Skip to content
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

[mono] Basic SIMD support for System.Numerics.Vector2 on arm64 #91659

Merged
merged 17 commits into from
Sep 21, 2023
Merged

[mono] Basic SIMD support for System.Numerics.Vector2 on arm64 #91659

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
17 commits
Select commit Hold shift + click to select a range
a1ead5f
enable Vector2 on arm64
matouskozak Aug 30, 2023
429bd7f
Vector2 fabs arm64 instrinsics
matouskozak Aug 30, 2023
cfee109
Vector2 fmin and fmax arm64 intrinsics
matouskozak Aug 30, 2023
13e08d0
Vector 2 fdiv arm64 intrinsics
matouskozak Aug 31, 2023
f223d29
Vector2 Dot product intrinsics arm64
matouskozak Aug 31, 2023
23c0003
Vector2 sub and sqrt intrinsics support for arm64
matouskozak Aug 31, 2023
c12429a
Vector2 op_equality fix or arm64
matouskozak Aug 31, 2023
e29c5ef
Support
matouskozak Sep 1, 2023
9cab6af
OP_XZERO 64-bit vector support arm64
matouskozak Sep 1, 2023
27d244b
OP_EXPAND/EXTRACT_R4/R8 for 64 bit vector arm64
matouskozak Sep 1, 2023
9493095
Revert "OP_EXPAND/EXTRACT_R4/R8 for 64 bit vector arm64"
matouskozak Sep 1, 2023
b191c22
fix get/set_Item for Vector2
matouskozak Sep 5, 2023
0691b2b
remove Vector2 SIMD support from amd64 and wasm if
matouskozak Sep 5, 2023
fe66c97
faddv 64-bit, extract VREG_FULL only
matouskozak Sep 5, 2023
cf43f87
Merge branch 'main' into arm64-vector2-intrinsics
matouskozak Sep 6, 2023
763eb0c
fix incorrect 64bit support for 64bit elements
matouskozak Sep 6, 2023
a2873b4
rename methods table
matouskozak Sep 6, 2023
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
23 changes: 16 additions & 7 deletions src/mono/mono/mini/mini-arm64.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3917,14 +3917,17 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
break;

case OP_XZERO:
arm_neon_eor_16b (code, dreg, dreg, dreg);
if (ins->klass && mono_class_value_size (ins->klass, NULL) == 8)
arm_neon_eor_8b (code, dreg, dreg, dreg);
else
arm_neon_eor_16b (code, dreg, dreg, dreg);
break;
case OP_XONES:
arm_neon_eor_16b (code, dreg, dreg, dreg);
arm_neon_not_16b (code, dreg, dreg);
break;
case OP_XEXTRACT:
code = emit_xextract (code, VREG_FULL, GTMREG_TO_INT (ins->inst_c0), dreg, sreg1);
code = emit_xextract (code, (ins->inst_c1 == 8) ? VREG_LOW : VREG_FULL, GTMREG_TO_INT (ins->inst_c0), dreg, sreg1);
break;
case OP_STOREX_MEMBASE:
if (ins->klass && mono_class_value_size (ins->klass, NULL) == 8)
Expand Down Expand Up @@ -3966,7 +3969,10 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
case OP_EXPAND_R4:
case OP_EXPAND_R8: {
const int t = get_type_size_macro (ins->inst_c1);
arm_neon_fdup_e (code, VREG_FULL, t, dreg, sreg1, 0);
if (ins->opcode == OP_EXPAND_R8)
arm_neon_fdup_e (code, VREG_FULL, t, dreg, sreg1, 0);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OP_EXPAND_R8 can be simplified to a mov dreg, sreg1 or nothing if dreg == sreg1.

