Skip to content

[AIE2P] Fix broadcast intrinsics. #503

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

Open
wants to merge 1 commit into
base: aie-public
Choose a base branch
from
Open

[AIE2P] Fix broadcast intrinsics. #503

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 0 additions & 8 deletions clang/lib/Headers/aie2p_aie_api_compat.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -327,16 +327,8 @@ inline __attribute__((always_inline)) v16cint16 shift_bytes(v16cint16 , v16cint1
inline __attribute__((always_inline)) v8cint32 shift_bytes(v8cint32 , v8cint32 , unsigned int );
inline __attribute__((always_inline)) v16cint16 shift(v16cint16 , v16cint16 , unsigned int );
inline __attribute__((always_inline)) v8cint32 shift(v8cint32 , v8cint32 , unsigned int );
inline __attribute__((always_inline)) v64int8 broadcast_s8(int );
inline __attribute__((always_inline)) v32int16 broadcast_s16(int );
inline __attribute__((always_inline)) v64uint8 broadcast_u8(unsigned int );
inline __attribute__((always_inline)) v32uint16 broadcast_u16(unsigned int );
inline __attribute__((always_inline)) v16cint16 broadcast_c16(cint16 );
inline __attribute__((always_inline)) v8cint32 broadcast_c32(cint32 );
inline __attribute__((always_inline)) v32int16 broadcast_to_v32int16(int );
inline __attribute__((always_inline)) v64uint8 broadcast_to_v64uint8(unsigned int );
inline __attribute__((always_inline)) v32uint16 broadcast_to_v32uint16(unsigned int );
inline __attribute__((always_inline)) v16uint32 broadcast_to_v16uint32(v2uint32 );
inline __attribute__((always_inline)) v16cint16 broadcast_to_v16cint16(cint16 );
inline __attribute__((always_inline)) v16cint16 broadcast_to_v16cint16(v2cint16 );
inline __attribute__((always_inline)) v8cint32 broadcast_to_v8cint32(cint32 );
Expand Down
84 changes: 37 additions & 47 deletions clang/lib/Headers/aie2p_scl2vec.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,14 +31,14 @@ inline decltype(auto) vector_extract64(T a, int idx, int sign) {
return sign ? (v2int32){a[idx], a[idx + 1]} : (v2uint32){a[idx], a[idx + 1]};
}

inline v16int32 vector_broadcast64(v2int32 b) {
return {b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1],
b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1]};
}
inline v16uint32 vector_broadcast64(v2uint32 b) {
return {b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1],
b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1]};
}
#define VECTOR_BROADCAST64_FUNC(outType, inType) \
inline outType vector_broadcast64(inType b) { \
return {b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1], \
b[0], b[1], b[0], b[1], b[0], b[1], b[0], b[1]}; \
}

VECTOR_BROADCAST64_FUNC(v16int32, v2int32)
VECTOR_BROADCAST64_FUNC(v16uint32, v2uint32)

INTRINSIC(v128int4)
shiftx(v128int4 a, v128int4 b, int step, unsigned int shift) {
Expand Down Expand Up @@ -1354,10 +1354,10 @@ INTRINSIC(void *) extract_address(v16int32 v, int idx) {
}
// broadcast from scalar (alternative syntax to broadcast to vector)
INTRINSIC(v64int8)
broadcast_s8(char b) { return b - v64int8{0}; }
broadcast_s8(int b) { return (char)b - v64int8{0}; }
Copy link
Collaborator

Choose a reason for hiding this comment

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

Can we update the tests as well?

Copy link
Collaborator

Choose a reason for hiding this comment

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

As I remember, we reused aie2 tests for aie2p as well. Since at it, can you please fix it for aie2 as well.

Copy link
Collaborator

Choose a reason for hiding this comment

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

v64int8{b} doesn't work?

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

No, that will only initialize the first element of the vector with b.
But the subtraction (char)b - v64int8{0} is performed element-wise due to how vector types work, effectively resulting in a vector where each element is b.

We don't even need to specify 0 when initializing a v64int8 vector, since all elements of a vector are initialized to 0 by default.
So (char)b - v64int8{} will also work.


INTRINSIC(v32int16)
broadcast_s16(short b) { return b - v32int16{0}; }
broadcast_s16(int b) { return (short)b - v32int16{0}; }
Copy link
Collaborator

Choose a reason for hiding this comment

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

why weren't we happy with implicit casting?

