SVE detection fixes, submodule fixes, other cleanup

openvinotoolkit · Oct 29, 2024 · 2ea2be3 · 2ea2be3
1 parent 77b2980
commit 2ea2be3
Show file tree

Hide file tree

Showing 4 changed files with 6 additions and 9 deletions.
diff --git a/src/inference/src/system_conf.cpp b/src/inference/src/system_conf.cpp
@@ -22,6 +22,7 @@
 #    include <sys/auxv.h>
 #    define ARM_COMPUTE_CPU_FEATURE_HWCAP_FPHP    (1 << 9)
 #    define ARM_COMPUTE_CPU_FEATURE_HWCAP_ASIMDHP (1 << 10)
+#    define ARM_COMPUTE_CPU_FEATURE_HWCAP_SVE     (1 << 24)
 #elif defined(__APPLE__) && defined(__aarch64__)
 #    include <sys/sysctl.h>
 #    include <sys/types.h>
@@ -181,7 +182,7 @@ bool with_cpu_sve() {
 #    if !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
         !defined(__arm__) && defined(__aarch64__)
     const uint32_t hwcaps = getauxval(AT_HWCAP);
-    return hwcaps & HWCAP_SVE;
+    return hwcaps & ARM_COMPUTE_CPU_FEATURE_HWCAP_SVE;
 #    elif !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
         !defined(__aarch64__) && defined(__arm__)
     return false;

diff --git a/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/common.hpp b/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/common.hpp
@@ -34,6 +34,10 @@ static constexpr size_t vec_len_f32_avx2 = vec_len_avx2 / sizeof(float);
 static constexpr size_t vec_len_f32_neon = vec_len_neon / sizeof(float);
 static constexpr size_t vec_len_f16_neon = vec_len_neon / sizeof(ov::float16);
 
+#if defined(HAVE_SVE)
+static constexpr size_t vec_len_f32_sve = svcntw();
+#endif
+
 #ifdef HAVE_AVX512F
     inline __m512 cvt_bf16_to_fp32(const __m256i src) {
         __m512i y = _mm512_cvtepu16_epi32(src);

diff --git a/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/mha_single_token.cpp b/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/mha_single_token.cpp
@@ -63,7 +63,6 @@ void cvt_copy(TA* dst, TB* src, size_t n) {
     }
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     auto _dst = reinterpret_cast<float32_t*>(dst);
     size_t inc = vec_len_f32_sve;
     svbool_t pg = svptrue_b32();
@@ -86,7 +85,6 @@ void cvt_copy(TA* dst, TB* src, size_t n) {
         }
     }
 #else
-    int vec_len_f32_neon = 4;
     auto _dst = reinterpret_cast<float32_t*>(dst);
     if (std::is_same<TA, float>::value && std::is_same<TB, float>::value) {
         for (; i + vec_len_f32_neon <= n; i += vec_len_f32_neon) {
@@ -123,7 +121,6 @@ static void attn_acc_value(float* out, float weight, T* v, size_t S, float* scal
     }
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     auto _v = reinterpret_cast<float32_t*>(v);
     svfloat32_t attn_w_vec_fp32 = svdup_n_f32(weight);
     size_t inc = vec_len_f32_sve;
@@ -404,7 +401,6 @@ static float sum_q_head(T* a, size_t n) {
     sum = _mm256_cvtss_f32(vsum0);
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     svfloat32_t sum0 = svdup_n_f32(0.0f);
     svfloat32_t sum1 = svdup_n_f32(0.0f);
     svfloat32_t sum2 = svdup_n_f32(0.0f);
@@ -448,7 +444,6 @@ static float sum_q_head(T* a, size_t n) {
     float32_t sum_3 = svaddv_f32(pg, sum3);
     sum = static_cast<float>(sum_0 + sum_1 + sum_2 + sum_3);
 #else
-    size_t vec_len_f32_neon = 4;
     float32x4_t vsum0 = vdupq_n_f32(0.0f);
     float32x4_t vsum1 = vdupq_n_f32(0.0f);
     float32x4_t vsum2 = vdupq_n_f32(0.0f);
@@ -589,7 +584,6 @@ static float dot_product(TA* a, TB* b, size_t n, float* scale, float* zp, float*
 
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     svbool_t pg = svptrue_b32();
     svfloat32_t sum0 = svdup_n_f32(0.0f);
     svfloat32_t sum1 = svdup_n_f32(0.0f);
@@ -940,7 +934,6 @@ static void attn_reduce(T* dst, float* temp, size_t M, size_t S, size_t temp_str
     }
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     auto _dst = reinterpret_cast<float32_t*>(dst);
     size_t inc = vec_len_f32_sve;
     svbool_t pg = svptrue_b32();

diff --git a/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/softmax_kernel.hpp b/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/softmax_kernel.hpp
@@ -806,7 +806,6 @@ inline void multiply_scalar(float* a, float* a_dst, const float val, const size_
     }
 #elif defined(OPENVINO_ARCH_ARM64)
 #if defined(HAVE_SVE)
-    size_t vec_len_f32_sve = svcntw();
     svfloat32_t v_scale = svdup_n_f32(val);
     size_t inc = vec_len_f32_sve;
     svbool_t pg = svptrue_b32();