Merge branch 'develop' into betterPowerGEMVTail

.cirrus.yml

@@ -89,20 +89,21 @@ task:
       type: text/plain
 
 macos_instance:
-  image: ghcr.io/cirruslabs/macos-monterey-xcode:latest
+  image: ghcr.io/cirruslabs/macos-sonoma-xcode:latest
 task:
   name: AppleM1/LLVM armv7-androidndk xbuild
   compile_script:
-  - brew install android-ndk
+  - brew install --cask android-ndk
   - export #PATH=/opt/homebrew/opt/llvm/bin:$PATH
   - export #LDFLAGS="-L/opt/homebrew/opt/llvm/lib"
-  - export #CPPFLAGS="-I/opt/homebrew/opt/llvm/include" 
-  - ls /System/Volumes/Data/opt/homebrew
+  - export #CPPFLAGS="-I/opt/homebrew/opt/llvm/include"
+  - export ANDROID_NDK_HOME="/opt/homebrew/share/android-ndk"
+  - ls /opt/homebrew
   - ls -l /System/Volumes/Data/opt/homebrew/Caskroom/android-ndk
-  - find /System/Volumes/Data/opt/homebrew -name "armv7a-linux-androideabi*-ranlib"
+  - find /opt/homebrew -name "armv7a-linux-androideabi*-ranlib"
   - #export CC=/Applications/Xcode-13.4.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang
   - #export CFLAGS="-O2 -unwindlib=none -Wno-macro-redefined -isysroot /Applications/Xcode-13.4.1.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS16.0.sdk -arch arm64 -miphoneos-version-min=10.0"
-  - export CC=/System/Volumes/Data/opt/homebrew/Caskroom/android-ndk/26d/AndroidNDK*.app/Contents/NDK/toolchains/llvm/prebuilt/darwin-x86_64/bin/armv7a-linux-androideabi23-clang
+  - export CC=/System/Volumes/Data/opt/homebrew/Caskroom/android-ndk/27/AndroidNDK*.app/Contents/NDK/toolchains/llvm/prebuilt/darwin-x86_64/bin/armv7a-linux-androideabi23-clang
   - make TARGET=ARMV7 ARM_SOFTFP_ABI=1 NUM_THREADS=32 HOSTCC=clang NOFORTRAN=1 RANLIB="ls -l"
   always:
     config_artifacts:

README.md

@@ -85,6 +85,8 @@ Examples:
   make CC=loongcc FC=loongf95 HOSTCC=gcc TARGET=LOONGSON3A CROSS=1 CROSS_SUFFIX=mips64el-st-linux-gnu-   NO_LAPACKE=1 NO_SHARED=1 BINARY=32
   ```
 
+When compiling for a more modern CPU TARGET of the same architecture, e.g. TARGET=SKYLAKEX on a HASWELL host, option "CROSS=1" can be used to suppress the automatic invocation of the tests at the end of the build.
+
 ### Debug version
 
 A debug version can be built using `make DEBUG=1`.

cpuid_x86.c

@@ -1527,6 +1527,19 @@ int get_cpuname(void){
       break;
       case 10: //family 6 exmodel 10
         switch (model) {
+	  case 13: // Granite Rapids
+	  if(support_amx_bf16())
+	    return CPUTYPE_SAPPHIRERAPIDS;
+	  if(support_avx512_bf16())
+            return CPUTYPE_COOPERLAKE;	
+          if(support_avx512())
+            return CPUTYPE_SKYLAKEX;
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+            return CPUTYPE_SANDYBRIDGE;
+          else
+          return CPUTYPE_NEHALEM;
           case 5: // Comet Lake H and S
           case 6: // Comet Lake U
 	  case 10: // Meteor Lake
@@ -2352,8 +2365,22 @@ int get_coretype(void){
 
       case 10:
         switch (model) {
+	  case 13: // Granite Rapids
+	  if(support_amx_bf16())
+	    return CORE_SAPPHIRERAPIDS;
+	  if(support_avx512_bf16())
+            return CORE_COOPERLAKE;	
+          if(support_avx512())
+            return CORE_SKYLAKEX;
+          if(support_avx2())
+            return CORE_HASWELL;
+          if(support_avx())
+	    return CORE_SANDYBRIDGE;
+	  else
+	    return CORE_NEHALEM;
 	  case 5: // Comet Lake H and S
     	  case 6: // Comet Lake U
+	  case 10: // Meteor Lake
             if(support_avx())
   #ifndef NO_AVX2
               return CORE_HASWELL;
@@ -2362,6 +2389,7 @@ int get_coretype(void){
   #endif
             else
               return CORE_NEHALEM;
+	  case 0: // Meteor Lake
 	  case 7:// Rocket Lake
 #ifndef NO_AVX512
 	  if(support_avx512())

driver/others/blas_server.c

@@ -1076,6 +1076,8 @@ fprintf(STDERR, "Server[%2ld] Calculation started.  Mode = 0x%03x M = %3ld N=%3l
       main_status[cpu] = MAIN_RUNNING1;
 #endif
 
+if (buffer == NULL) blas_thread_buffer[cpu] = blas_memory_alloc(2);
+
 //For target LOONGSON3R5, applying an offset to the buffer is essential
 //for minimizing cache conflicts and optimizing performance.
 #if defined(ARCH_LOONGARCH64) && !defined(NO_AFFINITY)

exports/gensymbol

@@ -880,10 +880,8 @@ lapackobjs2c="$lapackobjs2c
 #    clatrs3
 
 lapackobjs2d="$lapackobjs2d
-    dgelqs
     dgelst
     dgeqp3rk
-    dgeqrs
     dlaqp2rk
     dlaqp3rk
     dlarmm
@@ -897,10 +895,8 @@ lapackobjs2d="$lapackobjs2d
 #    dlaqz4
 
 lapackobjs2z="$lapackobjs2z
-    zgelqs
     zgelst
     zgeqp3rk
-    zgeqrs
     zlaqp2rk
     zlaqp3rk
     zlatrs3
@@ -918,6 +914,7 @@ lapack_extendedprecision_objs="
 "
 
 lapack_deprecated_objsc="
+    cgelqs  cgeqrs
     cgegs   cggsvd
     cgegv   cggsvp
     cgelsx  clahrd
@@ -926,13 +923,16 @@ lapack_deprecated_objsc="
     "
 
 lapack_deprecated_objsd="
+    dgelqs  dgeqrs
     dgegs   dgeqpf
     dgegv   dggsvd
     dgelsx  dggsvp
              dlahrd
   dlatzm  dtzrqf"
 
 lapack_deprecated_objss="
+  sgelqs
+  sgeqrs
   sgelsx
   sgegs
   sgegv
@@ -945,6 +945,8 @@ lapack_deprecated_objss="
   "
 
 lapack_deprecated_objsz="
+  zgelqs
+  zgeqrs
   zgegs
   zgegv
   zgelsx

kernel/mips64/cgemm_kernel_loongson3a_4x2_ps.S

@@ -131,11 +131,11 @@
 	sd	$21,  40($sp)
 	sd	$22,  48($sp)
 
-	ST	$f24, 56($sp)
-	ST	$f25, 64($sp)
-	ST	$f26, 72($sp)
-	ST	$f27, 80($sp)
-	ST	$f28, 88($sp)