else
arm_neon_fdup_e (code, get_vector_size_macro (ins), t, dreg, sreg1, 0);
break;
}
case OP_EXTRACT_I1:
Expand All @@ -3989,6 +3995,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
// Technically, this broadcasts element #inst_c0 to all dest XREG elements; whereas it should
// set the FREG to the said element. Since FREG and XREG pool is the same on arm64 and the rest
// of the F/XREG is ignored in FREG mode, this operation remains valid.
// FIXME: pass VREG_LOW for 64-bit vectors
arm_neon_fdup_e (code, VREG_FULL, t, dreg, sreg1, GTMREG_TO_UINT32 (ins->inst_c0));
}
break;
Expand Down Expand Up @@ -4083,17 +4090,19 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)

case OP_ARM64_XADDV: {
switch (ins->inst_c0) {
case INTRINS_AARCH64_ADV_SIMD_FADDV:
case INTRINS_AARCH64_ADV_SIMD_FADDV: {
const int width = get_vector_size_macro (ins);
if (ins->inst_c1 == MONO_TYPE_R8) {
arm_neon_faddp (code, VREG_FULL, TYPE_F64, dreg, sreg1, sreg1);
} else if (ins->inst_c1 == MONO_TYPE_R4) {
arm_neon_faddp (code, VREG_FULL, TYPE_F32, dreg, sreg1, sreg1);
arm_neon_faddp (code, VREG_FULL, TYPE_F32, dreg, dreg, dreg);
arm_neon_faddp (code, width, TYPE_F32, dreg, sreg1, sreg1);
if (width == VREG_FULL)
arm_neon_faddp (code, width, TYPE_F32, dreg, dreg, dreg);
} else {
g_assert_not_reached ();
}
break;

}
case INTRINS_AARCH64_ADV_SIMD_UADDV:
case INTRINS_AARCH64_ADV_SIMD_SADDV:
if (get_type_size_macro (ins->inst_c1) == TYPE_I64)
Expand Down
10 changes: 9 additions & 1 deletion src/mono/mono/mini/mini-runtime.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4469,7 +4469,7 @@ init_class (MonoClass *klass)

const char *name = m_class_get_name (klass);

#if defined(TARGET_AMD64) || defined(TARGET_ARM64) || defined(TARGET_WASM)
#if defined(TARGET_AMD64) || defined(TARGET_WASM)
/*
* Some of the intrinsics used by the VectorX classes are only implemented on amd64.
* The JIT can't handle SIMD types with != 16 size yet.
Expand All @@ -4481,6 +4481,14 @@ init_class (MonoClass *klass)
}
#endif

#ifdef TARGET_ARM64
if (!strcmp (m_class_get_name_space (klass), "System.Numerics")) {
// FIXME: Support Vector3 https://github.com/dotnet/runtime/issues/81501
if (!strcmp (name, "Vector2") || !strcmp (name, "Vector4") || !strcmp (name, "Quaternion") || !strcmp (name, "Plane"))
mono_class_set_is_simd_type (klass, TRUE);
}
#endif

if (m_class_is_ginst (klass)) {
if (!strcmp (name, "Vector`1") || !strcmp (name, "Vector64`1") || !strcmp (name, "Vector128`1") || !strcmp (name, "Vector256`1") || !strcmp (name, "Vector512`1")) {
MonoGenericClass *gclass = mono_class_try_get_generic_class (klass);
Expand Down
3 changes: 3 additions & 0 deletions src/mono/mono/mini/mini.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4573,6 +4573,9 @@ mini_get_simd_type_info (MonoClass *klass, guint32 *nelems)
if (!strcmp (klass_name, "Vector4") || !strcmp (klass_name, "Quaternion") || !strcmp (klass_name, "Plane")) {
*nelems = 4;
return MONO_TYPE_R4;
} else if (!strcmp (klass_name, "Vector2")) {
*nelems = 2;
return MONO_TYPE_R4;
} else if (!strcmp (klass_name, "Vector`1") || !strcmp (klass_name, "Vector64`1") || !strcmp (klass_name, "Vector128`1") || !strcmp (klass_name, "Vector256`1") || !strcmp (klass_name, "Vector512`1")) {
MonoType *etype = mono_class_get_generic_class (klass)->context.class_inst->type_argv [0];
int size = mono_class_value_size (klass, NULL);
Expand Down
8 changes: 8 additions & 0 deletions src/mono/mono/mini/simd-arm64.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,14 @@ SIMD_OP (64, OP_XCOMPARE_FP, CMP_LE, WTDSS_REV, _UNDEF

