[dev.simd] simd, cmd/compile: add float -> float conversions

This should mark the end of the conversion table, except for float16
which does not exist on Go yet.

The rounding logic documentation of float64 -> float32 is based on
abi-internal default MXCSR:
| RC | 14/13 | 0 (RN) | Round to nearest |

Change-Id: I27a86560e8d74d20f21350bf78314b4eada20ec0
Reviewed-on: https://go-review.googlesource.com/c/go/+/724440
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: David Chase <drchase@google.com>

4 files changed, 55 insertions, 3 deletions

diff --git a/src/simd/_gen/simdgen/gen_utility.go b/src/simd/_gen/simdgen/gen_utility.go
index cf91f4a65e..424ba41466 100644
--- a/src/simd/_gen/simdgen/gen_utility.go
+++ b/src/simd/_gen/simdgen/gen_utility.go
@@ -337,7 +337,9 @@ func (op *Operation) sortOperand() {
 
 // adjustAsm adjusts the asm to make it align with Go's assembler.
 func (op *Operation) adjustAsm() {
-	if op.Asm == "VCVTTPD2DQ" || op.Asm == "VCVTTPD2UDQ" || op.Asm == "VCVTQQ2PS" || op.Asm == "VCVTUQQ2PS" {
+	if op.Asm == "VCVTTPD2DQ" || op.Asm == "VCVTTPD2UDQ" ||
+		op.Asm == "VCVTQQ2PS" || op.Asm == "VCVTUQQ2PS" ||
+		op.Asm == "VCVTPD2PS" {
 		switch *op.In[0].Bits {
 		case 128:
 			op.Asm += "X"
diff --git a/src/simd/_gen/simdgen/ops/Converts/categories.yaml b/src/simd/_gen/simdgen/ops/Converts/categories.yaml
index 61ab195765..dd33284063 100644
--- a/src/simd/_gen/simdgen/ops/Converts/categories.yaml
+++ b/src/simd/_gen/simdgen/ops/Converts/categories.yaml
@@ -32,12 +32,12 @@
     // When a conversion is inexact, a truncated (round toward zero) value is returned.
     // If a converted result cannot be represented in uint64, an implementation-defined
     // architecture-specific value is returned.
-- go: "ConvertToFloat32"
+- go: "ConvertToFloat32" # Also float64 -> float32
   commutative: false
   regexpTag: "convert"
   documentation: !string |-
     // NAME converts element values to float32.
-- go: "ConvertToFloat64"
+- go: "ConvertToFloat64" # Also float32 -> float64
   commutative: false
   regexpTag: "convert"
   documentation: !string |-
diff --git a/src/simd/_gen/simdgen/ops/Converts/go.yaml b/src/simd/_gen/simdgen/ops/Converts/go.yaml
index ec28b57681..af058124fb 100644
--- a/src/simd/_gen/simdgen/ops/Converts/go.yaml
+++ b/src/simd/_gen/simdgen/ops/Converts/go.yaml
@@ -112,6 +112,28 @@
   out:
   - base: float
     bits: 256|512
+# float64 -> float32
+- go: ConvertToFloat32
+  regexpTag: "convert"
+  asm: "VCVTPD2PS"
+  addDoc:
+    !string |-
+    // The result vector's elements are rounded to the nearest value.
+  in: &fp64
+  - base: float
+    elemBits: 64
+  out: &fp32
+  - base: float
+    elemBits: 32
+# float32 -> float64
+- go: ConvertToFloat64
+  regexpTag: "convert"
+  asm: "VCVTPS2PD"
+  in: *fp32
+  out:
+  - base: float
+    elemBits: 64
+    bits: 256|512
 
 # Widening integer conversions.
 # uint8 -> uint16
diff --git a/src/simd/ops_amd64.go b/src/simd/ops_amd64.go
index 5722fb2543..7280e873a0 100644
--- a/src/simd/ops_amd64.go
+++ b/src/simd/ops_amd64.go
@@ -1547,6 +1547,24 @@ func (x Uint8x64) ConcatShiftBytesRightGrouped(constant uint8, y Uint8x64) Uint8
 /* ConvertToFloat32 */
 
 // ConvertToFloat32 converts element values to float32.
+// The result vector's elements are rounded to the nearest value.
+//
+// Asm: VCVTPD2PSX, CPU Feature: AVX
+func (x Float64x2) ConvertToFloat32() Float32x4
+
+// ConvertToFloat32 converts element values to float32.
+// The result vector's elements are rounded to the nearest value.
+//
+// Asm: VCVTPD2PSY, CPU Feature: AVX
+func (x Float64x4) ConvertToFloat32() Float32x4
+
+// ConvertToFloat32 converts element values to float32.
+// The result vector's elements are rounded to the nearest value.
+//
+// Asm: VCVTPD2PS, CPU Feature: AVX512
+func (x Float64x8) ConvertToFloat32() Float32x8
+
+// ConvertToFloat32 converts element values to float32.
 //
 // Asm: VCVTDQ2PS, CPU Feature: AVX
 func (x Int32x4) ConvertToFloat32() Float32x4
@@ -1610,6 +1628,16 @@ func (x Uint64x8) ConvertToFloat32() Float32x8
 
 // ConvertToFloat64 converts element values to float64.
 //
+// Asm: VCVTPS2PD, CPU Feature: AVX
+func (x Float32x4) ConvertToFloat64() Float64x4
+
+// ConvertToFloat64 converts element values to float64.
+//
+// Asm: VCVTPS2PD, CPU Feature: AVX512
+func (x Float32x8) ConvertToFloat64() Float64x8
+
+// ConvertToFloat64 converts element values to float64.
+//
 // Asm: VCVTDQ2PD, CPU Feature: AVX
 func (x Int32x4) ConvertToFloat64() Float64x4