Skip to content

Commit

Permalink
neon/ld4_lane: move private type usage to inside loop
Browse files Browse the repository at this point in the history 
Beyond requiring a bit less memory, this should make things a bit
easier for the autovectorize since the code it needs to worry about
is all contained inside of each loop iteration.
  • Loading branch information
@nemequ
nemequ committed May 28, 2021
1 parent a973cab commit 0d1ab79
Showing 1 changed file with 31 additions and 48 deletions.
79 changes: 31 additions & 48 deletions simde/arm/neon/ld4_lane.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
*
* Copyright:
* 2021 Zhi An Ng <[email protected]> (Copyright owned by Google, LLC)
* 2021 Evan Nemerson <[email protected]>
*/

#if !defined(SIMDE_ARM_NEON_LD4_LANE_H)
Expand All @@ -39,18 +40,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_int8x8x4_t
simde_vld4_lane_s8(int8_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_int8x8x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 7) {
simde_int8x8_private a_[4];
simde_int8x8x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_int8x8_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_int8x8_private tmp_ = simde_int8x8_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_int8x8_from_private(tmp_);
}

simde_int8x8x4_t s_ = { {
simde_int8x8_from_private(a_[0]), simde_int8x8_from_private(a_[1]),
simde_int8x8_from_private(a_[2]), simde_int8x8_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_s8(ptr, src, lane) vld4_lane_s8(ptr, src, lane)
Expand All @@ -64,18 +62,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_int16x4x4_t
simde_vld4_lane_s16(int16_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_int16x4x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 3) {
simde_int16x4_private a_[4];
simde_int16x4x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_int16x4_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_int16x4_private tmp_ = simde_int16x4_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_int16x4_from_private(tmp_);
}

simde_int16x4x4_t s_ = { {
simde_int16x4_from_private(a_[0]), simde_int16x4_from_private(a_[1]),
simde_int16x4_from_private(a_[2]), simde_int16x4_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_s16(ptr, src, lane) vld4_lane_s16(ptr, src, lane)
Expand All @@ -89,18 +84,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_int32x2x4_t
simde_vld4_lane_s32(int32_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_int32x2x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 1) {
simde_int32x2_private a_[4];
simde_int32x2x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_int32x2_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_int32x2_private tmp_ = simde_int32x2_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_int32x2_from_private(tmp_);
}

simde_int32x2x4_t s_ = { {
simde_int32x2_from_private(a_[0]), simde_int32x2_from_private(a_[1]),
simde_int32x2_from_private(a_[2]), simde_int32x2_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_s32(ptr, src, lane) vld4_lane_s32(ptr, src, lane)
Expand All @@ -114,18 +106,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_uint8x8x4_t
simde_vld4_lane_u8(uint8_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_uint8x8x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 7) {
simde_uint8x8_private a_[4];
simde_uint8x8x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_uint8x8_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_uint8x8_private tmp_ = simde_uint8x8_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_uint8x8_from_private(tmp_);
}

simde_uint8x8x4_t s_ = { {
simde_uint8x8_from_private(a_[0]), simde_uint8x8_from_private(a_[1]),
simde_uint8x8_from_private(a_[2]), simde_uint8x8_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_u8(ptr, src, lane) vld4_lane_u8(ptr, src, lane)
Expand All @@ -139,18 +128,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_uint16x4x4_t
simde_vld4_lane_u16(uint16_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_uint16x4x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 3) {
simde_uint16x4_private a_[4];
simde_uint16x4x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_uint16x4_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_uint16x4_private tmp_ = simde_uint16x4_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_uint16x4_from_private(tmp_);
}

simde_uint16x4x4_t s_ = { {
simde_uint16x4_from_private(a_[0]), simde_uint16x4_from_private(a_[1]),
simde_uint16x4_from_private(a_[2]), simde_uint16x4_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_u16(ptr, src, lane) vld4_lane_u16(ptr, src, lane)
Expand All @@ -164,18 +150,15 @@ SIMDE_FUNCTION_ATTRIBUTES
simde_uint32x2x4_t
simde_vld4_lane_u32(uint32_t const ptr[HEDLEY_ARRAY_PARAM(4)], simde_uint32x2x4_t src, const int lane)
SIMDE_REQUIRE_CONSTANT_RANGE(lane, 0, 1) {
simde_uint32x2_private a_[4];
simde_uint32x2x4_t r;

for (size_t i = 0 ; i < 4 ; i++) {
a_[i] = simde_uint32x2_to_private(src.val[i]);
a_[i].values[lane] = ptr[i];
simde_uint32x2_private tmp_ = simde_uint32x2_to_private(src.val[i]);
tmp_.values[lane] = ptr[i];
r.val[i] = simde_uint32x2_from_private(tmp_);
}

simde_uint32x2x4_t s_ = { {
simde_uint32x2_from_private(a_[0]), simde_uint32x2_from_private(a_[1]),
simde_uint32x2_from_private(a_[2]), simde_uint32x2_from_private(a_[3]) } };

return s_;
return r;
}
#if defined(SIMDE_ARM_NEON_A32V7_NATIVE)
#define simde_vld4_lane_u32(ptr, src, lane) vld4_lane_u32(ptr, src, lane)
Expand Down

0 comments on commit 0d1ab79

Please sign in to comment.