]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
sycl: remove separate fp32 type promotion in gemm non-oneDNN path (#26372)
authorIan Faust <redacted>
Thu, 13 Aug 2026 09:02:18 +0000 (11:02 +0200)
committerGitHub <redacted>
Thu, 13 Aug 2026 09:02:18 +0000 (12:02 +0300)
* sycl: use automatic fp16 promotion in gemm

* sycl: remove redundant comment

ggml/src/ggml-sycl/ggml-sycl.cpp

index e712c18bada1b76cfe829d6b4c9fbbbaa9508c68..a239d2f8d769c0a08133dea861e2fb8501e15313 100644 (file)
@@ -2687,21 +2687,15 @@ inline void ggml_sycl_op_mul_mat_sycl(
         else
 #endif
         {
-            ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
-
-            const sycl::half alpha_f16 = 1.0f;
-            const sycl::half beta_f16  = 0.0f;
+            const float alpha = 1.0f;
+            const float beta  = 0.0f;
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
                 *stream, oneapi::mkl::transpose::trans,
                 oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
-                &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
-                src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
-                dst_f16.get(), dpct::library_data_t::real_half, ldc,
-                dpct::library_data_t::real_half)));
-            scope_op_debug_print scope_dbg_print(__func__, "/to_fp32_sycl", dst, /*num_src=*/2,
-                                                 " : converting dst to fp32");
-            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
+                &alpha, src0_ptr, dpct::library_data_t::real_half, ne00,
+                src1_ptr, dpct::library_data_t::real_half, ne10, &beta,
+                dst_dd_i, dpct::library_data_t::real_float, ldc,
+                dpct::library_data_t::real_float)));
         }
     } else {
         ggml_sycl_pool_alloc<float> src0_ddq_as_f32(ctx.pool());