whispercpp 1.3.6 → 1.3.7

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

@@ -24,4 +24,20 @@ 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);
+
 #endif // GGML_SYCL_MMVQ_HPP

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);
 }

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

@@ -58,6 +58,31 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q3_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI3_K;
+        static constexpr uint32_t qr       = QR3_K;
+        static constexpr uint32_t vdr_mmvq = 1;
+    };
+
+    // Reordered layout: [qs (QK_K/4 per block)] [hmask (QK_K/8 per block)] [scales] [d]
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
+        auto qs_offset    = block_index * (QK_K / 4);
+        auto hmask_offset = n_blocks * (QK_K / 4) + block_index * (QK_K / 8);
+        return { qs_offset, hmask_offset };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / QK_K));
+        auto total_qs_bytes = nblocks * (QK_K / 4) + nblocks * (QK_K / 8);
+        return { total_qs_bytes + block_index * 12,
+                 total_qs_bytes + nblocks * 12 + block_index * sizeof(ggml_half) };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
+
 template <> struct block_q_t<GGML_TYPE_Q4_K> {
     struct traits {
         static constexpr uint32_t qk       = QK_K;
@@ -79,6 +104,31 @@ template <> struct block_q_t<GGML_TYPE_Q4_K> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q5_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI5_K;
+        static constexpr uint32_t qr       = QR5_K;
+        static constexpr uint32_t vdr_mmvq = 2;
+    };
+
+    // Reordered layout: [qs (QK_K/2 per block)] [qh (QK_K/8 per block)] [scales] [dm]
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
+        auto qs_offset = block_index * (QK_K / 2);
+        auto qh_offset = n_blocks * (QK_K / 2) + block_index * (QK_K / 8);
+        return { qs_offset, qh_offset };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / QK_K));
+        auto total_qs_bytes = nblocks * (QK_K / 2) + nblocks * (QK_K / 8);
+        return { total_qs_bytes + block_index * K_SCALE_SIZE,
+                 total_qs_bytes + nblocks * K_SCALE_SIZE + block_index * sizeof(ggml_half2) };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
+
 template <> struct block_q_t<GGML_TYPE_Q6_K> {
     struct traits {
         static constexpr uint32_t qk       = QK_K;
@@ -105,6 +155,27 @@ template <> struct block_q_t<GGML_TYPE_Q6_K> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q8_0> {
+    struct traits {
+        static constexpr uint32_t qk       = QK8_0;      // 32
+        static constexpr uint32_t qi       = QI8_0;      // 8
+        static constexpr uint32_t qr       = QR8_0;      // 1
+        static constexpr uint32_t vdr_mmvq = 4;
+    };
+
+    // Q8_0 reorder layout: [qs0|qs1|...|qsN][d0|d1|...|dN]
+    // Each block has 32 int8 weights (32 bytes) followed by all scales
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * QK8_0, 0 };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        return { (ncols * nrows) + block_index * sizeof(ggml_half), 0 };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }  // 1
+};
+
 }  // namespace ggml_sycl_reordered
 
 #endif  // GGML_SYCL_QUANTS_HPP

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

@@ -4,7 +4,11 @@
 namespace utils {
 template<typename T>
 static constexpr bool is_arithmetic_v() {
-    return std::is_arithmetic_v<T> || std::is_same_v<T, sycl::half> || std::is_same_v<T, sycl::ext::oneapi::bfloat16>;
+    return std::is_arithmetic_v<T> || std::is_same_v<T, sycl::half>
+#ifdef GGML_SYCL_HAS_BF16
+        || std::is_same_v<T, sycl::ext::oneapi::bfloat16>
+#endif
+        ;
 }
 }
 
@@ -181,6 +185,7 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
                 stream
             );
             break;
+#ifdef GGML_SYCL_HAS_BF16
         case GGML_TYPE_BF16:
             set_rows_sycl<TIn, TIdx, sycl::ext::oneapi::bfloat16>(
                 src0_d, src1_d, (char *)dst->data,
@@ -193,6 +198,7 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
                 stream
             );
             break;
+#endif
         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;