SIMD_OP (64, OP_XBINOP, OP_IADD, WTDSS, arm_neon_add, arm_neon_add, arm_neon_add, _UNDEF, _UNDEF, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FADD, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fadd, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FSUB, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fsub, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FMAX, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fmax, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FMIN, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fmin, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FMUL, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fmul, _UNDEF)
SIMD_OP (64, OP_XBINOP, OP_FDIV, WTDSS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fdiv, _UNDEF)
SIMD_OP (64, OP_ARM64_XADDV, INTRINS_AARCH64_ADV_SIMD_FADDV, WTDS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, _SKIP, _UNDEF)
SIMD_OP (64, OP_XOP_OVR_X_X, INTRINS_AARCH64_ADV_SIMD_FSQRT, WTDS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fsqrt, _UNDEF)
SIMD_OP (64, OP_XOP_OVR_X_X, INTRINS_AARCH64_ADV_SIMD_FABS, WTDS, _UNDEF, _UNDEF, _UNDEF, _UNDEF, arm_neon_fabs, _UNDEF)

/* 128-bit vectors */
/* Width Opcode Function Operand config I8 I16 I32 I64 F32 F64 */
Expand Down
36 changes: 18 additions & 18 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 @@ -1825,7 +1825,7 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
arg_class = mono_class_from_mono_type_internal (fsig->params [0]);
}

// FIXME: Add support for Vector64 on arm64
// FIXME: Add support for Vector64 on arm64 https://github.com/dotnet/runtime/issues/90402
int size = mono_class_value_size (arg_class, NULL);
if (size != 16)
return NULL;
Expand Down Expand Up @@ -2588,7 +2588,7 @@ emit_vector64_vector128_t (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
}

// System.Numerics.Vector2/Vector3/Vector4, Quaternion, and Plane
static guint16 vector2_methods[] = {
static guint16 vector_2_3_4_methods[] = {
SN_ctor,
SN_Abs,
SN_Add,
Expand Down Expand Up @@ -2631,12 +2631,12 @@ static G_GNUC_UNUSED MonoInst*
emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
{
MonoInst *ins;
int id;
int id, len;
MonoClass *klass;
MonoType *type, *etype;


id = lookup_intrins (vector2_methods, sizeof (vector2_methods), cmethod);
id = lookup_intrins (vector_2_3_4_methods, sizeof (vector_2_3_4_methods), cmethod);
if (id == -1) {
// https://github.com/dotnet/runtime/issues/81961
// check_no_intrinsic_cattr (cmethod);
Expand All @@ -2657,6 +2657,7 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
klass = cmethod->klass;
type = m_class_get_byval_arg (klass);
etype = m_class_get_byval_arg (mono_defaults.single_class);
len = mono_class_value_size (klass, NULL) / 4;

// Similar to the cases in emit_sys_numerics_vector_t ()
switch (id) {
Expand Down Expand Up @@ -2686,7 +2687,6 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
}
// FIXME: These don't work since Vector2/Vector3 are not handled as SIMD
#if 0
int len = mono_class_value_size (klass, NULL) / 4;
} else if (len == 3 && fsig->param_count == 2 && fsig->params [0]->type == MONO_TYPE_VALUETYPE && fsig->params [1]->type == etype->type) {
/* Vector3 (Vector2, float) */
int dreg = load_simd_vreg (cfg, cmethod, args [0], NULL);
Expand Down Expand Up @@ -2719,26 +2719,25 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
case SN_get_Item: {
// GetElement is marked as Intrinsic, but handling this in get_Item leads to better code
int src1 = load_simd_vreg (cfg, cmethod, args [0], NULL);
int elems = 4;
MonoTypeEnum ty = etype->type;