Copy link
Collaborator

Choose a reason for hiding this comment

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

short b is promoted to int


INTRINSIC(v16int32)
broadcast_s32(int b) { return b - v16int32{0}; }
Expand All @@ -1369,10 +1369,10 @@ INTRINSIC(v16int32)
broadcast_v2s32(v2int32 b) { return vector_broadcast64(b); }

INTRINSIC(v64uint8)
broadcast_u8(unsigned char b) { return b - v64uint8{0}; }
broadcast_u8(unsigned int b) { return (unsigned char)b - v64uint8{0}; }

INTRINSIC(v32uint16)
broadcast_u16(unsigned short b) { return b - v32uint16{0}; }
broadcast_u16(unsigned int b) { return (unsigned short)b - v32uint16{0}; }

INTRINSIC(v16uint32)
broadcast_u32(unsigned int b) { return b - v16uint32{0}; }
Expand Down Expand Up @@ -1402,10 +1402,10 @@ broadcast_c32 (cint32 b) { return vector_broadcast64(b);}

// broadcast to vector (alternative syntax to broadcast from scalar)
INTRINSIC(v128int4)
broadcast_to_v128int4(v2int4 b) { return b - v128int4{0}; }
broadcast_to_v128int4(v2int4 b) { return broadcast_s8((int)(char)b); }

INTRINSIC(v128int4)
broadcast_to_v128int4(v4int4 b) { return broadcast_s16((short)b); }
broadcast_to_v128int4(v4int4 b) { return broadcast_s16((int)(short)b); }

INTRINSIC(v128int4)
broadcast_to_v128int4(v8int4 b) { return broadcast_s32((int)b); }
Expand All @@ -1414,10 +1414,10 @@ INTRINSIC(v128int4)
broadcast_to_v128int4(v16int4 b) { return vector_broadcast64((v2int32)b); }

INTRINSIC(v64int8)
broadcast_to_v64int8(char b) { return b - v64int8{0}; }
broadcast_to_v64int8(int b) { return broadcast_s8(b); }

INTRINSIC(v64int8)
broadcast_to_v64int8(v2int8 b) { return broadcast_s16((short)b); }
broadcast_to_v64int8(v2int8 b) { return broadcast_s16((int)(short)b); }

INTRINSIC(v64int8)
broadcast_to_v64int8(v4int8 b) { return broadcast_s32((int)b); }
Expand All @@ -1426,7 +1426,7 @@ INTRINSIC(v64int8)
broadcast_to_v64int8(v8int8 b) { return vector_broadcast64((v2int32)b); }

INTRINSIC(v32int16)
broadcast_to_v32int16(short b) { return b - v32int16{0}; }
broadcast_to_v32int16(int b) { return broadcast_s16(b); }

INTRINSIC(v32int16)
broadcast_to_v32int16(v2int16 b) { return broadcast_s32((int)b); }
Expand All @@ -1435,7 +1435,7 @@ INTRINSIC(v32int16)
broadcast_to_v32int16(v4int16 b) { return vector_broadcast64((v2int32)b); }

INTRINSIC(v16int32)
broadcast_to_v16int32(int b) { return b - v16int32{0}; }
broadcast_to_v16int32(int b) { return broadcast_s32(b); }

INTRINSIC(v16int32)
broadcast_to_v16int32(mask64 b) { return vector_broadcast64((v2int32)b); }
Expand All @@ -1444,10 +1444,14 @@ INTRINSIC(v16int32)
broadcast_to_v16int32(v2int32 b) { return vector_broadcast64(b); }

INTRINSIC(v128uint4)
broadcast_to_v128uint4(v2uint4 b) { return b - v128uint4{0}; }
broadcast_to_v128uint4(v2uint4 b) {
return broadcast_u8((unsigned int)(unsigned char)b);
}

INTRINSIC(v128uint4)
broadcast_to_v128uint4(v4uint4 b) { return broadcast_u16((unsigned short)b); }
broadcast_to_v128uint4(v4uint4 b) {
return broadcast_u16((unsigned int)(unsigned short)b);
}

INTRINSIC(v128uint4)
broadcast_to_v128uint4(v8uint4 b) { return broadcast_u32((unsigned int)b); }
Expand All @@ -1456,10 +1460,12 @@ INTRINSIC(v128uint4)
broadcast_to_v128uint4(v16uint4 b) { return vector_broadcast64((v2uint32)b); }

INTRINSIC(v64uint8)
broadcast_to_v64uint8(unsigned char b) { return b - v64uint8{0}; }
broadcast_to_v64uint8(unsigned int b) { return broadcast_u8(b); }

INTRINSIC(v64uint8)
broadcast_to_v64uint8(v2uint8 b) { return broadcast_u16((unsigned short)b); }
broadcast_to_v64uint8(v2uint8 b) {
return broadcast_u16((unsigned int)(unsigned short)b);
}

