Detect SSE version and enable the correct set of SIMD intrinsics

diff --git a/basis/cpu/x86/32/32.factor b/basis/cpu/x86/32/32.factor
index e9388e300d0acf9f37a8e2fcb2de2af36222bd73..2ed461927c7b1a081cc02dc506df410e0e3a98e3 100755 (executable)
--- a/basis/cpu/x86/32/32.factor
+++ b/basis/cpu/x86/32/32.factor
@@ -295,22 +295,6 @@ os windows? [
     4 "double" c-type (>>align)
 ] unless
 
-USING: cpu.x86.features cpu.x86.features.private ;
-
-"-no-sse2" (command-line) member? [
-    [ { check_sse2 } compile ] with-optimizer
-
-    "Checking if your CPU supports SSE2..." print flush
-    sse2? [
-        " - yes" print
-        enable-sse2
-        [
-            sse2? [
-                "This image was built to use SSE2, which your CPU does not support." print
-                "You will need to bootstrap Factor again." print
-                flush
-                1 exit
-            ] unless
-        ] "cpu.x86" add-init-hook
-    ] [ " - no" print ] if
-] unless
+USE: vocabs.loader
+
+"cpu.x86.features" require

diff --git a/basis/cpu/x86/64/64.factor b/basis/cpu/x86/64/64.factor
index a7a4e783c3f56bb9e2f50a07adc4d8afb287b353..df9287162cfcea7d030d569e65cf0e0ab7a6ca94 100644 (file)
--- a/basis/cpu/x86/64/64.factor
+++ b/basis/cpu/x86/64/64.factor
@@ -221,12 +221,11 @@ enable-alien-4-intrinsics
 ! Enable fast calling of libc math functions
 enable-float-functions
 
-! SSE2 is always available on x86-64.
-enable-sse2
-
 USE: vocabs.loader
 
 {
     { [ os unix? ] [ "cpu.x86.64.unix" require ] }
     { [ os winnt? ] [ "cpu.x86.64.winnt" require ] }
 } cond
+
+"cpu.x86.features" require

diff --git a/basis/cpu/x86/features/features.factor b/basis/cpu/x86/features/features.factor
index bc4818d6af239ec98219a7220cacbf23a10ceaa8..878fb592b9611f4b51084f9292c002b422b897d8 100644 (file)
--- a/basis/cpu/x86/features/features.factor
+++ b/basis/cpu/x86/features/features.factor
@@ -1,21 +1,29 @@
 ! Copyright (C) 2009 Slava Pestov.
 ! See http://factorcode.org/license.txt for BSD license.
-USING: system kernel math alien.syntax ;
+USING: system kernel math alien.syntax combinators locals init io
+cpu.x86 compiler compiler.units accessors ;
 IN: cpu.x86.features
 
 <PRIVATE
 
-FUNCTION: bool check_sse2 ( ) ;
+FUNCTION: int sse_version ( ) ;
 
 FUNCTION: longlong read_timestamp_counter ( ) ;
 
 PRIVATE>
 
-HOOK: sse2? cpu ( -- ? )
+ALIAS: sse-version sse_version
 
-M: x86.32 sse2? check_sse2 ;
-
-M: x86.64 sse2? t ;
+: sse-string ( version -- string )
+    {
+        { 00 [ "no SSE" ] }
+        { 10 [ "SSE1" ] }
+        { 20 [ "SSE2" ] }
+        { 30 [ "SSE3" ] }
+        { 33 [ "SSSE3" ] }
+        { 41 [ "SSE4.1" ] }
+        { 42 [ "SSE4.2" ] }
+    } case ;
 
 HOOK: instruction-count cpu ( -- n )
 
@@ -23,3 +31,36 @@ M: x86 instruction-count read_timestamp_counter ;
 
 : count-instructions ( quot -- n )
     instruction-count [ call ] dip instruction-count swap - ; inline
+
+USING: cpu.x86.features cpu.x86.features.private ;
+
+:: install-sse-check ( version -- )
+    [
+        sse-version version < [
+            "This image was built to use " write
+            version sse-string write
+            " but your CPU supports " write
+            sse-version sse-string write "." print
+            "You will need to bootstrap Factor again." print
+            flush
+            1 exit
+        ] when
+    ] "cpu.x86" add-init-hook ;
+
+: enable-sse ( version -- )
+    {
+        { 00 [ ] }
+        { 10 [ ] }
+        { 20 [ enable-sse2 ] }
+        { 30 [ enable-sse3 ] }
+        { 33 [ enable-sse3 ] }
+        { 41 [ enable-sse3 ] }
+        { 42 [ enable-sse3 ] }
+    } case ;
+
+[ { sse_version } compile ] with-optimizer
+
+"Checking for multimedia extensions: " write sse-version
+[ sse-string write " detected" print ]
+[ install-sse-check ]
+[ enable-sse ] tri

diff --git a/basis/cpu/x86/x86.factor b/basis/cpu/x86/x86.factor
index af548a1f2b6a398400a156a560e18d2bb99900d1..602dfd65412f28bd12adfea2dbfef256861513b5 100644 (file)
--- a/basis/cpu/x86/x86.factor
+++ b/basis/cpu/x86/x86.factor
@@ -700,6 +700,11 @@ M: x86 small-enough? ( n -- ? )
 : enable-sse2 ( -- )
     enable-float-intrinsics
     enable-fsqrt
-    enable-float-min/max ;
+    enable-float-min/max
+    enable-sse2-simd ;
+
+: enable-sse3 ( -- )
+    enable-sse2
+    enable-sse3-simd ;
 
