[Fix] : invalid argument type 'simde_float16' to unary expression

simd-everywhere · Oct 16, 2023 · 0046c2e · 0046c2e
1 parent dcea734
commit 0046c2e
Show file tree

Hide file tree

Showing 5 changed files with 72 additions and 34 deletions.
diff --git a/simde/arm/neon/cadd_rot270.h b/simde/arm/neon/cadd_rot270.h
@@ -48,8 +48,9 @@ simde_float16x4_t simde_vcadd_rot270_f16(simde_float16x4_t a,
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(b_.values[2 * i + 1], a_.values[2 * i]);
-    r_.values[2 * i + 1] =
-        simde_vaddh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]);
+    r_.values[2 * i + 1] = simde_vaddh_f16(
+        simde_float16_from_float32(-simde_float16_to_float32(b_.values[2 * i])),
+        a_.values[2 * i + 1]);
   }
 
   return simde_float16x4_from_private(r_);
@@ -75,8 +76,9 @@ simde_float16x8_t simde_vcaddq_rot270_f16(simde_float16x8_t a,
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(b_.values[2 * i + 1], a_.values[2 * i]);
-    r_.values[2 * i + 1] =
-        simde_vaddh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]);
+    r_.values[2 * i + 1] = simde_vaddh_f16(
+        simde_float16_from_float32(-simde_float16_to_float32(b_.values[2 * i])),
+        a_.values[2 * i + 1]);
   }
 
   return simde_float16x8_from_private(r_);

diff --git a/simde/arm/neon/cadd_rot90.h b/simde/arm/neon/cadd_rot90.h
@@ -48,7 +48,9 @@ simde_float16x4_t simde_vcadd_rot90_f16(simde_float16x4_t a,
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
     r_.values[2 * i] =
-        simde_vaddh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i]);
+        simde_vaddh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]);
     r_.values[2 * i + 1] =
         simde_vaddh_f16(b_.values[2 * i], a_.values[2 * i + 1]);
   }
@@ -74,7 +76,9 @@ simde_float16x8_t simde_vcaddq_rot90_f16(simde_float16x8_t a,
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
     r_.values[2 * i] =
-        simde_vaddh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i]);
+        simde_vaddh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]);
     r_.values[2 * i + 1] =
         simde_vaddh_f16(b_.values[2 * i], a_.values[2 * i + 1]);
   }

diff --git a/simde/arm/neon/cmla_rot180_lane.h b/simde/arm/neon/cmla_rot180_lane.h
@@ -57,12 +57,16 @@ simde_float16x4_t simde_vcmla_rot180_lane_f16(simde_float16x4_t r,
   SIMDE_VECTORIZE
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
-    r_.values[2 * i] =
-        simde_vaddh_f16(r_.values[2 * i],
-                        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i]));
-    r_.values[2 * i + 1] += = simde_vaddh_f16(
+    r_.values[2 * i] = simde_vaddh_f16(
+        r_.values[2 * i],
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i]));
+    r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]) * a_.values[2 * i]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]));
   }
 
   result = simde_float16x4_from_private(r_);
@@ -134,12 +138,16 @@ simde_float16x8_t simde_vcmlaq_rot180_lane_f16(simde_float16x8_t r,
   SIMDE_VECTORIZE
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
-    r_.values[2 * i] =
-        simde_vaddh_f16(r_.values[2 * i],
-                        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i]));
-    r_.values[2 * i + 1] += = simde_vaddh_f16(
+    r_.values[2 * i] = simde_vaddh_f16(
+        r_.values[2 * i],
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i]));
+    r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]) * a_.values[2 * i]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]));
   }
 
   result = simde_float16x8_from_private(r_);
@@ -211,12 +219,16 @@ simde_float16x4_t simde_vcmla_rot180_laneq_f16(simde_float16x4_t r,
   SIMDE_VECTORIZE
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
-    r_.values[2 * i] =
-        simde_vaddh_f16(r_.values[2 * i],
-                        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i]));
-    r_.values[2 * i + 1] += = simde_vaddh_f16(
+    r_.values[2 * i] = simde_vaddh_f16(
+        r_.values[2 * i],
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i]));
+    r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]) * a_.values[2 * i]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]));
   }
 