INTRINSIC(v64uint8)
broadcast_to_v64uint8(v4uint8 b) { return broadcast_u32((unsigned int)b); }
Expand All @@ -1468,7 +1474,7 @@ INTRINSIC(v64uint8)
broadcast_to_v64uint8(v8uint8 b) { return vector_broadcast64((v2uint32)b); }

INTRINSIC(v32uint16)
broadcast_to_v32uint16(unsigned short b) { return b - v32uint16{0}; }
broadcast_to_v32uint16(unsigned int b) { return broadcast_u16(b); }

INTRINSIC(v32uint16)
broadcast_to_v32uint16(v2uint16 b) { return broadcast_u32((unsigned int)b); }
Expand All @@ -1477,7 +1483,7 @@ INTRINSIC(v32uint16)
broadcast_to_v32uint16(v4uint16 b) { return vector_broadcast64((v2uint32)b); }

INTRINSIC(v16uint32)
broadcast_to_v16uint32(unsigned int b) { return b - v16uint32{0}; }
broadcast_to_v16uint32(unsigned int b) { return broadcast_u32(b); }

INTRINSIC(v16uint32)
broadcast_to_v16uint32(mask64 b) { return vector_broadcast64((v2uint32)b); }
Expand Down Expand Up @@ -1530,15 +1536,6 @@ broadcast_to_v16float(v2float b) {
return broadcast_s64(as_mask64);
}

INTRINSIC(v32bfloat16)
broadcast_zero_to_v32bfloat16() { return broadcast_to_v32bfloat16(0); }

INTRINSIC(v32bfloat16)
broadcast_one_to_v32bfloat16() { return broadcast_to_v32bfloat16(1); }

INTRINSIC(v16float)
broadcast_one_to_v16float() { return broadcast_to_v16float(1); }

// Right-most insertion (left shift)
INTRINSIC(v64int8) shiftl_elem(v64int8 v, int s) {
return shift_bytes(v, broadcast_s8(s), 1);
Expand Down Expand Up @@ -1644,13 +1641,16 @@ INTRINSIC(v16uint32) broadcast_one_to_v16uint32() {
INTRINSIC(v32bfloat16) broadcast_one_bfloat16() {
return broadcast_bfloat16(1);
}
INTRINSIC(v32bfloat16)
broadcast_one_to_v32bfloat16() { return broadcast_one_bfloat16(); }

INTRINSIC(v16float)
broadcast_one_to_v16float() { return broadcast_to_v16float(1); }

#if 0
INTRINSIC(v16cint16) broadcast_one_c16() { return broadcast_c16(1); }

INTRINSIC(v8cint32) broadcast_one_c32() { return broadcast_c32(1); }

INTRINSIC(v16float) broadcast_one_float() { return broadcast_float(1); }
#endif

// broadcast value zero(0) to all vector lanes
Expand Down Expand Up @@ -1708,13 +1708,9 @@ INTRINSIC(v16uint32) broadcast_zero_to_v16uint32() { return broadcast_u32(0); }
broadcast_zero_bfloat16() {
return broadcast_bfloat16(0);
}
INTRINSIC(v32bfloat16)
broadcast_zero_to_v32bfloat16() { return broadcast_to_v32bfloat16(0); }

[[deprecated(
"Function 'broadcast_zero_float' is deprecated. Please use the "
"'broadcast_zero_to_v16float' variant instead.")]] INTRINSIC(v16float)
broadcast_zero_float() {
return broadcast_float(0);
}
INTRINSIC(v16float) broadcast_zero_to_v16float() { return broadcast_float(0); }

#if 0
Expand Down Expand Up @@ -1791,25 +1787,19 @@ broadcast_elem(v16float v, int idx) {
return vector_broadcast64(ext_v2int32(v, idx, 0));
}

INTRINSIC(v64int8)
broadcast_to_v64int8(int b) { return broadcast_s8((int)b); }

INTRINSIC(v16acc64) broadcast_zero_to_v16acc64() { return v16acc64{}; }

[[deprecated("Function 'clr' is deprecated. Please use the 'broadcast_zero_to' "
"variant instead.")]] INTRINSIC(v16acc64) clr16() {
return broadcast_zero_to_v16acc64();
}

INTRINSIC(v32acc64) broadcast_zero_to_v32acc64() { return v32acc64{}; }

[[deprecated("Function 'clr' is deprecated. Please use the 'broadcast_zero_to' "
"variant instead.")]] INTRINSIC(v32acc64) clr32() {
return broadcast_zero_to_v32acc64();
}

INTRINSIC(v64acc32) broadcast_zero_to_v64acc32() { return v64acc32{}; }

[[deprecated("Function 'clr' is deprecated. Please use the 'broadcast_zero_to' "
"variant instead.")]] INTRINSIC(v64acc32) clr64() {
return broadcast_zero_to_v64acc32();
Expand Down