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

@@ -0,0 +1,172 @@
+#include "solve_tri.hpp"
+#include "common.hpp"
+#include <oneapi/mkl/blas.hpp>
+
+template <int n_template, int k_template>
+static void solve_tri_f32_fast(const float * __restrict__ A,
+                               const float * __restrict__ B,
+                               float * __restrict__ X,
+                               const int64_t ne02, [[maybe_unused]] const int64_t ne03,
+                               const int64_t nb02, const int64_t nb03,
+                               const int64_t nb12, const int64_t nb13,
+                               const int64_t nb2,  const int64_t nb3,
+                               const int n_arg, const int k_arg,
+                               const sycl::nd_item<2> & item, float * sA) {
+
+    const int n = n_template == 0 ? n_arg : n_template;
+    const int k = k_template == 0 ? k_arg : k_template;
+
+    const int batch_idx = item.get_group(1);
+    const int lane      = item.get_local_id(1) % WARP_SIZE;
+    const int col_idx   = item.get_local_id(0);
+
+    if (col_idx >= k) {
+        return;
+    }
+
+    const int64_t i03 = batch_idx / ne02;
+    const int64_t i02 = batch_idx % ne02;
+
+    const float * A_batch = (const float *) ((const char *) A + i02 * nb02 + i03 * nb03);
+    const float * B_batch = (const float *) ((const char *) B + i02 * nb12 + i03 * nb13);
+    float *       X_batch = (float *)       ((char *)       X + i02 * nb2  + i03 * nb3);
+
+    const int offset = item.get_local_id(1) + item.get_local_id(0) * item.get_local_range(1);
+
+#pragma unroll
+    for (int i = 0; i < n * n; i += k * WARP_SIZE) {
+        const int i0 = i + offset;
+        if (i0 < n * n) {
+            sA[i0] = A_batch[i0];
+        }
+    }
+
+    item.barrier(sycl::access::fence_space::local_space);
+
+    float x_low  = (lane < n) ? B_batch[lane * k + col_idx] : 0.0f;
+    float x_high = (WARP_SIZE + lane < n) ? B_batch[(WARP_SIZE + lane) * k + col_idx] : 0.0f;
+
+    const int half      = WARP_SIZE;
+    const int nrows_low = (n < half) ? n : half;
+
+#pragma unroll
+    for (int row = 0; row < nrows_low; ++row) {
+        float sum = 0.0f;
+        if (lane < row) {
+            sum += sA[row * n + lane] * x_low;
+        }
+        sum = warp_reduce_sum<WARP_SIZE>(sum);
+        if (lane == row) {
+            x_low = (x_low - sum) / sA[row * n + row];
+        }
+    }
+
+#pragma unroll
+    for (int row = half; row < n; ++row) {
+        float     sum = sA[row * n + lane] * x_low;
+        const int j   = half + lane;
+        if (j < row) {
+            sum += sA[row * n + j] * x_high;
+        }
+        sum = warp_reduce_sum<WARP_SIZE>(sum);
+        if (lane == row - half) {
+            x_high = (x_high - sum) / sA[row * n + row];
+        }
+    }
+
+#pragma unroll
+    for (int rr = 0; rr < 2; ++rr) {
+        const int row = rr * WARP_SIZE + lane;
+        if (row < n) {
+            const float val            = (row < half) ? x_low : x_high;
+            X_batch[row * k + col_idx] = val;
+        }
+    }
+}
+
+static void solve_tri_f32_mkl(dpct::queue_ptr stream,
+                               const float * A, float * X,
+                               int n, int k,
+                               int64_t ne02, [[maybe_unused]] int64_t ne03,
+                               int64_t nb02, [[maybe_unused]] int64_t nb03,
+                               int64_t nb2,  [[maybe_unused]] int64_t nb3) {
+    const float alpha = 1.0f;
+    const int64_t total_batches = ne02 * ne03;
+    if (total_batches == 0) {
+        return;
+    }
+
+    const int64_t stride_a = nb02 / sizeof(float);
+    const int64_t stride_x = nb2 / sizeof(float);
+
+    oneapi::mkl::blas::trsm_batch(
+        *stream,
+        oneapi::mkl::side::right,
+        oneapi::mkl::uplo::upper,
+        oneapi::mkl::transpose::nontrans,
+        oneapi::mkl::diag::nonunit,
+        k, n, alpha,
+        A, n, stride_a,
+        X, k, stride_x,
+        total_batches);
+}
+
+inline void ggml_sycl_op_solve_tri(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+    dpct::queue_ptr stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int n    = src0->ne[0];
+    const int k    = src1->ne[0];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    GGML_ASSERT(n <= SYCL_SOLVE_TRI_MAX_N && k <= SYCL_SOLVE_TRI_MAX_K);
+
+    const float * A_d = static_cast<const float *>(src0->data);
+    const float * B_d = static_cast<const float *>(src1->data);
+    float * X_d       = static_cast<float *>(dst->data);
+
+    if (X_d != B_d) {
+        const int64_t total_elements = (int64_t)n * k * ne02 * ne03;
+        stream->memcpy(X_d, B_d, total_elements * sizeof(float));
+    }
+
+    const int64_t nb02 = src0->nb[2];
+    const int64_t nb03 = src0->nb[3];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+    const int64_t nb2  = dst->nb[2];
+    const int64_t nb3  = dst->nb[3];
+
+    const int64_t total_batches = ne02 * ne03;
+
+    if (n <= 2 * WARP_SIZE && k <= 32) {
+        const int smem_size = 2 * WARP_SIZE * 2 * WARP_SIZE;
+        const sycl::range<2> grid(1, total_batches);
+        const sycl::range<2> block(k, WARP_SIZE);
+        stream->submit([&](sycl::handler & cgh) {
+            sycl::local_accessor<float, 1> smem_acc(sycl::range<1>(smem_size), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<2>(grid * block, block),
+                [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    solve_tri_f32_fast<0, 0>(A_d, B_d, X_d, ne02, ne03,
+                                              nb02, nb03, nb12, nb13, nb2, nb3,
+                                              n, k, item, get_pointer(smem_acc));
+                });
+        });
+    } else {
+        solve_tri_f32_mkl(stream, A_d, X_d, n, k, ne02, ne03, nb02, nb03, nb2, nb3);
+    }
+}
+
+void ggml_sycl_solve_tri(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_solve_tri(ctx, dst);
+}