   result = simde_float16x4_from_private(r_);
@@ -291,12 +303,16 @@ simde_float16x8_t simde_vcmlaq_rot180_laneq_f16(simde_float16x8_t r,
   SIMDE_VECTORIZE
   for (size_t i = 0; i < (sizeof(r_.values) / (2 * sizeof(r_.values[0])));
        i++) {
-    r_.values[2 * i] =
-        simde_vaddh_f16(r_.values[2 * i],
-                        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i]));
-    r_.values[2 * i + 1] += = simde_vaddh_f16(
+    r_.values[2 * i] = simde_vaddh_f16(
+        r_.values[2 * i],
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i]));
+    r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]) * a_.values[2 * i]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i]));
   }
 
   result = simde_float16x8_from_private(r_);

diff --git a/simde/arm/neon/cmla_rot270_lane.h b/simde/arm/neon/cmla_rot270_lane.h
@@ -61,7 +61,9 @@ simde_float16x4_t simde_vcmla_rot270_lane_f16(simde_float16x4_t r,
         simde_vmulh_f16(b_.values[2 * i + 1], a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i + 1]));
   }
 
   result = simde_float16x4_from_private(r_);
@@ -138,7 +140,9 @@ simde_float16x8_t simde_vcmlaq_rot270_lane_f16(simde_float16x8_t r,
         simde_vmulh_f16(b_.values[2 * i + 1], a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i + 1]));
   }
 
   result = simde_float16x8_from_private(r_);
@@ -215,7 +219,9 @@ simde_float16x4_t simde_vcmla_rot270_laneq_f16(simde_float16x4_t r,
         simde_vmulh_f16(b_.values[2 * i + 1], a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i + 1]));
   }
 
   result = simde_float16x4_from_private(r_);
@@ -295,7 +301,9 @@ simde_float16x8_t simde_vcmlaq_rot270_laneq_f16(simde_float16x8_t r,
         simde_vmulh_f16(b_.values[2 * i + 1], a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
-        simde_vmulh_f16(-(b_.values[2 * i]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i])),
+                        a_.values[2 * i + 1]));
   }
 
   result = simde_float16x8_from_private(r_);

diff --git a/simde/arm/neon/cmla_rot90_lane.h b/simde/arm/neon/cmla_rot90_lane.h
@@ -58,7 +58,9 @@ simde_float16x4_t simde_vcmla_rot90_lane_f16(simde_float16x4_t r,
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(
         r_.values[2 * i],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
         simde_vmulh_f16(b_.values[2 * i], a_.values[2 * i + 1]));
@@ -134,7 +136,9 @@ simde_float16x8_t simde_vcmlaq_rot90_lane_f16(simde_float16x8_t r,
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(
         r_.values[2 * i],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
         simde_vmulh_f16(b_.values[2 * i], a_.values[2 * i + 1]));
@@ -211,7 +215,9 @@ simde_float16x4_t simde_vcmla_rot90_laneq_f16(simde_float16x4_t r,
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(
         r_.values[2 * i],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
         simde_vmulh_f16(b_.values[2 * i], a_.values[2 * i + 1]));
@@ -291,7 +297,9 @@ simde_float16x8_t simde_vcmlaq_rot90_laneq_f16(simde_float16x8_t r,
        i++) {
     r_.values[2 * i] = simde_vaddh_f16(
         r_.values[2 * i],
-        simde_vmulh_f16(-(b_.values[2 * i + 1]), a_.values[2 * i + 1]));
+        simde_vmulh_f16(simde_float16_from_float32(
+                            -simde_float16_to_float32(b_.values[2 * i + 1])),
+                        a_.values[2 * i + 1]));
     r_.values[2 * i + 1] = simde_vaddh_f16(
         r_.values[2 * i + 1],
         simde_vmulh_f16(b_.values[2 * i], a_.values[2 * i + 1]));