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

feat(arch): add the packed FP16 arithmetic to the amd64 extension layer

Browse Source 
Assisted-by: GLM 5.3 Flash
This commit is contained in:
petrbalvin committed 2026-10-07 00:07:58 +02:00
1 parent a2301b52de
commit 22055b9bf3
3 files changed
+55 -4
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

arch/amd64_ext.go
+25
View File
@@ -406,4 +406,29 @@ var amd64Extensions = []ExtInstr{
{Name: "VCVTSH2USI", Summary: "Convert a low FP16 value to an unsigned 64-bit integer",
Bytes: []byte{0x62, 0x05, 0x86, 0x00, 0x79, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VCVTSH2USI (EVEX.LIG.F3.MAP5.W1 79 /r)"},
// AVX512-FP16 packed, the 512-bit arithmetic the scalar core mirrors:
// full ZMM lanes, EVEX.NDS.MAP5 with no mandatory prefix, rounding
// control left to MXCSR.
{Name: "VADDPH", Summary: "Add packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x58, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.512.MAP5.W0 58 /r)"},
{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.512.MAP5.W0 5C /r)"},
{Name: "VMULPH", Summary: "Multiply packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x59, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.512.MAP5.W0 59 /r)"},
{Name: "VDIVPH", Summary: "Divide packed FP16 values",
Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.512.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,
Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.512.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,
Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.512.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,
Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.512.MAP5.W0 51 /r)"},
}