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

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "common.hpp"
+
+#define SYCL_SOLVE_TRI_MAX_N 64
+#define SYCL_SOLVE_TRI_MAX_K 64
+
+void ggml_sycl_solve_tri(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

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

@@ -63,7 +63,7 @@ static void kernel_ssm_conv(
     });
 }
 
-void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+inline void ggml_sycl_op_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
 
@@ -125,3 +125,8 @@ void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
         throw;
     }
 }
+
+void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_ssm_conv(ctx, dst);
+}

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

@@ -0,0 +1,156 @@
+#include "ssm_scan.hpp"
+#include "common.hpp"
+
+template <int c_factor, int d_state>
+static void ssm_scan_f32_group(
+        const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
+        const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+        const int32_t * __restrict__ src6, float * __restrict__ dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+        const int src2_nb1, const int src2_nb2, const int src3_nb1,
+        const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+        const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok,
+        const sycl::nd_item<2> & item) {
+
+    const int lane     = item.get_local_id(1) % WARP_SIZE;
+    const int warp     = item.get_local_id(1) / WARP_SIZE;
+    const int warp_idx = item.get_group(1) * c_factor + warp;
+    const int seq_idx  = item.get_group(0);
+
+    const int head_idx = warp_idx / d_head;
+    const int head_off = (warp_idx % d_head) * sizeof(float);
+    const int group_off = (head_idx / (n_head / n_group)) * d_state * sizeof(float);
+
+    const float * s0_warp = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+    const float * x_warp  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + (warp_idx * sizeof(float)));
+    const float * dt_warp = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
+    const float * A_warp  = (const float *) ((const char *) src3 + head_idx * src3_nb1);
+    const float * B_warp  = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
+    const float * C_warp  = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
+    float *       y_warp  = dst + (seq_idx * n_tok * n_head * d_head) + warp_idx;
+    float *       s_warp  = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+
+    const int stride_x  = src1_nb2 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
+    const int stride_y  = n_head * d_head;
+
+    float state[c_factor];
+    float state_sum = 0.0f;
+
+#pragma unroll
+    for (int j = 0; j < c_factor; j++) {
+        state[j] = s0_warp[WARP_SIZE * j + lane];
+    }
+
+    for (int64_t i = 0; i < n_tok; i++) {
+        const float dt_val = dt_warp[i * stride_dt];
+        const float dt_soft_plus = (dt_val <= 20.0f ? sycl::log1p(sycl::exp(dt_val)) : dt_val);
+
+        state_sum = 0.0f;
+        const float dA   = sycl::exp(dt_soft_plus * A_warp[0]);
+        const float x_dt = x_warp[i * stride_x] * dt_soft_plus;
+#pragma unroll
+        for (int j = 0; j < c_factor; j++) {
+            const float B_val = B_warp[i * stride_B + WARP_SIZE * j + lane];
+            const float C_val = C_warp[i * stride_C + WARP_SIZE * j + lane];
+            state[j] = (state[j] * dA) + (B_val * x_dt);
+            state_sum += state[j] * C_val;
+        }
+
+        state_sum = warp_reduce_sum<WARP_SIZE>(state_sum);
+
+        if (lane == 0) {
+            y_warp[i * stride_y] = state_sum;
