ggml, ci : Windows ARM runner and build fixes (#5979)

* windows arm ci

* fix `error C2078: too many initializers` with ggml_vld1q_u32 macro for MSVC ARM64

* fix `warning C4146: unary minus operator applied to unsigned type, result still unsigned`

* fix `error C2065: '__fp16': undeclared identifier`

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 9144f926628a51a40946c1418d22578b3f525f92..d39cd6bc338e0896f9ed31215d58bb9266eaaa4c 100644 (file)
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -425,6 +425,8 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'arm64'
+            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -520,7 +522,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900

diff --git a/ggml-impl.h b/ggml-impl.h
index c5637e4d45d8c5174472ed4c8cd86ab10f5d6dbe..e68b728775c414e4d2a24eaa81f2034c0141374f 100644 (file)
--- a/ggml-impl.h
+++ b/ggml-impl.h
@@ -53,26 +53,30 @@ extern "C" {
 //
 #include <arm_neon.h>
 
+typedef __fp16 ggml_fp16_internal_t;
+
 #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
 #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
 
 static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    __fp16 tmp;
+    ggml_fp16_internal_t tmp;
     memcpy(&tmp, &h, sizeof(ggml_fp16_t));
     return (float)tmp;
 }
 
 static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
     ggml_fp16_t res;
-    __fp16 tmp = f;
+    ggml_fp16_internal_t tmp = f;
     memcpy(&res, &tmp, sizeof(ggml_fp16_t));
     return res;
 }
 
 #else
 
+typedef uint16_t ggml_fp16_internal_t;
+
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else

diff --git a/ggml-quants.c b/ggml-quants.c
index f9a3d9fd229e186315d392502b7afa417e03a333..86b0764cbae18d9596227979c9cb85e4f875ce17 100644 (file)
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -9374,15 +9374,15 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 
             const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
             idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
-            const uint32x4_t aux32x4_0 = {iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                          iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]};
-            const uint32x4_t aux32x4_1 = {iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                          iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]};
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
             idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
-            const uint32x4_t aux32x4_2 = {iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                          iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]};
-            const uint32x4_t aux32x4_3 = {iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                          iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]};
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
 
 
             vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));

diff --git a/ggml.c b/ggml.c
index 80efa6f2ac9000df8567c50ee204584b4f1fed89..9a7bd1d8c527b19168e4769220eb9b847a45e165 100644 (file)
--- a/ggml.c
+++ b/ggml.c
@@ -857,7 +857,7 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F16x8              float16x8_t
     #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
     #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
-    #define GGML_F16x8_LOAD(x)      vld1q_f16((const __fp16 *)(x))
+    #define GGML_F16x8_LOAD(x)      vld1q_f16((const ggml_fp16_internal_t *)(x))
     #define GGML_F16x8_STORE        vst1q_f16
     #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
     #define GGML_F16x8_ADD          vaddq_f16
@@ -900,7 +900,7 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F32Cx4              float32x4_t
     #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
     #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
-    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
+    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const ggml_fp16_internal_t *)(x)))
     #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
     #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
     #define GGML_F32Cx4_ADD          vaddq_f32

diff --git a/llama.cpp b/llama.cpp
index 110e509ccd5b6cbcb7177e3bf66bf3dfc360ee00..f61ea791b5be23c298aad16f6246899dc43d0c1d 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -13980,7 +13980,7 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
             } else if (llama_is_user_defined_token(model->vocab, token)) {
                 std::string result = model->vocab.id_to_token[token].text;
                 if (length < (int) result.length()) {
-                    return -result.length();
+                    return -(int) result.length();
                 }
                 memcpy(buf, result.c_str(), result.length());
                 return result.length();
@@ -14015,7 +14015,7 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
             } else if (llama_is_user_defined_token(model->vocab, token)) {
                 std::string result = model->vocab.id_to_token[token].text;
                 if (length < (int) result.length()) {
-                    return -result.length();
+                    return -(int) result.length();
                 }
                 memcpy(buf, result.c_str(), result.length());
                 return result.length();