out_ch = i % oc;
+        const int64_t p_in_batch = patch_idx % (dst_w * dst_h * dst_d);
+        const int64_t batch_idx = patch_idx / (dst_w * dst_h * dst_d);
+        const int64_t dst_z = p_in_batch / (dst_w * dst_h);
+        const int64_t dst_y = (p_in_batch % (dst_w * dst_h)) / dst_w;
+        const int64_t dst_x = p_in_batch % dst_w;
+        const int64_t ocn_idx = batch_idx * oc + out_ch;
+
+        const int64_t dst_offset = dst_x * dst_nb0 + dst_y * dst_nb1 + dst_z * dst_nb2 + ocn_idx * dst_nb3;
+        // `src` is a column-major (m x n) GEMM output where m == patch_total, n == oc.
+        // GEMM stores element (row, col) at index `row + col*m`, so compute index accordingly.
+        const int64_t src_index = patch_idx + out_ch * patch_total;
+        const float value = src[src_index];
+        *(float *)((char *)dst_data + dst_offset) = value;
+    });
+}
+
+void ggml_sycl_op_conv_3d(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_F16 || 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));
+
+    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int32_t s0 = opts[0];
+    const int32_t s1 = opts[1];
+    const int32_t s2 = opts[2];
+    const int32_t p0 = opts[3];
+    const int32_t p1 = opts[4];
+    const int32_t p2 = opts[5];
+    const int32_t d0 = opts[6];
+    const int32_t d1 = opts[7];
+    const int32_t d2 = opts[8];
+    const int32_t c  = opts[9];
+    const int32_t n  = opts[10];
+    const int32_t oc = opts[11];
+
+    const int64_t knl_w = src0->ne[0];
+    const int64_t knl_h = src0->ne[1];
+    const int64_t knl_d = src0->ne[2];
+
+    const int64_t patch_total = ggml_sycl_conv3d_calc_patch_total(dst, n);
+    const int64_t knl_n_total = ggml_sycl_conv3d_calc_knl_n_total(src0, c);
+
+    const size_t kernel_type_size = ggml_element_size(src0);
+
+    ggml_sycl_pool_alloc<float> gemm_output(ctx.pool());
+    gemm_output.alloc((size_t) patch_total * oc);
+
+    ggml_tensor dst_mat = {};
+    dst_mat.type = GGML_TYPE_F32;
+    dst_mat.ne[0] = patch_total;
+    dst_mat.ne[1] = oc;
+    dst_mat.ne[2] = 1;
+    dst_mat.ne[3] = 1;
+    dst_mat.nb[0] = sizeof(float);
+    dst_mat.nb[1] = dst_mat.nb[0] * dst_mat.ne[0];
+    dst_mat.nb[2] = dst_mat.nb[1];
+    dst_mat.nb[3] = dst_mat.nb[2];
+    dst_mat.data = gemm_output.get();
+    dst_mat.buffer = dst->buffer;
+    dst_mat.extra = dst->extra;
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    // allocate packed arrays: A_packed (k x m), B_packed (k x n)
+    ggml_sycl_pool_alloc<float> A_packed_alloc(ctx.pool());
+    ggml_sycl_pool_alloc<float> B_packed_alloc(ctx.pool());
+    A_packed_alloc.alloc((size_t) knl_n_total * patch_total);
+    B_packed_alloc.alloc((size_t) knl_n_total * oc);
+
+    float * A_packed = A_packed_alloc.get();
+    float * B_packed = B_packed_alloc.get();
+
+    const int m = (int) patch_total;
+    const int n_gemm = (int) oc;
+    const int k = (int) knl_n_total;
+
+    // Combined kernel: im2col -> pack A, and pack B simultaneously
+    const char * src1_base = (const char *) src1->data;
+    const char * src0_base = (const char *) src0->data;
+    const int64_t src1_nb0 = src1->nb[0];
+    const int64_t src1_nb1 = src1->nb[1];
+    const int64_t src1_nb2 = src1->nb[2];
+    const int64_t src1_nb3 = src1->nb[3];
+    const int64_t src1_w = src1->ne[0];
+    const int64_t src1_h = src1->ne[1];
+    const int64_t src1_d = src1->ne[2];
+
+    const bool src0_is_f32 = (src0->type == GGML_TYPE_F32);
+
+    // Compute correct strides for src0 as (knl_n_total, oc) matrix
+    const int64_t src0_packed_nb0 = kernel_type_size;
+    const int64_t src0_packed_nb1 = kernel_type_size * knl_n_total;