if (args [1]->opcode == OP_ICONST) {
// If the index is provably a constant, we can generate vastly better code.
int index = GTMREG_TO_INT (args[1]->inst_c0);

if (index < 0 || index >= elems) {
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, elems);
if (index < 0 || index >= len) {
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, len);
MONO_EMIT_NEW_COND_EXC (cfg, GE_UN, "ArgumentOutOfRangeException");
}

int opcode = type_to_extract_op (ty);
ins = emit_simd_ins (cfg, klass, opcode, src1, -1);
ins->inst_c0 = args[1]->inst_c0;
ins->inst_c0 = index;
ins->inst_c1 = ty;
return ins;
}

MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, elems);
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, len);
MONO_EMIT_NEW_COND_EXC (cfg, GE_UN, "ArgumentOutOfRangeException");

if (COMPILE_LLVM (cfg)) {
Expand Down Expand Up @@ -2806,14 +2805,14 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
g_assert (fsig->hasthis && fsig->param_count == 2 && fsig->params [0]->type == MONO_TYPE_I4 && fsig->params [1]->type == MONO_TYPE_R4);

gboolean indirect = FALSE;
int elems = 4, index = GTMREG_TO_INT (args [1]->inst_c0);
int index = GTMREG_TO_INT (args [1]->inst_c0);
int dreg = load_simd_vreg (cfg, cmethod, args [0], &indirect);

if (args [1]->opcode == OP_ICONST) {
// If the index is provably a constant, we can generate vastly better code.
// Bounds check only if the index is out of range
if (index < 0 || index >= elems) {
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, elems);
if (index < 0 || index >= len) {
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, len);
MONO_EMIT_NEW_COND_EXC (cfg, GE_UN, "ArgumentOutOfRangeException");
}

Expand All @@ -2830,7 +2829,7 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
return ins;
}

MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, elems);
MONO_EMIT_NEW_BIALU_IMM (cfg, OP_COMPARE_IMM, -1, args [1]->dreg, len);
MONO_EMIT_NEW_COND_EXC (cfg, GE_UN, "ArgumentOutOfRangeException");

if (COMPILE_LLVM (cfg)) {
Expand Down Expand Up @@ -2948,7 +2947,7 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
#endif
}
case SN_CopyTo:
// FIXME:
// FIXME: https://github.com/dotnet/runtime/issues/91394
return NULL;
case SN_Clamp: {
if (!(!fsig->hasthis && fsig->param_count == 3 && mono_metadata_type_equal (fsig->ret, type) && mono_metadata_type_equal (fsig->params [0], type) && mono_metadata_type_equal (fsig->params [1], type) && mono_metadata_type_equal (fsig->params [2], type)))
Expand All @@ -2971,7 +2970,7 @@ emit_vector_2_3_4 (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
case SN_LengthSquared:
case SN_Lerp:
case SN_Normalize: {
// FIXME:
// FIXME: https://github.com/dotnet/runtime/issues/91394
return NULL;
}
default:
Expand Down Expand Up @@ -5832,8 +5831,9 @@ arch_emit_simd_intrinsics (const char *class_ns, const char *class_name, MonoCom
}

if (!strcmp (class_ns, "System.Numerics")) {
// FIXME: Support Vector2/Vector3
if (!strcmp (class_name, "Vector4") || !strcmp (class_name, "Quaternion") || !strcmp (class_name, "Plane"))
// FIXME: Support Vector2 https://github.com/dotnet/runtime/issues/81501
if (!strcmp (class_name, "Vector2") || !strcmp (class_name, "Vector4") ||
!strcmp (class_name, "Quaternion") || !strcmp (class_name, "Plane"))
return emit_vector_2_3_4 (cfg, cmethod, fsig, args);
}

Expand Down