petrbalvin/gasm-sdk
0
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 37 Wiki Activity

feat(arch): add the packed FP16 and BF16 memory forms to the extension layer

Browse Source 
Assisted-by: GLM 5.3 Flash
This commit is contained in:
petrbalvin committed 2026-10-07 01:35:48 +02:00
1 parent d275dee3ae
commit 454a21f5b7
2 files changed
+137 -30
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

arch/amd64_ext.go
+53 -30
View File
@@ -316,13 +316,36 @@ func (in ExtInstr) encodeAmdVecMem(ops []ExtOperand) ([]byte, error) {
// encodeAmdVec2 fills the two-vector form: src, dest. The half form narrows
// the destination: VCVTNEPS2BF16 converts 512 bits of source into 256 bits
// of destination, and at 128 bits the companion stays the class itself.
// of destination, and at 128 bits the companion stays the class itself. An
// entry with Mem set takes the memory shape of that position too: the
// compares and the packed square root read their source from memory, and
// the narrow BF16 convert reads its full-width source there.
func (in ExtInstr) encodeAmdVec2(ops []ExtOperand) ([]byte, error) {
class := amd64LengthClass(in.Bytes)
destClass := class
if in.Form == ExtFormAmdVec2Half {
destClass = amd64HalfClass(class)
}
if in.Mem == 1 && ops[0].Kind == ExtMem {
base, disp, err := in.amd64Memory(ops[0], 1)
if err != nil {
return nil, err
}
if err := in.amd64Vector(ops[1], destClass, 2); err != nil {
return nil, err
}
return amd64EncodeMemory(in.Bytes, ops[1].Reg, -1, base, disp), nil
}
if in.Mem == 2 && ops[1].Kind == ExtMem {
if err := in.amd64Vector(ops[0], class, 1); err != nil {
return nil, err
}
base, disp, err := in.amd64Memory(ops[1], 2)
if err != nil {
return nil, err
}
return amd64EncodeMemory(in.Bytes, ops[0].Reg, -1, base, disp), nil
}
if err := in.amd64Vector(ops[0], class, 1); err != nil {
return nil, err
}
@@ -509,22 +532,22 @@ var amd64Extensions = []ExtInstr{
Bytes: []byte{0x62, 0x02, 0x07, 0x00, 0x72, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VCVTNE2PS2BF16 (EVEX.NDS.128.F2.0F38.W0 72 /r)"},
{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.512.F3.0F38.W0 72 /r, YMM destination)"},
{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.256.F3.0F38.W0 72 /r, XMM destination)"},
{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.128.F3.0F38.W0 72 /r, XMM destination)"},
{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x52, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.512.F3.0F38.W0 52 /r)"},
{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x52, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.256.F3.0F38.W0 52 /r)"},
{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x52, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.128.F3.0F38.W0 52 /r)"},
// AVX512-VP2INTERSECT: the pairwise intersection indices, one opmask
@@ -605,10 +628,10 @@ var amd64Extensions = []ExtInstr{
Bytes: []byte{0x62, 0x06, 0x05, 0x00, 0x43, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VGETEXPSH (EVEX.NDS.LIG.66.MAP6.W0 43 /r)"},
{Name: "VCOMISH", Summary: "Compare a scalar FP16 value and set EFLAGS",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2F, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2F, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VCOMISH (EVEX.LIG.MAP5.W0 2F /r)"},
{Name: "VUCOMISH", Summary: "Unordered-compare a scalar FP16 value and set EFLAGS",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2E, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2E, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VUCOMISH (EVEX.LIG.MAP5.W0 2E /r)"},
{Name: "VCVTSS2SH", Summary: "Convert one FP32 value to one FP16 value",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x1D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
@@ -653,67 +676,67 @@ var amd64Extensions = []ExtInstr{
// quoted from x86-64-avx512_fp16.d, the 256- and 128-bit ones from
// avx512_fp16_vl.d, on the same low registers the suite uses.
{Name: "VADDPH", Summary: "Add packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x58, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.512.MAP5.W0 58 /r)"},
{Name: "VADDPH", Summary: "Add packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x58, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.256.MAP5.W0 58 /r)"},
{Name: "VADDPH", Summary: "Add packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x58, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.128.MAP5.W0 58 /r)"},
{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.512.MAP5.W0 5C /r)"},
{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.256.MAP5.W0 5C /r)"},
{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.128.MAP5.W0 5C /r)"},
{Name: "VMULPH", Summary: "Multiply packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x59, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.512.MAP5.W0 59 /r)"},
{Name: "VMULPH", Summary: "Multiply packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x59, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.256.MAP5.W0 59 /r)"},
{Name: "VMULPH", Summary: "Multiply packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x59, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.128.MAP5.W0 59 /r)"},
{Name: "VDIVPH", Summary: "Divide packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.512.MAP5.W0 5E /r)"},
{Name: "VDIVPH", Summary: "Divide packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.256.MAP5.W0 5E /r)"},
{Name: "VDIVPH", Summary: "Divide packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.128.MAP5.W0 5E /r)"},
{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.512.MAP5.W0 5D /r)"},
{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.256.MAP5.W0 5D /r)"},
{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.128.MAP5.W0 5D /r)"},
{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.512.MAP5.W0 5F /r)"},
{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.256.MAP5.W0 5F /r)"},
{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.128.MAP5.W0 5F /r)"},
{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x51, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.512.MAP5.W0 51 /r)"},
{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x51, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.256.MAP5.W0 51 /r)"},
{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x51, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.128.MAP5.W0 51 /r)"},
// AVX512-FP16 scalar, the imm8-control group: mantissa extraction,