AVX-512 detection for AMD

lsp-plugins · Dec 12, 2024 · b2cf1ea · b2cf1ea
1 parent 5fda461
commit b2cf1ea
Show file tree

Hide file tree

Showing 4 changed files with 54 additions and 0 deletions.
diff --git a/include/private/dsp/arch/x86/cpuid.h b/include/private/dsp/arch/x86/cpuid.h
@@ -83,6 +83,16 @@
 #define X86_CPUID7_INTEL_ECX_AVX512VBMI         (1 << 1)
 
 #define X86_CPUID7_AMD_EBX_AVX2                 (1 << 5)
+#define X86_CPUID7_AMD_EBX_AVX512F              (1 << 16)
+#define X86_CPUID7_AMD_EBX_AVX512DQ             (1 << 17)
+#define X86_CPUID7_AMD_EBX_AVX512IFMA           (1 << 21)
+#define X86_CPUID7_AMD_EBX_AVX512PF             (1 << 26)
+#define X86_CPUID7_AMD_EBX_AVX512ER             (1 << 27)
+#define X86_CPUID7_AMD_EBX_AVX512CD             (1 << 28)
+#define X86_CPUID7_AMD_EBX_AVX512BW             (1 << 30)
+#define X86_CPUID7_AMD_EBX_AVX512VL             (1 << 31)
+
+#define X86_CPUID7_AMD_ECX_AVX512VBMI           (1 << 1)
 
 //-------------------------------------------------------------------------
 // Function 80000001

diff --git a/include/private/dsp/arch/x86/features.h b/include/private/dsp/arch/x86/features.h
@@ -90,6 +90,7 @@
                 FEAT_FAST_MOVS,         // Processor implements optimized MOVS instruction
                 FEAT_FAST_AVX,          // Fast AVX implementation
                 FEAT_FAST_FMA3,         // Fast FMA3 implementation
+                FEAT_FAST_AVX512,       // Fast AVX-512 implementation
                 FEAT_BELOW_ZEN3         // CPU has AMD architecture and is below Zen3
             };
 

diff --git a/src/main/x86/avx512.cpp b/src/main/x86/avx512.cpp
@@ -84,6 +84,11 @@
 
             void dsp_init(const cpu_features_t *f)
             {
+                // Enable AVX-512 only for CPUs that really support it well
+                const bool favx512  = feature_check(f, FEAT_FAST_AVX512);
+                if (!favx512)
+                    return;
+
                 const bool vl = (f->features & (CPU_OPTION_AVX512F | CPU_OPTION_AVX512VL)) ==
                                 (CPU_OPTION_AVX512F | CPU_OPTION_AVX512VL);
 

diff --git a/src/main/x86/x86.cpp b/src/main/x86/x86.cpp
@@ -299,6 +299,30 @@
                         if (info.ebx & X86_CPUID7_AMD_EBX_AVX2)
                             f->features     |= CPU_OPTION_AVX2;
                     }
+
+                    // Additional check for AVX512 support
+                    if ((xcr0 & XCR_FLAGS_AVX512) == XCR_FLAGS_AVX512)
+                    {
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512F)
+                            f->features     |= CPU_OPTION_AVX512F;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512DQ)
+                            f->features     |= CPU_OPTION_AVX512DQ;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512IFMA)
+                            f->features     |= CPU_OPTION_AVX512IFMA;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512PF)
+                            f->features     |= CPU_OPTION_AVX512PF;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512ER)
+                            f->features     |= CPU_OPTION_AVX512ER;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512CD)
+                            f->features     |= CPU_OPTION_AVX512CD;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512BW)
+                            f->features     |= CPU_OPTION_AVX512BW;
+                        if (info.ebx & X86_CPUID7_AMD_EBX_AVX512VL)
+                            f->features     |= CPU_OPTION_AVX512VL;
+
+                        if (info.ecx & X86_CPUID7_AMD_ECX_AVX512VBMI)
+                            f->features     |= CPU_OPTION_AVX512VBMI;
+                    }
                 }
 
                 // FUNCTION 0x80000001
@@ -542,6 +566,20 @@
                         }
                         break;
 
+                    case FEAT_FAST_AVX512:
+                        if (f->vendor == CPU_VENDOR_INTEL) // Any Intel CPU seems to be good enough with AVX-512
+                            return true;
+                        // Only starting with ZEN 1 architecture AMD's implementation of AVX is fast enough
+                        if ((f->vendor == CPU_VENDOR_AMD) || (f->vendor == CPU_VENDOR_HYGON))
+                        {
+                            if (f->family < AMD_FAMILY_ZEN_3_4)
+                                return false;
+                            if (f->family == AMD_FAMILY_DHYANA)
+                                return false;
+                            return true;
+                        }
+                        break;
+
                     case FEAT_BELOW_ZEN3: // Test that this is AMD and below Zen 3 architecture
                         if ((f->vendor == CPU_VENDOR_AMD) || (f->vendor == CPU_VENDOR_HYGON))
                         {