whispercpp 1.3.0 → 1.3.1

data/ext/ggml/src/ggml-sycl/common.cpp

@@ -0,0 +1,95 @@
+//
+// MIT license
+// Copyright (C) 2024 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 "common.hpp"
+#include "ggml-impl.h"
+
+int get_current_device_id() {
+  return dpct::dev_mgr::instance().current_device_id();
+}
+
+void* ggml_sycl_host_malloc(size_t size) try {
+  if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
+    return nullptr;
+  }
+
+  void* ptr = nullptr;
+  // allow to use dpct::get_in_order_queue() for host malloc
+  dpct::err0 err = CHECK_TRY_ERROR(
+      ptr = (void*)sycl::malloc_host(size, dpct::get_in_order_queue()));
+
+  if (err != 0) {
+    // clear the error
+    GGML_LOG_ERROR("WARNING: failed to allocate %.2f MB of pinned memory: %s\n", size / 1024.0 / 1024.0,    "syclGetErrorString is not supported");
+    return nullptr;
+  }
+
+  return ptr;
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+void ggml_sycl_host_free(void* ptr) try {
+  // allow to use dpct::get_in_order_queue() for host malloc
+  SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue())));
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size) {
+  const int64_t max_range = std::numeric_limits<int>::max();
+  int64_t sycl_down_blk_size = block_size;
+  int64_t global_range = accumulate_block_num * sycl_down_blk_size;
+  while(global_range > max_range) {
+      sycl_down_blk_size /= 2;
+      global_range = accumulate_block_num * sycl_down_blk_size;
+  }
+  return sycl_down_blk_size;
+}
+
+void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const ggml_sycl_op_flatten_t op) try {
+
+    const bool use_src1 = src1 != nullptr;
+
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+
+    // dd = data device
+    float * src0_ddf = (float *) src0->data;
+    float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
+    float *  dst_ddf = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> src0_f(ctx.pool());
+    ggml_sycl_pool_alloc<float> src1_f(ctx.pool());
+    ggml_sycl_pool_alloc<float>  dst_f(ctx.pool());
+
+    ggml_sycl_set_device(ctx.device);
+    queue_ptr main_stream = ctx.stream();
+    // GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n",
+        // ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device);
+
+    // do the computation
+    op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
+    // print_ggml_tensor("tensor", dst);
+}
+catch (sycl::exception const &exc) {
+
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}

