LLM inference in C/C++
C++ ggml
ggml : fix conv 2d dw (#25490)
Georgi Gerganov5c3a58686036
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cppindex e1f47e5ab0..df0028cf15 100644--- a/ggml/src/ggml-cpu/ops.cpp+++ b/ggml/src/ggml-cpu/ops.cpp@@ -7324,11 +7324,12 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(#ifdef GGML_SIMDint64_t c_pkg_end = 0;+ int64_t pkg_size = GGML_F32_EPR;if (knl_type == GGML_TYPE_F32) {#if defined(__ARM_FEATURE_SVE)- const int64_t pkg_size = svcntw();+ pkg_size = svcntw();#else- const int64_t pkg_size = GGML_F32_EPR;+ pkg_size = GGML_F32_EPR;#endifc_pkg_end = (c / pkg_size) * pkg_size;}@@ -7345,7 +7346,7 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;#ifdef GGML_SIMD- for (int64_t c_i = 0; c_i < c_pkg_end; c_i += GGML_F32_EPR) {+ for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {GGML_F32_VEC sum = GGML_F32_VEC_ZERO;for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {const int64_t src_y = src_y_base + knl_y * p.dilation_y;diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cuindex bfe449878e..98816f885c 100644--- a/ggml/src/ggml-cuda/ggml-cuda.cu+++ b/ggml/src/ggml-cuda/ggml-cuda.cu@@ -4917,7 +4917,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const gcase GGML_OP_IM2COL:case GGML_OP_IM2COL_3D:case GGML_OP_CONV_2D:+ return true;case GGML_OP_CONV_2D_DW:+ return op->src[0]->type == GGML_TYPE_F32;case GGML_OP_CONV_TRANSPOSE_2D:case GGML_OP_POOL_2D:return true;