Help! Failed to build, encounter some "invalid conversion " errors

[bluven]

I executed make tiny.en and got a lot of 'invalid conversion' errors:

https_proxy=http://192.168.1.121:1080 make tiny.en
bash ./models/download-ggml-model.sh tiny.en
Downloading ggml model tiny.en from 'https://huggingface.co/ggerganov/whisper.cpp' ...
ggml-tiny.en.bin                                            100%[=========================================================================================================================================>]  74.10M   232KB/s    in 6m 41s  
Done! Model 'tiny.en' saved in '/home/bluven/Workspace/whisper.cpp/models/ggml-tiny.en.bin'
You can now use it like this:

  $ ./main -m /home/bluven/Workspace/whisper.cpp/models/ggml-tiny.en.bin -f samples/jfk.wav

cmake -B build
-- ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- Including CPU backend
-- x86 detected
-- Adding CPU backend variant ggml-cpu: -march=native 
-- Configuring done (0.1s)
-- Generating done (0.0s)
-- Build files have been written to: /home/bluven/Workspace/whisper.cpp/build
cmake --build build --config Release
gmake[1]: Entering directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[2]: Entering directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[3]: Entering directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[3]: Leaving directory '/home/bluven/Workspace/whisper.cpp/build'
[ 17%] Built target ggml-base
gmake[3]: Entering directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[3]: Leaving directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[3]: Entering directory '/home/bluven/Workspace/whisper.cpp/build'
[ 20%] Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu-aarch64.cpp.o
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp: In function ‘void ggml_gemv_q4_0_8x8_q8_0(int, float*, size_t, const void*, const void*, int, int)’:
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:860:82: error: invalid conversion from ‘const ggml_half*’ {aka ‘const short unsigned int*’} to ‘ggml_fp16_t*’ {aka ‘short unsigned int*’} [-fpermissive]
  860 |                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
      |                                                                         ~~~~~~~~~^
      |                                                                                  |
      |                                                                                  const ggml_half* {aka const short unsigned int*}
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:150:85: note: in definition of macro ‘GGML_F32Cx8_REARRANGE_LOAD’
  150 | #define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     __avx_rearranged_f32cx8_load(x, arrangeMask)
      |                                                                                     ^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:136:64: note:   initializing argument 1 of ‘__m256 __avx_rearranged_f32cx8_load(ggml_fp16_t*, __m128i)’
  136 | static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrangeMask) {
      |                                                   ~~~~~~~~~~~~~^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp: In function ‘void ggml_gemm_q4_0_8x8_q8_0(int, float*, size_t, const void*, const void*, int, int)’:
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:2988:76: error: invalid conversion from ‘const ggml_half*’ {aka ‘const short unsigned int*’} to ‘ggml_fp16_t*’ {aka ‘short unsigned int*’} [-fpermissive]
 2988 |                     const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
      |                                                                   ~~~~~~~~~^
      |                                                                            |
      |                                                                            const ggml_half* {aka const short unsigned int*}
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:148:51: note: in definition of macro ‘GGML_F32Cx8_LOAD’
  148 | #define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
      |                                                   ^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:117:53: note:   initializing argument 1 of ‘__m256 __avx_f32cx8_load(ggml_fp16_t*)’
  117 | static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
      |                                        ~~~~~~~~~~~~~^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:3065:92: error: invalid conversion from ‘const ggml_half*’ {aka ‘const short unsigned int*’} to ‘ggml_fp16_t*’ {aka ‘short unsigned int*’} [-fpermissive]
 3065 |                         const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
      |                                                                              ~~~~~~~~~~~~~~^
      |                                                                                            |
      |                                                                                            const ggml_half* {aka const short unsigned int*}
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:149:75: note: in definition of macro ‘GGML_F32Cx8_REPEAT_LOAD’
  149 | #define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
      |                                                                           ^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:126:60: note:   initializing argument 1 of ‘__m256 __avx_repeat_f32cx8_load(ggml_fp16_t*)’
  126 | static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
      |                                               ~~~~~~~~~~~~~^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:3152:76: error: invalid conversion from ‘const ggml_half*’ {aka ‘const short unsigned int*’} to ‘ggml_fp16_t*’ {aka ‘short unsigned int*’} [-fpermissive]
 3152 |                     const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
      |                                                                   ~~~~~~~~~^
      |                                                                            |
      |                                                                            const ggml_half* {aka const short unsigned int*}
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:148:51: note: in definition of macro ‘GGML_F32Cx8_LOAD’
  148 | #define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
      |                                                   ^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:117:53: note:   initializing argument 1 of ‘__m256 __avx_f32cx8_load(ggml_fp16_t*)’
  117 | static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
      |                                        ~~~~~~~~~~~~~^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:3230:83: error: invalid conversion from ‘const ggml_half*’ {aka ‘const short unsigned int*’} to ‘ggml_fp16_t*’ {aka ‘short unsigned int*’} [-fpermissive]
 3230 |                     const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
      |                                                                          ~~~~~~~~~^
      |                                                                                   |
      |                                                                                   const ggml_half* {aka const short unsigned int*}
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:149:75: note: in definition of macro ‘GGML_F32Cx8_REPEAT_LOAD’
  149 | #define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
      |                                                                           ^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:126:60: note:   initializing argument 1 of ‘__m256 __avx_repeat_f32cx8_load(ggml_fp16_t*)’
  126 | static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
      |                                               ~~~~~~~~~~~~~^
/home/bluven/Workspace/whisper.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp:2504:23: warning: unused variable ‘loadMask’ [-Wunused-variable]
 2504 |         const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
      |                       ^~~~~~~~
gmake[3]: *** [ggml/src/CMakeFiles/ggml-cpu.dir/build.make:104: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu-aarch64.cpp.o] Error 1
gmake[3]: Leaving directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[2]: *** [CMakeFiles/Makefile2:318: ggml/src/CMakeFiles/ggml-cpu.dir/all] Error 2
gmake[2]: Leaving directory '/home/bluven/Workspace/whisper.cpp/build'
gmake[1]: *** [Makefile:136: all] Error 2
gmake[1]: Leaving directory '/home/bluven/Workspace/whisper.cpp/build'
make: *** [Makefile:56: tiny.en] Error 2