+
+    const int64_t KW = knl_w;
+    const int64_t KH = knl_h;
+    const int64_t KD = knl_d;
+    const int64_t PW = dst->ne[0];
+    const int64_t PH = dst->ne[1];
+    const int64_t PD = dst->ne[2];
+
+    // Pack A (with inline im2col): for each (row, col) in k x m matrix
+    const int64_t A_total = (int64_t)k * m;
+    const int64_t A_block_size = 256;
+    const int64_t A_num_work = ((A_total + A_block_size - 1) / A_block_size) * A_block_size;
+
+    stream->parallel_for(sycl::range<1>(A_num_work), [=](sycl::id<1> id) {
+        const int64_t t = id[0];
+        if (t >= A_total) return;
+
+        const int64_t row = t % k;
+        const int64_t col = t / k;
+
+        // Inline im2col for this element
+        const int64_t k_index = row;
+        const int64_t patch_idx = col;
+
+        const int64_t ic = k_index / (KD * KH * KW);
+        const int64_t rem = k_index - ic * (KD * KH * KW);
+        const int64_t kz = rem / (KH * KW);
+        const int64_t rem2 = rem - kz * (KH * KW);
+        const int64_t ky = rem2 / KW;
+        const int64_t kx = rem2 % KW;
+
+        const int64_t p_in_batch = patch_idx % (PW * PH * PD);
+        const int64_t batch_idx = patch_idx / (PW * PH * PD);
+        const int64_t dst_z = p_in_batch / (PW * PH);
+        const int64_t dst_y = (p_in_batch % (PW * PH)) / PW;
+        const int64_t dst_x = p_in_batch % PW;
+
+        const int64_t sx = dst_x * s0 + kx * d0 - p0;
+        const int64_t sy = dst_y * s1 + ky * d1 - p1;
+        const int64_t sz = dst_z * s2 + kz * d2 - p2;
+
+        float val = 0.0f;
+        if (sx >= 0 && sx < src1_w && sy >= 0 && sy < src1_h && sz >= 0 && sz < src1_d) {
+            const int64_t channel_idx = batch_idx * c + ic;
+            const char * ptr = src1_base + sx * src1_nb0 + sy * src1_nb1 + sz * src1_nb2 + channel_idx * src1_nb3;
+            val = *(const float *) ptr;
+        }
+        A_packed[row + col * (int64_t)k] = val;
+    });
+
+    // Pack B: for each (row, col) in k x n_gemm matrix
+    const int64_t B_total = (int64_t)k * n_gemm;
+    const int64_t B_block_size = 256;
+    const int64_t B_num_work = ((B_total + B_block_size - 1) / B_block_size) * B_block_size;
+
+    stream->parallel_for(sycl::range<1>(B_num_work), [=](sycl::id<1> id) {
+        const int64_t t = id[0];
+        if (t >= B_total) return;
+
+        const int64_t row = t % k;
+        const int64_t col = t / k;
+        const char * src_ptr = src0_base + row * src0_packed_nb0 + col * src0_packed_nb1;
+        float v;
+        if (src0_is_f32) {
+            v = *(const float *) src_ptr;
+        } else {
+            v = sycl::vec<sycl::half, 1>(*(const sycl::half *) src_ptr).convert<float, sycl::rounding_mode::automatic>()[0];
+        }
+        B_packed[row + col * (int64_t)k] = v;
+    });
+
+    // GEMM: C = A^T * B where A is (k x m), B is (k x n), C is (m x n)
+    const float alpha = 1.0f;
+    const float beta  = 0.0f;
+    const int lda = k;
+    const int ldb = k;
+    const int ldc = m;
+
+    SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+        *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans,
+        m, n_gemm, k,
+        dpct::get_value(&alpha, *stream),
+        (const float *) A_packed, lda,
+        (const float *) B_packed, ldb,
+        dpct::get_value(&beta, *stream),
+        (float *) dst_mat.data, ldc)));
+
+    const float * gemm_data = (const float *) dst_mat.data;
+    float * dst_data = (float *) dst->data;
+
+    ggml_sycl_conv3d_write_output(dst, gemm_data, dst_data, patch_total, oc,
+                                  dst->ne[0], dst->ne[1], dst->ne[2], stream);
+}