 enable-min/max

diff --git a/basis/math/vectors/simd/simd.factor b/basis/math/vectors/simd/simd.factor
index 39ce72356bf40155c91863e424520e667d8b6ab1..a36eee9081a021a1328ad6ac00986dd8cc613810 100644 (file)
--- a/basis/math/vectors/simd/simd.factor
+++ b/basis/math/vectors/simd/simd.factor
@@ -4,7 +4,7 @@ USING: accessors alien.c-types byte-arrays cpu.architecture
 generalizations kernel math math.functions math.vectors
 math.vectors.simd.functor math.vectors.specialization parser
 prettyprint.custom sequences sequences.private
-specialized-arrays.double locals assocs literals ;
+specialized-arrays.double locals assocs literals words fry ;
 IN: math.vectors.simd
 
 <PRIVATE
@@ -69,8 +69,20 @@ PRIVATE>
 
 <PRIVATE
 
-:: simd-vector-words ( class ctor elt-type assoc -- )
-    class elt-type assoc {
+: supported-simd-ops ( assoc -- assoc' )
+    {
+        { v+ (simd-v+) }
+        { v- (simd-v-) }
+        { v* (simd-v*) }
+        { v/ (simd-v/) }
+        { vmin (simd-vmin) }
+        { vmax (simd-vmax) }
+        { sum (simd-sum) }
+    } [ nip "intrinsic" word-prop ] assoc-filter
+    '[ drop _ key? ] assoc-filter ;
+
+:: high-level-ops ( ctor -- assoc )
+    {
         { vneg [ [ dup v- ] keep v- ] }
         { v. [ v* sum ] }
         { n+v [ [ ctor execute ] dip v+ ] }
@@ -85,7 +97,10 @@ PRIVATE>
         { norm [ norm-sq sqrt ] }
         { normalize [ dup norm v/n ] }
         { distance [ v- norm ] }
-    } assoc-union
+    } ;
+
+:: simd-vector-words ( class ctor elt-type assoc -- )
+    class elt-type assoc supported-simd-ops ctor high-level-ops assoc-union
     specialize-vector-words ;
 
 PRIVATE>

diff --git a/vm/cpu-x86.32.S b/vm/cpu-x86.32.S
index a8797121901162c5a957dc78387287f76cb7c4ad..10cd7f23833b01a51320e3b38bc6682a886b031d 100644 (file)
--- a/vm/cpu-x86.32.S
+++ b/vm/cpu-x86.32.S
@@ -44,17 +44,6 @@ DEF(void,set_callstack,(F_STACK_FRAME *to, F_STACK_FRAME *from, CELL length, voi
        add $12,%esp                       /* pop args from the stack */
        ret                                /* return _with new stack_ */
 
-/* cpu.x86.32 calls this */
-DEF(bool,check_sse2,(void)):
-       push %ebx
-       mov $1,%eax
-       cpuid
-       shr $26,%edx
-       and $1,%edx
-       pop %ebx
-       mov %edx,%eax
-       ret
-
 DEF(long long,read_timestamp_counter,(void)):
        rdtsc
        ret
@@ -72,6 +61,5 @@ DEF(void,primitive_inline_cache_miss_tail,(void)):
 
 #ifdef WINDOWS
        .section .drectve
-       .ascii " -export:check_sse2"
        .ascii " -export:read_timestamp_counter"
 #endif

diff --git a/vm/cpu-x86.S b/vm/cpu-x86.S
index e83bb0fd7d97e9ab2860dec5086fe933fa7df8a5..09a423d6b13ac244d610d3373218deeb7e1a15d3 100644 (file)
--- a/vm/cpu-x86.S
+++ b/vm/cpu-x86.S
@@ -68,7 +68,44 @@ DEF(F_FASTCALL void,lazy_jit_compile,(CELL quot)):
        add $STACK_PADDING,STACK_REG
         jmp *QUOT_XT_OFFSET(ARG0)    /* Call the quotation */
 
+/* cpu.x86.features calls this */
+DEF(bool,sse_version,(void)):
+       mov $0x1,RETURN_REG
+       cpuid
+    test $0x100000,%ecx
+    jnz sse_42
+    test $0x80000,%ecx
+    jnz sse_41
+    test $0x200,%ecx
+    jnz ssse_3
+    test $0x1,%ecx
+    jnz sse_3
+    test $0x4000000,%edx
+    jnz sse_2
+    test $0x2000000,%edx
+    jnz sse_1
+    mov $0,%eax
+    ret
+sse_42:
+    mov $42,RETURN_REG
+    ret
+sse_41:
+    mov $41,RETURN_REG
+    ret
+ssse_3:
+    mov $33,RETURN_REG
+    ret
+sse_3:
+    mov $30,RETURN_REG
+    ret
+sse_2:
+    mov $20,RETURN_REG
+    ret
+sse_1:
+    mov $10,RETURN_REG
+    ret
 #ifdef WINDOWS
        .section .drectve
+       .ascii " -export:sse_version"
        .ascii " -export:c_to_factor"
 #endif