commit e14e791244e1ffae0285b442af7cf12a9db1321a
author Marat Dukhan <maratek@google.com> Thu Dec 30 07:10:36 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Dec 30 07:11:27 2021 -0800
tree ef446fedda97828404506d846dbb971140a693d9
parent 134f9846ebc3294333527849d06aef0487cf7156

Move mask_table into VHSWISH AVX microkernel parameters

Unblock amalgamation for F32 VHSWISH AVX microkernels

PiperOrigin-RevId: 418954150

src/f32-vhswish/avx.c.in

diff --git a/src/f32-vhswish/avx.c.in b/src/f32-vhswish/avx.c.in
index 72a48ff..2848d35 100644
--- a/src/f32-vhswish/avx.c.in
+++ b/src/f32-vhswish/avx.c.in
@@ -14,8 +14,6 @@
 #include <xnnpack/vunary.h>
 
 
-static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
-
 $ISA = {0: "avx", 3: "fma3"}[FMA]
 void xnn_f32_vhswish_ukernel__${ISA}_x${BATCH_TILE}(
     size_t n,
@@ -79,7 +77,7 @@
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
-    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
+    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
 
     const __m256 vx = _mm256_maskload_ps(x, vmask);
     $if FMA == 3:

src/f32-vhswish/gen/vhswish-avx-x16.c

diff --git a/src/f32-vhswish/gen/vhswish-avx-x16.c b/src/f32-vhswish/gen/vhswish-avx-x16.c
index e96a321..83f712e 100644
--- a/src/f32-vhswish/gen/vhswish-avx-x16.c
+++ b/src/f32-vhswish/gen/vhswish-avx-x16.c
@@ -15,8 +15,6 @@
 #include <xnnpack/vunary.h>
 
 
-static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
-
 void xnn_f32_vhswish_ukernel__avx_x16(
     size_t n,
     const float* x,
@@ -69,7 +67,7 @@
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
-    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
+    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
 
     const __m256 vx = _mm256_maskload_ps(x, vmask);
     __m256 vacc = _mm256_mul_ps(vx, vsixth);

src/f32-vhswish/gen/vhswish-avx-x8.c

diff --git a/src/f32-vhswish/gen/vhswish-avx-x8.c b/src/f32-vhswish/gen/vhswish-avx-x8.c
index c6e7ebc..1f4945f 100644
--- a/src/f32-vhswish/gen/vhswish-avx-x8.c
+++ b/src/f32-vhswish/gen/vhswish-avx-x8.c
@@ -15,8 +15,6 @@
 #include <xnnpack/vunary.h>
 
 
-static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
-
 void xnn_f32_vhswish_ukernel__avx_x8(
     size_t n,
     const float* x,
@@ -45,7 +43,7 @@
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
-    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
+    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
 
     const __m256 vx = _mm256_maskload_ps(x, vmask);
     __m256 vacc = _mm256_mul_ps(vx, vsixth);

src/f32-vhswish/gen/vhswish-fma3-x16.c

diff --git a/src/f32-vhswish/gen/vhswish-fma3-x16.c b/src/f32-vhswish/gen/vhswish-fma3-x16.c
index 92f7de1..223905c 100644
--- a/src/f32-vhswish/gen/vhswish-fma3-x16.c
+++ b/src/f32-vhswish/gen/vhswish-fma3-x16.c
@@ -15,8 +15,6 @@
 #include <xnnpack/vunary.h>
 
 
-static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
-
 void xnn_f32_vhswish_ukernel__fma3_x16(
     size_t n,
     const float* x,
@@ -65,7 +63,7 @@
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
-    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
+    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
 
     const __m256 vx = _mm256_maskload_ps(x, vmask);
     __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);

src/f32-vhswish/gen/vhswish-fma3-x8.c

diff --git a/src/f32-vhswish/gen/vhswish-fma3-x8.c b/src/f32-vhswish/gen/vhswish-fma3-x8.c
index 22baf10..a45f636 100644
--- a/src/f32-vhswish/gen/vhswish-fma3-x8.c
+++ b/src/f32-vhswish/gen/vhswish-fma3-x8.c
@@ -15,8 +15,6 @@
 #include <xnnpack/vunary.h>
 
 
-static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
-
 void xnn_f32_vhswish_ukernel__fma3_x8(
     size_t n,
     const float* x,
@@ -44,7 +42,7 @@
   if XNN_UNLIKELY(n != 0) {
     assert(n >= 1 * sizeof(float));
     assert(n <= 7 * sizeof(float));
-    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
+    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
 
     const __m256 vx = _mm256_maskload_ps(x, vmask);
     __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);

src/params-init.c

diff --git a/src/params-init.c b/src/params-init.c
index 472f0b3..459d1b6 100644
--- a/src/params-init.c
+++ b/src/params-init.c
@@ -1223,6 +1223,12 @@
     params->avx.half[i] = 0.5f;
     params->avx.one[i] = 1.0f;
   }
+  for (uint32_t i = 0; i < 7; i++) {
+    params->avx.mask_table[i] = -1;
+  }
+  for (uint32_t i = 7; i < 14; i++) {
+    params->avx.mask_table[i] = 0;
+  }
 }
 
 void xnn_init_f32_hswish_avx512_params(

src/xnnpack/params.h

diff --git a/src/xnnpack/params.h b/src/xnnpack/params.h
index a976c78..824b2f4 100644
--- a/src/xnnpack/params.h
+++ b/src/xnnpack/params.h
@@ -284,6 +284,7 @@
     XNN_ALIGN(32) float sixth[8];
     XNN_ALIGN(32) float half[8];
     XNN_ALIGN(32) float one[8];
+    int32_t mask_table[14];
   } avx;
   struct {
     float sixth;