data/ext/ggml/src/ggml-sycl/concat.cpp

@@ -0,0 +1,196 @@
+//
+// MIT license
+// Copyright (C) 2024 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 "concat.hpp"
+#include "common.hpp"
+
+static void concat_f32_dim0(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne00,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (nidx < ne00) { // src0
+    int offset_src = nidx + item_ct1.get_group(1) * ne00 +
+                     item_ct1.get_group(0) * ne00 * item_ct1.get_group_range(1);
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx - ne00 + item_ct1.get_group(1) * (ne0 - ne00) +
+        item_ct1.get_group(0) * (ne0 - ne00) * item_ct1.get_group_range(1);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_dim1(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne01,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (item_ct1.get_group(1) < (size_t) ne01) { // src0
+    int offset_src =
+        nidx + item_ct1.get_group(1) * ne0 + item_ct1.get_group(0) * ne0 * ne01;
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx + (item_ct1.get_group(1) - ne01) * ne0 +
+        item_ct1.get_group(0) * ne0 * (item_ct1.get_group_range(1) - ne01);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_dim2(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne02,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (item_ct1.get_group(0) < (size_t) ne02) { // src0
+    int offset_src = nidx + item_ct1.get_group(1) * ne0 +
+                     item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx + item_ct1.get_group(1) * ne0 +
+        (item_ct1.get_group(0) - ne02) * ne0 * item_ct1.get_group_range(1);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_sycl(const float *x, const float *y, float *dst,
+                            int ne00, int ne01, int ne02, int ne0, int ne1,
+                            int ne2, int dim, queue_ptr stream) {
+  int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE;
+  sycl::range<3> gridDim(ne2, ne1, num_blocks);
+  switch (dim) {
+  case 0:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
+        });
+    break;
+  case 1:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
+        });
+    break;
+  // dim >=2 will be dispatched to the default path
+  default:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
+        });
+    break;
+  }
+}
+
+// non-contiguous kernel (slow)
+static void concat_f32_sycl_non_cont(
+    queue_ptr stream, const char *src0, const char *src1, char *dst,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03, uint64_t nb00,
+    uint64_t nb01, uint64_t nb02, uint64_t nb03, int64_t /*ne10*/,
+    int64_t /*ne11*/, int64_t /*ne12*/, int64_t /*ne13*/, uint64_t nb10,
+    uint64_t nb11, uint64_t nb12, uint64_t nb13, int64_t ne0, int64_t ne1,
+    int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
+    uint64_t nb3, int32_t dim) {
+  sycl::range<3> gridDim(ne3, ne2, ne1);
+  stream->parallel_for(
+      sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)),
+      [=](sycl::nd_item<3> item_ct1) {
+        int64_t i3 = item_ct1.get_group(0);
+        int64_t i2 = item_ct1.get_group(1);
+        int64_t i1 = item_ct1.get_group(2);
+
+        int64_t o[4] = {0, 0, 0, 0};
+        o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+        const float *x;
+
+        for (int i0 = item_ct1.get_local_id(2); i0 < ne0;
+             i0 += item_ct1.get_local_range(2)) {
+          if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 +
+                                (i0)*nb00);
+          } else {
+            x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 +
+                                (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10);
+          }
+
+          float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
+
+          *y = *x;
+        }
+      });
+}
+
+void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst) {
+  queue_ptr stream = ctx.stream();
+
+  const int32_t dim = ((int32_t *)dst->op_params)[0];
+
+  if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+    const float *src0_d = (const float *)src0->data;
+    const float *src1_d = (const float *)src1->data;
+
+    float *dst_d = (float *)dst->data;
+
+    if (dim != 3) {
+      for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+        concat_f32_sycl(
+            src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4),
+            dst_d + i3 * (dst->nb[3] / 4), src0->ne[0], src0->ne[1],
+            src0->ne[2], dst->ne[0], dst->ne[1], dst->ne[2], dim, stream);
+      }
+    } else {
+      const size_t size0 = ggml_nbytes(src0);
+      const size_t size1 = ggml_nbytes(src1);
+
+      SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0).wait()));
+      SYCL_CHECK(CHECK_TRY_ERROR(
+          stream->memcpy(dst_d + size0 / 4, src1_d, size1).wait()));
+    }
+  } else
+    concat_f32_sycl_non_cont(
+        stream, (const char *)src0->data, (const char *)src1->data,
+        (char *)dst->data, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], src1->ne[0],
+        src1->ne[1], src1->ne[2], src1->ne[3], src1->nb[0], src1->nb[1],
+        src1->nb[2], src1->nb[3], dst->ne[0], dst->ne[1], dst->ne[2],
+        dst->ne[3], dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+}

data/ext/ggml/src/ggml-sycl/conv.cpp

@@ -0,0 +1,99 @@
+//
+// MIT license
+// Copyright (C) 2024 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 "conv.hpp"
+
+static  void conv_transpose_1d_kernel(
+        const int s0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2,
+        const int src1_ne0, const int dst_ne0,
+        const float * src0, const float * src1,  float * dst,
+        const sycl::nd_item<3> &item_ct1) {
+    int global_index = item_ct1.get_local_id(2) +
+                       item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (global_index >= output_size) {
+        return;
+    }
+
+    int out_index = global_index / dst_ne0;
+
+    float accumulator = 0;
+
+    for (int c = 0; c < src0_ne2; c++) {
+        int idx = global_index % dst_ne0;
+
+        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
+        int input_offset = src1_ne0 * c;
+
+        for (int i = 0; i < src1_ne0; i++) {
+            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+                continue;
+            }
+            int weight_idx = idx - i*s0;
+
+            float kernel_weight = src0[kernel_offset + weight_idx];
+            float input_value =  src1[input_offset+i];
+
+            accumulator += kernel_weight * input_value;
+        }
+    }
+    dst[global_index] = accumulator;
+}
+
+static void conv_transpose_1d_f32_f32_sycl(
+    const int s0, const int output_size,
+    const int src0_ne0, const int src0_ne1, const int src0_ne2,
+    const int src1_ne0, const int dst_ne0,
+    const float *src0, const float *src1, float *dst,
+    const queue_ptr& stream) {
+
+    const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_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) {
+            conv_transpose_1d_kernel(
+                s0, output_size,
+                src0_ne0, src0_ne1, src0_ne2,
+                src1_ne0, dst_ne0,
+                src0, src1, dst, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor *dst) {
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    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));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+
+    const int s0 = opts[0];
+
+    const int64_t output_size = ggml_nelements(dst);
+
+    conv_transpose_1d_f32_f32_sycl(s0, output_size,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        src1->ne[0], dst->ne[0],
+        src0_d, src1_d, dst_d, stream);
+}
+