diff options
| author | Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> | 2024-02-08 22:39:10 +0530 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-02-08 22:39:10 +0530 |
| commit | 6e99f2a04f1871d637dd77eb4d81de31a5510253 (patch) | |
| tree | ef11166429d2b6bb6163ec07bdd58047db78b40b /ggml-sycl.cpp | |
| parent | ff4ff05c5ff4311c05a8ce1f984c7d8def4f07a5 (diff) | |
Fix f16_sycl cpy call from Arc (#5411)
* fix f16_sycl cpy call
* rm old logic
* add fp16 build CI
* use macro
* format fix
Diffstat (limited to 'ggml-sycl.cpp')
| -rw-r--r-- | ggml-sycl.cpp | 8 |
1 files changed, 5 insertions, 3 deletions
diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index a03df4c6..dd562a89 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -12148,7 +12148,8 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec( const int64_t src1_ncols, const int64_t src1_padded_row_size, const dpct::queue_ptr &stream) { - const int64_t ne00 = src0->ne[0]; + GGML_TENSOR_BINARY_OP_LOCALS + const int64_t row_diff = row_high - row_low; // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics @@ -12167,8 +12168,9 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec( } else { src1_dfloat = src1_dfloat_a.alloc(ne00); ggml_cpy_f32_f16_sycl((const char *)src1_ddf_i, (char *)src1_dfloat, - ne00, ne00, 1, sizeof(float), 0, 0, ne00, 1, - sizeof(sycl::half), 0, 0, stream); + ne00, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, + nb13, stream); } } #else |