+        }
+    }
+
+#pragma unroll
+    for (int j = 0; j < c_factor; j++) {
+        s_warp[WARP_SIZE * j + lane] = state[j];
+    }
+}
+
+static void ssm_scan_f32_sycl(
+        const float * src0, const float * src1, const float * src2, const float * src3,
+        const float * src4, const float * src5, const int32_t * src6, float * dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
+        const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
+        const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+        const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
+        dpct::queue_ptr stream) {
+
+    // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
+    GGML_ASSERT(src3_nb1 == sizeof(float));
+    if (d_state == 128) {
+        constexpr int threads   = 128;
+        constexpr int num_warps = threads / WARP_SIZE;
+        const sycl::range<2> grid(n_seq, (n_head * head_dim + num_warps - 1) / num_warps);
+        const sycl::range<2> block(1, threads);
+        stream->parallel_for(
+            sycl::nd_range<2>(grid * block, block),
+            [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                ssm_scan_f32_group<128 / WARP_SIZE, 128>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok, item);
+            });
+    } else if (d_state == 256) {
+        constexpr int threads   = 256;
+        constexpr int num_warps = threads / WARP_SIZE;
+        const sycl::range<2> grid(n_seq, (n_head * head_dim + num_warps - 1) / num_warps);
+        const sycl::range<2> block(1, threads);
+        stream->parallel_for(
+            sycl::nd_range<2>(grid * block, block),
+            [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                ssm_scan_f32_group<256 / WARP_SIZE, 256>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok, item);
+            });
+    } else {
+        GGML_ABORT("ssm_scan: unsupported d_state (must be 128 or 256)");
+    }
+}
+
+inline void ggml_sycl_op_ssm_scan(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+    const ggml_tensor * src3 = dst->src[3];
+    const ggml_tensor * src4 = dst->src[4];
+    const ggml_tensor * src5 = dst->src[5];
+    const ggml_tensor * src6 = dst->src[6];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src6->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t nc  = src0->ne[0];
+    const int64_t nr  = src0->ne[1];
+    const int64_t nh  = src1->ne[1];
+    const int64_t ng  = src4->ne[1];
+    const int64_t n_t = src1->ne[2];
+    const int64_t n_s = src1->ne[3];
+    const int64_t s_off = ggml_nelements(src1) * sizeof(float);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc * nr * nh * n_s == ggml_nelements(dst));
+
+    dpct::queue_ptr stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    ssm_scan_f32_sycl(
+        static_cast<const float *>(src0->data), static_cast<const float *>(src1->data),
+        static_cast<const float *>(src2->data), static_cast<const float *>(src3->data),
+        static_cast<const float *>(src4->data), static_cast<const float *>(src5->data),
+        static_cast<const int32_t *>(src6->data), static_cast<float *>(dst->data),
+        src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
+        src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
+        s_off, nc, nr, nh, ng, n_t, n_s, stream);
+}
+
+void ggml_sycl_ssm_scan(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/7);
+    ggml_sycl_op_ssm_scan(ctx, dst);
+}

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

@@ -0,0 +1,5 @@
+#pragma once
+
+#include "common.hpp"
+
+void ggml_sycl_ssm_scan(ggml_backend_sycl_context & ctx, ggml_tensor * dst);