SSE RSUM/RDSUM use correct headers for SSE2, SSSE3 and SSE4

- Replace python floor with // and remove math header PiperOrigin-RevId: 676263295
google · Sep 19, 2024 · 619b3fc · 619b3fc
1 parent 424d318
commit 619b3fc
Show file tree

Hide file tree

Showing 22 changed files with 23 additions and 35 deletions.
diff --git a/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c16.c b/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c16.c
@@ -6,11 +6,10 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <smmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c32.c b/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c32.c
@@ -6,11 +6,10 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <smmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c64.c b/src/qs8-rdsum/gen/qs8-rdsum-7p7x-minmax-fp32-sse41-c64.c
@@ -6,11 +6,10 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <smmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qs8-rdsum/sse41.c.in b/src/qs8-rdsum/sse41.c.in
@@ -2,12 +2,10 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
-$import math
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <smmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"
@@ -47,7 +45,7 @@ void xnn_qs8_rdsum_ukernel_${ACCUMULATORS}p${ACCUMULATORS}x__sse41_c${CHANNELS}(
     // 256 int8s may be summed into an int16 before overflowing
     // To prevent handling the tails of the inner 256 loop, we round 256 down to
     // the nearest integer multiple of ACCUMULATORS.
-    $OVERFLOW = math.floor(256 / ACCUMULATORS) * ACCUMULATORS
+    $OVERFLOW = (256 // ACCUMULATORS) * ACCUMULATORS
     int r = rows;
     while (r > 0) {
       $for C in range(0, CHANNELS, 8):

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u16.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u16.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u32-acc2.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u32-acc2.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u32.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u32.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u64-acc2.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u64-acc2.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u64-acc4.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u64-acc4.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/gen/qs8-rsum-ssse3-u64.c b/src/qs8-rsum/gen/qs8-rsum-ssse3-u64.c
@@ -6,10 +6,9 @@
 //
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qs8-rsum/ssse3.c.in b/src/qs8-rsum/ssse3.c.in
@@ -6,10 +6,9 @@ $assert CHANNEL_TILE % 16 == 0
 $assert CHANNEL_TILE >= 16
 $SIMD_TILE = CHANNEL_TILE // 16
 $assert ACCUMULATORS <= SIMD_TILE
-
 #include <assert.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c16.c b/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c16.c
@@ -10,7 +10,7 @@
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c32.c b/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c32.c
@@ -10,7 +10,7 @@
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c64.c b/src/qu8-rdsum/gen/qu8-rdsum-7p7x-ssse3-c64.c
@@ -10,7 +10,7 @@
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qu8-rdsum/ssse3.c.in b/src/qu8-rdsum/ssse3.c.in
@@ -7,7 +7,7 @@ $import math
 #include <assert.h>
 #include <math.h>
 
-#include <immintrin.h>
+#include <tmmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/intrinsics-polyfill.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u16.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u16.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u32-acc2.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u32-acc2.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u32.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u32.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u64-acc2.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u64-acc2.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u64-acc4.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u64-acc4.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/gen/qu8-rsum-sse2-u64.c b/src/qu8-rsum/gen/qu8-rsum-sse2-u64.c
@@ -8,7 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"

diff --git a/src/qu8-rsum/sse2.c.in b/src/qu8-rsum/sse2.c.in
@@ -8,7 +8,7 @@ $SIMD_TILE = CHANNEL_TILE // 16
 $assert ACCUMULATORS <= SIMD_TILE
 #include <assert.h>
 
-#include <immintrin.h>
+#include <emmintrin.h>
 
 #include "xnnpack/common.h"
 #include "xnnpack/reduce.h"