Remove deprecated tensor->backend uses
This commit is contained in:
0cc4m
parent
179103c23d
commit
8dbde1f033
1 changed files with 21 additions and 107 deletions
128
ggml-vulkan.cpp
128
ggml-vulkan.cpp
|
|
@ -376,7 +376,6 @@ struct ggml_backend_vk_context {
|
|||
vk_context * compute_ctx;
|
||||
vk_context * transfer_ctx;
|
||||
|
||||
bool disable;
|
||||
bool initialized;
|
||||
|
||||
size_t idx;
|
||||
|
|
@ -1875,7 +1874,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
|
|||
ctx->compute_ctx = nullptr;
|
||||
ctx->transfer_ctx = nullptr;
|
||||
|
||||
ctx->disable = false;
|
||||
ctx->initialized = true;
|
||||
|
||||
ctx->idx = idx;
|
||||
|
|
@ -2811,7 +2809,7 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_
|
|||
|
||||
static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
|
||||
std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
|
||||
std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")" << std::endl;
|
||||
#endif
|
||||
const int tensor_type_size = ggml_type_size(tensor->type);
|
||||
|
|
@ -2831,9 +2829,9 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context
|
|||
|
||||
static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
std::cerr << "ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT
|
||||
GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT
|
||||
|
|
@ -3001,19 +2999,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
|
|||
ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21,
|
||||
0, 0, 0, 0, 1
|
||||
); // NOLINT
|
||||
|
||||
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
||||
// copy dst to host
|
||||
float * d = (float *) ((char *) dst->data);
|
||||
ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, sizeof(float) * d_ne * ne12 * ne13);
|
||||
}
|
||||
}
|
||||
|
||||
static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
std::cerr << "ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT
|
||||
GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT
|
||||
|
|
@ -3166,12 +3158,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
|
|||
|
||||
static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_mul_mat_p021_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
std::cerr << "ggml_vk_mul_mat_p021_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
|
||||
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
|
||||
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // NOLINT
|
||||
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // NOLINT
|
||||
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
||||
|
|
@ -3236,25 +3227,17 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
|
|||
const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
|
||||
|
||||
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
||||
// copy dst to host
|
||||
float * d = (float *) dst->data;
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
|
||||
}
|
||||
}
|
||||
|
||||
static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
std::cerr << "ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(!ggml_is_transposed(src0));
|
||||
GGML_ASSERT(!ggml_is_transposed(src1));
|
||||
GGML_ASSERT(!ggml_is_permuted(src0));
|
||||
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
|
||||
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
||||
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
||||
|
||||
|
|
@ -3321,26 +3304,6 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
|
|||
const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
|
||||
|
||||
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
||||
// copy dst to host
|
||||
float * d = (float *) dst->data;
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
|
||||
}
|
||||
}
|
||||
|
||||
static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * dst) {
|
||||
const uint64_t ne10 = src1->ne[0];
|
||||
|
||||
const uint64_t ne0 = dst->ne[0];
|
||||
const uint64_t ne1 = dst->ne[1];
|
||||
|
||||
// TODO: find the optimal values for these
|
||||
return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
|
||||
(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || ggml_is_quantized(src1->type)) &&
|
||||
dst->type == GGML_TYPE_F32 &&
|
||||
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU);
|
||||
}
|
||||
|
||||
static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
|
|
@ -3730,8 +3693,6 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
|
|||
// TODO: support for transposed / permuted tensors
|
||||
GGML_ASSERT(nb0 == sizeof(float));
|
||||
GGML_ASSERT(nb00 == sizeof(float));
|
||||
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
|
||||
GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
||||
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
||||
|
|
@ -3921,11 +3882,11 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
|
|||
template<typename PC>
|
||||
static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op, const PC&& pc) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
||||
if (src1 != nullptr) {
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
||||
}
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
|
||||
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type)))); // NOLINT
|
||||
GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0)); // NOLINT
|
||||
|
|
@ -3942,8 +3903,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
const uint64_t ne13 = use_src1 ? src1->ne[3] : 0;
|
||||
const uint64_t ne1 = ne10 * ne11;
|
||||
// const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
|
||||
const uint64_t nb2 = dst->nb[2];
|
||||
const uint64_t nb3 = dst->nb[3];
|
||||
|
||||
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, dst, op);
|
||||
ggml_vk_func_t op_func;
|
||||
|
|
@ -3994,7 +3953,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
|
||||
// Workaround for tiny tensor inputs on ROPE
|
||||
if (use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU && y_sz > d_D->size) {
|
||||
if (use_src1 && y_sz > d_D->size) {
|
||||
y_sz = VK_WHOLE_SIZE;
|
||||
}
|
||||
|
||||
|
|
@ -4085,13 +4044,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
|
||||
}
|
||||
if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) {
|
||||
ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst);
|
||||
} else if(dst->backend == GGML_BACKEND_TYPE_CPU) {
|
||||
// copy dst to host
|
