Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Update mono to support shuffle for constant inputs #105299

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 15 commits into from
Jul 25, 2024
Merged

Update mono to support shuffle for constant inputs #105299

Changes from 1 commit
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
4820a21
Support mono creating xconst in a few more places
tannergooding Jul 23, 2024
fcdbd18
Update mono to support shuffle for constant inputs
tannergooding Jul 23, 2024
095873b
Ensure that arm64 also accelerates shuffle for non-constant inputs
tannergooding Jul 23, 2024
d87a444
Ensure OP_XZERO and OP_XONES are recognized as being constant
tannergooding Jul 23, 2024
71708fe
Ensure shuffle creates a correct instruction when the fsig doesn't ma…
tannergooding Jul 23, 2024
f85e513
Ensure that getting the index for floating-point shuffle is possible
tannergooding Jul 23, 2024
2995754
Ensure the right class handle is passed down to LLVM so overload reso…
tannergooding Jul 23, 2024
b35b1e3
Make sure we update the original xconst if we mutate it
tannergooding Jul 23, 2024
1d55132
Return a new constant and instead of mutating the existing one
tannergooding Jul 24, 2024
3c327b0
Insert relevant xcast nodes
tannergooding Jul 24, 2024
50cec20
Add some asserts around the ecount
tannergooding Jul 24, 2024
fe451d7
Ensure we get the right element type
tannergooding Jul 24, 2024
08aa477
Ensure we don't create nodes unnecessarily for create_elementwise
tannergooding Jul 24, 2024
4a49172
Ensure that create_elementwise still works for other vector sizes
tannergooding Jul 24, 2024
5a8a221
Ensure indentation of switch cases is correct for Mono
tannergooding Jul 25, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Ensure that arm64 also accelerates shuffle for non-constant inputs
  • Loading branch information
@tannergooding
tannergooding committed Jul 23, 2024
commit 095873b426c9fb36829c7e121edb35062bf28f84
44 changes: 39 additions & 5 deletions src/mono/mono/mini/simd-intrinsics.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2659,9 +2659,43 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
#ifdef TARGET_WASM
return emit_simd_ins_for_sig (cfg, klass, OP_WASM_SIMD_SWIZZLE, -1, -1, fsig, args);
#elif defined(TARGET_ARM64)
if (vector_size == 128 && (arg0_type == MONO_TYPE_I1 || arg0_type == MONO_TYPE_U1))
if (vector_size != 128) {
return NULL;
}
if ((arg0_type == MONO_TYPE_I1 || arg0_type == MONO_TYPE_U1)) {
return emit_simd_ins_for_sig (cfg, klass, OP_XOP_OVR_X_X_X, INTRINS_AARCH64_ADV_SIMD_TBL1, 0, fsig, args);
return NULL;
}
if (args [1]->opcode != OP_XCONST) {
return NULL;
}
MonoInst *new_args[2];
new_args [0] = args [0];
if (COMPILE_LLVM (cfg)) {
if ((get_xconst_int_elem (cfg, args [1], MONO_TYPE_U8, 0) == 0x300000002) &&
(get_xconst_int_elem (cfg, args [1], MONO_TYPE_U8, 1) == 0x100000000)) {
new_args [1] = args [0];
return emit_simd_ins_for_sig (cfg, klass, OP_ARM64_EXT, 0, MONO_TYPE_U8, fsig, new_args);
}
}
int esize = mono_class_value_size (mono_class_from_mono_type_internal (etype), NULL);
int ecount = (vector_size / 8) / esize;
guint64 value = 0;
guint8 vec_cns[16];
for (int index = 0; index < ecount; index++) {
value = get_xconst_int_elem (cfg, args [1], arg0_type, index);

if (value < ecount) {
for (int i = 0; i < esize; i++) {
vec_cns [(index * esize) + i] = (guint8)((value * esize) + i);
}
} else {
for (int i = 0; i < esize; i++) {
vec_cns [(index * esize) + i] = 0xFF;
}
}
}
new_args [1] = emit_xconst_v128 (cfg, klass, vec_cns);
return emit_simd_ins_for_sig (cfg, klass, OP_XOP_OVR_X_X_X, INTRINS_AARCH64_ADV_SIMD_TBL1, 0, fsig, new_args);
#elif defined(TARGET_AMD64)
if (COMPILE_LLVM (cfg)) {
if (vector_size != 128) {
Expand All @@ -2682,7 +2716,7 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
needs_zero = true;
}
for (int index = 0; index < ecount; index++) {
value = get_xconst_int_elem (cfg, args [1], etype->type, index);
value = get_xconst_int_elem (cfg, args [1], arg0_type, index);
if (value < ecount) {
// Setting the control for byte/sbyte and short/ushort is unnecessary
// and will actually compute an incorrect control word. But it simplifies
Expand All @@ -2696,7 +2730,7 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
// sequential bytes.

for (int i = 0; i < esize; i++) {
vec_cns[(index * esize) + i] = (guint8)((value * esize) + i);
vec_cns [(index * esize) + i] = (guint8)((value * esize) + i);
}
} else {
needs_zero = true;
Expand All @@ -2707,7 +2741,7 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
// set, but we use 0xFF instead since that will be the equivalent of AllBitsSet

for (int i = 0; i < esize; i++) {
vec_cns[(index * esize) + i] = 0xFF;
vec_cns [(index * esize) + i] = 0xFF;
}
}
}
Expand Down