The verson is v1.7.4 and I buiit it in a virtualbox machine which has no GPU. I know nothing about neither c++ programming nor AI, there were also no anwsers availabe from Google, what should I do to fix those errors?

Thanks for any help.

[personalizedrefrigerator]

I'm also experiencing this in VirtualBox. I was able to build whisper.cpp by:

1. Changing the first parameters to const in __avx_rearranged_f32cx8_load, __avx_repeat_f32cx8_load, and __avx_f32cx8_load (in ggml-cpu-aarch64.cpp).
   • This fixed the "invalid conversion" errors.
2. Undefining the __AVX2__ (advanced vector extensions) macro
   • This fixed another error: inlining failed in call to ‘always_inline’ ‘_m m256_fmadd_ps’: target specific option mismatch

Diff (based on the 1.7.4 tag)

diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h
index 3aa71ba..452e59c 100644
--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -3,6 +3,10 @@
 #include "ggml.h"
 #include "ggml-backend.h"
 
+// VirtualBox: __AVX2__ is defined, but certain extensions cause compiler errors.
+// For now, undefine it:
+#undef __AVX2__
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
index 622c63f..b2767b4 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
@@ -18,6 +18,11 @@
 
 #include "ggml-cpu-aarch64.h"
 
+
+// VirtualBox: __AVX2__ is defined, but certain extensions cause compiler errors.
+// For now, undefine it:
+#undef __AVX2__
+
 // TODO: move to include file?
 template <int K> constexpr int QK_0() {
     if constexpr (K == 4) {
@@ -114,7 +119,7 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
     return _mm512_loadu_ps(tmp);
 }
 #endif
-static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+static inline __m256 __avx_f32cx8_load(const ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
@@ -123,7 +128,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
 
     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++) {
@@ -133,7 +138,7 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
 
     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];
 
diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c
index 8e14722..63530f8 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-quants.c
+++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c
@@ -31,6 +31,10 @@
 // some compilers don't provide _mm256_set_m128i, e.g. gcc 7
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
+// VirtualBox: __AVX2__ is defined, but certain extensions cause compiler errors.
+// For now, undefine it:
+#undef __AVX2__
+
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 // multiply int8_t, add results pairwise twice
 static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {

Note: I haven't verified that these changes don't cause issues at runtime.