fix(ggml): repair builds without F16C

2 files changed, 56 insertions, 0 deletions

diff --git a/sci-ml/ggml/files/ggml-0.24.0-const-fp16-loads.patch b/sci-ml/ggml/files/ggml-0.24.0-const-fp16-loads.patch
new file mode 100644
index 000000000000..051d179aeb0c
--- /dev/null
+++ b/sci-ml/ggml/files/ggml-0.24.0-const-fp16-loads.patch
@@ -0,0 +1,54 @@
+Fix x86 FP16 fallback loads when F16C is disabled.
+
+These helpers only read their inputs, including const quantized block data.
+Accept const pointers in both AVX and AVX512 fallback implementations.
+Give the repack-local eight-value loader a distinct name to avoid colliding
+with the const overload already defined by simd-mappings.h for AVX.
+
+--- a/src/ggml-cpu/arch/x86/repack.cpp
++++ b/src/ggml-cpu/arch/x86/repack.cpp
+@@ -36,7 +36,7 @@
+ #define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     _mm256_cvtph_ps(_mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask))
+ #else
+ #if defined(__AVX512F__)
+-static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
++static inline __m512 __avx512_f32cx8x2_load(const ggml_fp16_t *x, const ggml_fp16_t *y) {
+     float tmp[16];
+ 
+     for (int i = 0; i < 8; i++) {
+@@ -64,7 +64,7 @@
+     return _mm512_loadu_ps(tmp);
+ }
+ #endif
+-static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
++static inline __m256 __avx_repack_f32cx8_load(const ggml_fp16_t *x) {
+     float tmp[8];
+ 
+     for (int i = 0; i < 8; i++) {
+@@ -73,7 +73,7 @@
+ 
+     return _mm256_loadu_ps(tmp);
+ }
+-static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
++static inline __m256 __avx_repeat_f32cx8_load(const ggml_fp16_t *x) {
+     float tmp[8];
+ 
+     for (int i = 0; i < 4; i++) {
+@@ -83,7 +83,7 @@
+ 
+     return _mm256_loadu_ps(tmp);
+ }
+-static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrangeMask) {
++static inline __m256 __avx_rearranged_f32cx8_load(const ggml_fp16_t *x, __m128i arrangeMask) {
+     uint16_t tmphalf[8];
+     float tmp[8];
+ 
+@@ -95,7 +95,7 @@
+     return _mm256_loadu_ps(tmp);
+ }
+ 
+-#define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
++#define GGML_F32Cx8_LOAD(x)     __avx_repack_f32cx8_load(x)
+ #define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
+ #define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     __avx_rearranged_f32cx8_load(x, arrangeMask)
+ #if defined(__AVX512F__)
diff --git a/sci-ml/ggml/ggml-0.24.0.ebuild b/sci-ml/ggml/ggml-0.24.0.ebuild
index 6ca110bcc73b..e4e4ef5d1ca9 100644
--- a/sci-ml/ggml/ggml-0.24.0.ebuild
+++ b/sci-ml/ggml/ggml-0.24.0.ebuild
@@ -56,6 +56,8 @@ DEPEND="${RDEPEND}
 "
 BDEPEND="vulkan? ( media-libs/shaderc )"
 
+PATCHES=( "${FILESDIR}/${P}-const-fp16-loads.patch" )
+
 pkg_pretend() {
 	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
 }