vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

eltwise: add back split buffer type checks

Browse source
This commit is contained in:
Akarshan Biswas 2025-02-01 11:45:18 +05:30
parent ecacff3f6e
commit 7d8d689d39
No known key found for this signature in database
GPG key ID: 52A578A14B32134D
1 changed files with 21 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

22
ggml/src/ggml-sycl/element_wise.cpp
View file

@ -512,6 +512,7 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -523,6 +524,7 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -534,6 +536,7 @@ inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -545,6 +548,7 @@ inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
@ -556,6 +560,7 @@ inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -566,6 +571,7 @@ inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tenso
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -576,6 +582,7 @@ inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
@ -588,6 +595,7 @@ inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -598,6 +606,7 @@ inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -609,6 +618,7 @@ inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor *d
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -620,6 +630,7 @@ inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -630,6 +641,7 @@ inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -641,7 +653,7 @@ inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -653,6 +665,7 @@ inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
dpct::queue_ptr main_stream = ctx.stream();
@ -664,6 +677,7 @@ inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
dpct::queue_ptr main_stream = ctx.stream();
@ -675,6 +689,7 @@ inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
float negative_slope;
memcpy(&negative_slope, dst->op_params, sizeof(float));
@ -690,6 +705,7 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
dpct::queue_ptr main_stream = ctx.stream();
@ -701,6 +717,7 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor *
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const float sf0 = (float)dst->ne[0]/dst->src[0]->ne[0];
const float sf1 = (float)dst->ne[1]/dst->src[0]->ne[1];
@ -721,6 +738,7 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
@ -738,6 +756,8 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx,
GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
GGML_ASSERT(strcmp(dst->src[1]->buffer->buft->iface.get_name(dst->src[1]->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const dpct::queue_ptr main_stream = ctx.stream();
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);