ggml-cpu : remove unnecesary arm feature detection (llama/14281)

Support for Arm runtime feature detection has now been added to GGML_CPU_ALL_VARIANTS. This removes the old and not very functional code.

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -256,7 +256,6 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
     const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
 
     for (int c = 0; c < nc; c += ncols_interleaved) {
@@ -294,7 +293,6 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         s += ncols_interleaved;
     }
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
     float sumf[4];
     int sumi;
@@ -341,7 +339,6 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
     const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
 
     for (int c = 0; c < nc; c += ncols_interleaved) {
@@ -384,7 +381,6 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         s += ncols_interleaved;
     }
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
     float sumf[4];
     int sumi;
@@ -432,7 +428,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -547,7 +543,6 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
     const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
     const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     float * res_ptr = s;
@@ -594,7 +589,6 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
         vst1q_f32(res_ptr + x * 4, sumf);
     }
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     {
         float sumf[4];
@@ -643,8 +637,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -1101,7 +1094,6 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     {
         float sumf[4][4];
@@ -1160,7 +1152,6 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -1557,7 +1548,6 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     float sumf[4][4];
     int sumi;
@@ -1615,7 +1605,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2083,7 +2073,6 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
     const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
 
     for (int y = 0; y < nr / 4; y++) {
@@ -2135,7 +2124,6 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
         }
     }
     return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     {
         float sumf[4][4];

ggml/src/ggml-cpu/ggml-cpu.c

@@ -74,13 +74,8 @@
 
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
-    int has_neon;
-    int has_dotprod;
-    int has_i8mm;
-    int has_sve;
     int sve_cnt;
-    int has_sme;
+} ggml_arm_arch_features = { 0 };
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
 #endif
 
 
@@ -678,87 +673,15 @@ bool ggml_is_numa(void) {
 
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__)
+#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-    uint32_t hwcap = getauxval(AT_HWCAP);
-    uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
-    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
-    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
-
-#if defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-#elif defined(__APPLE__)
-    int oldp = 0;
-    size_t size = sizeof(oldp);
-    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_neon = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_dotprod = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_i8mm = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_sme = oldp;
-
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#else
-// Run-time CPU feature detection not implemented for this platform, fallback to compile time
-#if defined(__ARM_NEON)
-    ggml_arm_arch_features.has_neon = 1;
-#else
-    ggml_arm_arch_features.has_neon = 0;
-#endif
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    ggml_arm_arch_features.has_i8mm = 1;
-#else
-    ggml_arm_arch_features.has_i8mm = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.has_sve = 1;
-    ggml_arm_arch_features.sve_cnt = 16;
-#else
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
-    ggml_arm_arch_features.has_sme = 1;
-#else
-    ggml_arm_arch_features.has_sme = 0;
-#endif
-#endif
 }
-#endif
+
+#endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -3448,7 +3371,7 @@ int ggml_cpu_has_vxe(void) {
 
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return ggml_arm_arch_features.has_neon;
+    return 1;
 #else
     return 0;
 #endif
@@ -3456,7 +3379,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
     return 0;
 #endif
@@ -3464,7 +3387,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
     return 0;
 #endif
@@ -3472,7 +3395,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
     return 0;
 #endif
@@ -3488,7 +3411,7 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
     return 0;
 #endif