LLM inference in C/C++
C++ ggml
vulkan: fix rdna4 mat_vec tuning (#29934)
Ruben Ortlam16c163d561b9
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cppindex bfc65bed8a..a4c7bcc8c3 100644--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp@@ -2889,11 +2889,11 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {rm_stdq = 2;rm_stdq_int = 2;}- // RDNA3: above four columns, static 4 rows for all types bench faster than the default- const bool is_rdna3 = device->vendor_id == VK_VENDOR_ID_AMD && device->architecture == AMD_RDNA3;- auto const &rm_int_n = [&](uint32_t rows, uint32_t i) { return (is_rdna3 && i >= 4) ? 4u : rows; };- // RDNA3: Static 4 rows for all types bench faster than the default- auto const &rm_id = [&](uint32_t rows) { return is_rdna3 ? 4u : rows; };+ // RDNA3/4: above four columns, static 4 rows for all types bench faster than the default+ const bool is_rdna3_or_4 = device->vendor_id == VK_VENDOR_ID_AMD && (device->architecture == AMD_RDNA3 || device->architecture == AMD_RDNA4);+ auto const &rm_int_n = [&](uint32_t rows, uint32_t i) { return (is_rdna3_or_4 && i >= 4) ? 4u : rows; };+ // RDNA3/4: Static 4 rows for all types bench faster than the default+ auto const &rm_id = [&](uint32_t rows) { return is_rdna3_or_4 ? 4u : rows; };uint32_t rm_iq = 2 * rm_kq;const bool use_subgroups = device->subgroup_arithmetic;@@ -3081,7 +3081,7 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {#if !defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)GGML_UNUSED(rm_stdq_int);GGML_UNUSED(rm_kq_int);- GGML_UNUSED(is_rdna3);+ GGML_UNUSED(is_rdna3_or_4);GGML_UNUSED(rm_int_n);GGML_UNUSED(rm_id);GGML_UNUSED(rm_iq_int);