+	sdc1	$f24, 56($sp)
+	sdc1	$f25, 64($sp)
+	sdc1	$f26, 72($sp)
+	sdc1	$f27, 80($sp)
+	sdc1	$f28, 88($sp)
 
 #if defined(TRMMKERNEL)
 	sd	$23,  96($sp)
@@ -146,10 +146,10 @@
 #endif
 
 #ifndef __64BIT__
-	ST	$f20,120($sp)
-	ST	$f21,128($sp)
-	ST	$f22,136($sp)
-	ST	$f23,144($sp)
+	sdc1	$f20,120($sp)
+	sdc1	$f21,128($sp)
+	sdc1	$f22,136($sp)
+	sdc1	$f23,144($sp)
 #endif
 
 	.align	4
@@ -4000,11 +4000,11 @@
 	ld	$21,  40($sp)
 	ld	$22,  48($sp)
 
-	LD	$f24, 56($sp)
-	LD	$f25, 64($sp)
-	LD	$f26, 72($sp)
-	LD	$f27, 80($sp)
-	LD	$f28, 88($sp)
+	ldc1	$f24, 56($sp)
+	ldc1	$f25, 64($sp)
+	ldc1	$f26, 72($sp)
+	ldc1	$f27, 80($sp)
+	ldc1	$f28, 88($sp)
 
 #if defined(TRMMKERNEL)
 	ld	$23,  96($sp)
@@ -4013,10 +4013,10 @@
 #endif
 
 #ifndef __64BIT__
-	LD	$f20,120($sp)
-	LD	$f21,128($sp)
-	LD	$f22,136($sp)
-	LD	$f23,144($sp)
+	ldc1	$f20,120($sp)
+	ldc1	$f21,128($sp)
+	ldc1	$f22,136($sp)
+	ldc1	$f23,144($sp)
 #endif
 
 	daddiu	$sp,$sp,STACKSIZE

test/compare_sgemm_sbgemm.c

@@ -29,6 +29,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "../common.h"
 #define SGEMM   BLASFUNC(sgemm)
 #define SBGEMM   BLASFUNC(sbgemm)
+#define SGEMV   BLASFUNC(sgemv)
+#define SBGEMV   BLASFUNC(sbgemv)
 typedef union
 {
   unsigned short v;
@@ -187,7 +189,79 @@ main (int argc, char *argv[])
     free(CC);
   }
 
-  if (ret != 0)
+  if (ret != 0) {
     fprintf (stderr, "FATAL ERROR SBGEMM - Return code: %d\n", ret);
+    return ret;
+  }
+
+  k = 1;
+  for (x = 1; x <= loop; x++)
+  {
+    float *A = (float *)malloc(x * x * sizeof(FLOAT));
+    float *B = (float *)malloc(x * sizeof(FLOAT));
+    float *C = (float *)malloc(x * sizeof(FLOAT));
+    bfloat16_bits *AA = (bfloat16_bits *)malloc(x * x * sizeof(bfloat16_bits));
+    bfloat16_bits *BB = (bfloat16_bits *)malloc(x * sizeof(bfloat16_bits));
+    float *DD = (float *)malloc(x * sizeof(FLOAT));
+    float *CC = (float *)malloc(x * sizeof(FLOAT));
+    if ((A == NULL) || (B == NULL) || (C == NULL) || (AA == NULL) || (BB == NULL) ||
+        (DD == NULL) || (CC == NULL))
+      return 1;
+    bfloat16 atmp, btmp;
+    blasint one = 1;
+
+    for (j = 0; j < x; j++)
+    {
+      for (i = 0; i < x; i++)
+      {
+        A[j * x + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+        sbstobf16_(&one, &A[j*x+i], &one, &atmp, &one);
+        AA[j * x + i].v = atmp;
+      }
+      B[j] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+      sbstobf16_(&one, &B[j], &one, &btmp, &one);
+      BB[j].v = btmp;
+    }
+    for (y = 0; y < 2; y++)
+    {
+      if (y == 0) {
+        transA = 'N';
+      } else {
+        transA = 'T';
+      }
+
+      memset(CC, 0, x * sizeof(FLOAT));
+      memset(DD, 0, x * sizeof(FLOAT));
+      memset(C, 0, x * sizeof(FLOAT));
+
+      SGEMV (&transA, &x, &x, &alpha, A, &x, B, &k, &beta, C, &k);
+      SBGEMV (&transA, &x, &x, &alpha, (bfloat16*) AA, &x, (bfloat16*) BB, &k, &beta, CC, &k);
+
+      for (j = 0; j < x; j++)
+        for (i = 0; i < x; i++)
+          if (transA == 'N') {
+            DD[i] += float16to32 (AA[j * x + i]) * float16to32 (BB[j]);
+          } else if (transA == 'T') {
+            DD[j] += float16to32 (AA[j * x + i]) * float16to32 (BB[i]);
+          }
+
+      for (j = 0; j < x; j++) {
+        if (fabs (CC[j] - C[j]) > 1.0)
+          ret++;
+        if (fabs (CC[j] - DD[j]) > 1.0)
+          ret++;
+      }
+    }
+    free(A);
+    free(B);
+    free(C);
+    free(AA);
+    free(BB);
+    free(DD);
+    free(CC);
+  }
+
+  if (ret != 0)
+    fprintf (stderr, "FATAL ERROR SBGEMV - Return code: %d\n", ret);
   return ret;
 }