||||
float * d = (float *) dst->data;
|
||||
ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
|
||||
}
|
||||
} else {
|
||||
GGML_ASSERT(op != GGML_OP_SOFT_MAX);
|
||||
GGML_ASSERT(op != GGML_OP_ARGSORT);
|
||||
|
|
@ -4130,10 +4082,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
|
||||
}
|
||||
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
||||
// copy dst to host
|
||||
ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset + d_offset, (char *) dst->data + i02*nb2 + i03*nb3, d_sz);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -4594,15 +4542,11 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
|
|||
src1_ggml->data = y;
|
||||
tensor_ggml->data = d_chk;
|
||||
|
||||
ctx->disable = true;
|
||||
|
||||
ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
|
||||
ggml_build_forward_expand(cgraph, tensor_ggml);
|
||||
|
||||
ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
|
||||
|
||||
ctx->disable = false;
|
||||
|
||||
ggml_free(ggml_ctx);
|
||||
|
||||
double avg_err = 0.0;
|
||||
|
|
@ -5082,15 +5026,11 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
|
|||
src1_ggml->data = y;
|
||||
tensor_ggml->data = d_chk;
|
||||
|
||||
ctx->disable = true;
|
||||
|
||||
ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
|
||||
ggml_build_forward_expand(cgraph, tensor_ggml);
|
||||
|
||||
ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
|
||||
|
||||
ctx->disable = false;
|
||||
|
||||
ggml_free(ggml_ctx);
|
||||
|
||||
double avg_err = 0.0;
|
||||
|
|
@ -5167,10 +5107,6 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
|
|||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
|
||||
#endif
|
||||
if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU) {
|
||||
return;
|
||||
}
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
|
||||
|
||||
ggml_tensor * src0 = node->src[0];
|
||||
|
|
@ -5277,9 +5213,6 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
|
|||
}
|
||||
|
||||
static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
|
||||
if (ctx->disable) {
|
||||
return;
|
||||
}
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_preallocate_buffers(x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
|
||||
#endif
|
||||
|
|
@ -5453,7 +5386,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
|
|||
}
|
||||
|
||||
static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
|
||||
if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(node)) {
|
||||
if (ggml_is_empty(node)) {
|
||||
return;
|
||||
}
|
||||
|
||||
|
|
@ -5612,7 +5545,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
last_node = true;
|
||||
#endif
|
||||
|
||||
if (node->backend == GGML_BACKEND_TYPE_CPU || last_node) {
|
||||
if (last_node) {
|
||||
ggml_vk_ctx_end(ctx->compute_ctx);
|
||||
ctx->compute_ctx->exit_tensor = node;
|
||||
ctx->compute_ctx = nullptr;
|
||||
|
|
@ -5620,10 +5553,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
}
|
||||
|
||||
static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor){
|
||||
if (ctx->disable) {
|
||||
return false;
|
||||
}
|
||||
|
||||
ggml_tensor_extra_gpu * extra = nullptr;
|
||||
|
||||
switch (tensor->op) {
|
||||
|
|
@ -5682,7 +5611,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
}
|
||||
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")" << std::endl;
|
||||
std::cerr << "ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")" << std::endl;
|
||||
#endif
|
||||
|
||||
#ifdef GGML_VULKAN_CHECK_RESULTS
|
||||
|
|
@ -5722,9 +5651,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
|
||||
// Clean up after graph processing is done
|
||||
static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
|
||||
if (ctx->disable) {
|
||||
return;
|
||||
}
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_vk_graph_cleanup()" << std::endl;
|
||||
#endif
|
||||
|
|
@ -5897,7 +5823,6 @@ GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t b
|
|||
extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
|
||||
}
|
||||
|
||||
tensor->backend = GGML_BACKEND_TYPE_GPU;
|
||||
tensor->extra = extra;
|
||||
}
|
||||
|
||||
|
|
@ -5905,8 +5830,6 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
|
|||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
||||
|
||||
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
|
@ -5920,8 +5843,6 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
|
|||
#ifdef GGML_VULKAN_DEBUG
|
||||
std::cerr << "ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
|
||||
#endif
|
||||
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
||||
|
||||
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
|
@ -6152,7 +6073,6 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
|
|||
#endif
|
||||
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
|
||||
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
|
||||
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
||||
|
|
@ -6173,7 +6093,6 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
|
|||
#endif
|
||||
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
|
||||
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
|
||||
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
||||
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
||||
|
|
@ -6253,7 +6172,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
|
|||
int last_node = cgraph->n_nodes - 1;
|
||||
|
||||
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
|
||||
while (last_node > 0 && (cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(cgraph->nodes[last_node]))) {
|
||||
while (last_node > 0 && ggml_is_empty(cgraph->nodes[last_node])) {
|
||||
last_node -= 1;
|
||||
}
|
||||
|
||||
|
|
@ -6911,16 +6830,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
GGML_ASSERT(false);
|
||||
}
|
||||
|
||||
// Disable vulkan here to avoid the hooks in ggml.c
|
||||
ctx->disable = true;
|
||||
|
||||
ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
|
||||
ggml_build_forward_expand(cgraph, tensor_clone);
|
||||
|
||||
ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
|
||||
|
||||
ctx->disable = false;
|
||||
|
||||
ggml_vk_check_tensor(ggml_op_name(tensor->op), tensor_clone);
|
||||
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
|
||||
ggml_vk_print_tensor(ctx, tensor_clone, "tensor_clone");
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue