update CMakeLists.txt

ggml-ci
ggml-org · Sep 26, 2024 · 7c371fa · 7c371fa
1 parent 98ddb38
commit 7c371fa
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 29 deletions.
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
@@ -266,15 +266,23 @@ if (GGML_LLAMAFILE)
 endif()
 
 if (GGML_AMX)
-    message(STATUS "Using AMX")
+    if (CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
+    else()
+        set(GGML_AMX OFF)
+        message(WARNING "AMX requires gcc version > 11.0. Turning off GGML_AMX.")
+    endif()
 
-    add_compile_definitions(GGML_USE_AMX)
+    if (GGML_AMX)
+        message(STATUS "Using AMX")
 
-    file(GLOB   GGML_HEADERS_AMX "ggml-amx/*.h")
-    list(APPEND GGML_HEADERS_AMX "../include/ggml-amx.h")
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_AMX)
 
-    file(GLOB   GGML_SOURCES_AMX "ggml-amx/*.cpp")
-    list(APPEND GGML_SOURCES_AMX "ggml-amx.cpp")
+        file(GLOB   GGML_HEADERS_AMX "ggml-amx/*.h")
+        list(APPEND GGML_HEADERS_AMX "../include/ggml-amx.h")
+
+        file(GLOB   GGML_SOURCES_AMX "ggml-amx/*.cpp")
+        list(APPEND GGML_SOURCES_AMX "ggml-amx.cpp")
+    endif()
 endif()
 
 if (GGML_CUDA)

diff --git a/ggml/src/ggml-amx/mmq.cpp b/ggml/src/ggml-amx/mmq.cpp
@@ -2326,13 +2326,21 @@ size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) {
 
 // pack weight to vnni format
 void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+
+  size_t alloc_size = ggml_backend_amx_get_alloc_size(tensor);
+  GGML_ASSERT(alloc_size == size);
+
   const enum ggml_type TYPE = tensor->type;
 
   const int K = tensor->ne[0]; // ne0: in_features
   const int N = tensor->ne[1]; // ne1: out_features
 
+#if defined(_OPENMP)
   // the buffer ctx is not initialized when .set_tensor is called
   int n_threads = omp_get_num_threads();
+#else
+  int n_threads = 1;
+#endif
 
   GGML_DISPATCH_QTYPES(TYPE, [&] {
      convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads);
@@ -2425,8 +2433,6 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
     }
   });
 
-  //printf("### using amx kernels ... n_threads = %d\n", n_threads);
-
   if (M == 1) {
     // MB = 1 and handle 8 tiles in each block
     constexpr int kTilesN = 4;

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
@@ -46,19 +46,10 @@ int ggml_sve_cnt_b = 0;
 #undef GGML_USE_LLAMAFILE
 #endif
 
-// enable AMX only with OPENMP
-//#if !defined(__AMX_INT8__) || !defined(GGML_USE_OPENMP)
-//#undef GGML_USE_AMX
-//#endif
-
 #ifdef GGML_USE_LLAMAFILE
 #include <llamafile/sgemm.h>
 #endif
 
-//#ifdef GGML_USE_AMX
-//#include <ggml-amx/mmq.h>
-//#endif
-
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -12904,13 +12895,6 @@ static void ggml_compute_forward_mul_mat(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-//#if GGML_USE_AMX
-//    if (ggml_compute_forward_mul_mat_use_amx(dst)) {
-//        ggml_mul_mat_amx(dst, nth, ith, params->wdata, params->wsize);
-//        return;
-//    }
-//#endif
-
 #if GGML_USE_LLAMAFILE
     // broadcast factors
     const int64_t r2 = ne12 / ne02;

diff --git a/src/llama.cpp b/src/llama.cpp
@@ -32,11 +32,6 @@
 #  include "ggml-metal.h"
 #endif
 
-// enable AMX only with OPENMP
-#if !defined(__AMX_INT8__) || !defined(GGML_USE_OPENMP)
-#  undef GGML_USE_AMX
-#endif
-
 #ifdef GGML_USE_AMX
 #  include "ggml-amx.h"
 #endif