adding __packed_store_active2 to generic targets

This commit is contained in:
Ilia Filippov
2013-12-19 14:13:55 +04:00
parent d666fc3f8f
commit 15816eb07e
9 changed files with 125 additions and 65 deletions

View File

@@ -3,6 +3,7 @@
define(`MASK',`i32')
define(`WIDTH',`1')
include(`util.m4')
rdrand_decls()
; Define some basics for a 1-wide target
stdlib_core()
packed_load_and_store()
@@ -655,7 +656,7 @@ define <1 x float> @__rsqrt_varying_float(<1 x float> %v) nounwind readonly alw
declare <1 x float> @__svml_sind(<1 x float>) nounwind readnone alwaysinline
declare <1 x float> @__svml_asind(<1 x float>) nounwind readnone alwaysinline
declare <1 x float> @__svml_cosd(<1 x float>) nounwind readnone alwaysinline
declare void @__svml_sincosd(<1 x float>, <1 x float> *, <1 x float> *) nounwind readnone alwaysinline
declare void @__svml_sincosd(<1 x float>, <1 x double> *, <1 x double> *) nounwind readnone alwaysinline
declare <1 x float> @__svml_tand(<1 x float>) nounwind readnone alwaysinline
declare <1 x float> @__svml_atand(<1 x float>) nounwind readnone alwaysinline
declare <1 x float> @__svml_atan2d(<1 x float>, <1 x float>) nounwind readnone alwaysinline

View File

@@ -371,6 +371,8 @@ declare i32 @__packed_load_active(i32 * nocapture, <WIDTH x i32> * nocapture,
<WIDTH x i1>) nounwind
declare i32 @__packed_store_active(i32 * nocapture, <WIDTH x i32> %vals,
<WIDTH x i1>) nounwind
declare i32 @__packed_store_active2(i32 * nocapture, <WIDTH x i32> %vals,
<WIDTH x i1>) nounwind
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

View File

@@ -1472,31 +1472,38 @@ static FORCEINLINE int32_t __packed_store_active(int32_t *ptr, __vec16_i32 val,
return count;
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *ptr, __vec16_i32 val,
__vec16_i1 mask) {
int count = 0;
int32_t *ptr_ = ptr;
for (int i = 0; i < 16; ++i) {
*ptr = val.v[i];
ptr += mask.v & 1;
mask.v = mask.v >> 1;
}
return ptr - ptr_;
}
static FORCEINLINE int32_t __packed_load_active(uint32_t *ptr,
__vec16_i32 *val,
__vec16_i1 mask) {
int count = 0;
for (int i = 0; i < 16; ++i) {
if ((mask.v & (1 << i)) != 0) {
val->v[i] = *ptr++;
++count;
}
}
return count;
return __packed_load_active((int32_t *)ptr, val, mask);
}
static FORCEINLINE int32_t __packed_store_active(uint32_t *ptr,
__vec16_i32 val,
__vec16_i1 mask) {
int count = 0;
for (int i = 0; i < 16; ++i) {
if ((mask.v & (1 << i)) != 0) {
*ptr++ = val.v[i];
++count;
return __packed_store_active((int32_t *)ptr, val, mask);
}
}
return count;
static FORCEINLINE int32_t __packed_store_active2(uint32_t *ptr,
__vec16_i32 val,
__vec16_i1 mask) {
return __packed_store_active2((int32_t *)ptr, val, mask);
}

View File

@@ -1523,31 +1523,38 @@ static FORCEINLINE int32_t __packed_store_active(int32_t *ptr, __vec32_i32 val,
return count;
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *ptr, __vec32_i32 val,
__vec32_i1 mask) {
int count = 0;
int32_t *ptr_ = ptr;
for (int i = 0; i < 32; ++i) {
*ptr = val.v[i];
ptr += mask.v & 1;
mask.v = mask.v >> 1;
}
return ptr - ptr_;
}
static FORCEINLINE int32_t __packed_load_active(uint32_t *ptr,
__vec32_i32 *val,
__vec32_i1 mask) {
int count = 0;
for (int i = 0; i < 32; ++i) {
if ((mask.v & (1 << i)) != 0) {
val->v[i] = *ptr++;
++count;
}
}
return count;
return __packed_load_active((int32_t *)ptr, val, mask);
}
static FORCEINLINE int32_t __packed_store_active(uint32_t *ptr,
__vec32_i32 val,
__vec32_i1 mask) {
int count = 0;
for (int i = 0; i < 32; ++i) {
if ((mask.v & (1 << i)) != 0) {
*ptr++ = val.v[i];
++count;
return __packed_store_active((int32_t *)ptr, val, mask);
}
}
return count;
static FORCEINLINE int32_t __packed_store_active2(uint32_t *ptr,
__vec32_i32 val,
__vec32_i1 mask) {
return __packed_store_active2((int32_t *)ptr, val, mask);
}

View File

@@ -1656,31 +1656,38 @@ static FORCEINLINE int32_t __packed_store_active(int32_t *ptr, __vec64_i32 val,
return count;
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *ptr, __vec64_i32 val,
__vec64_i1 mask) {
int count = 0;
int32_t *ptr_ = ptr;
for (int i = 0; i < 64; ++i) {
*ptr = val.v[i];
ptr += mask.v & 1;
mask.v = mask.v >> 1;
}
return ptr - ptr_;
}
static FORCEINLINE int32_t __packed_load_active(uint32_t *ptr,
__vec64_i32 *val,
__vec64_i1 mask) {
int count = 0;
for (int i = 0; i < 64; ++i) {
if ((mask.v & (1ull << i)) != 0) {
val->v[i] = *ptr++;
++count;
}
}
return count;
return __packed_load_active((int32_t *) ptr, val, mask);
}
static FORCEINLINE int32_t __packed_store_active(uint32_t *ptr,
__vec64_i32 val,
__vec64_i1 mask) {
int count = 0;
for (int i = 0; i < 64; ++i) {
if ((mask.v & (1ull << i)) != 0) {
*ptr++ = val.v[i];
++count;
return __packed_store_active((int32_t *) ptr, val, mask);
}
}
return count;
static FORCEINLINE int32_t __packed_store_active2(uint32_t *ptr,
__vec64_i32 val,
__vec64_i1 mask) {
return __packed_store_active2((int32_t *) ptr, val, mask);
}

View File

@@ -2451,20 +2451,24 @@ static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val, _
return _mm_countbits_32(uint32_t(mask));
}
static FORCEINLINE int32_t __packed_store_active2(uint32_t *p, __vec16_i32 val, __vec16_i1 mask)
{
return __packed_store_active(p, val, mask);
}
static FORCEINLINE int32_t __packed_load_active(int32_t *p, __vec16_i32 *val, __vec16_i1 mask)
{
__vec16_i32 v = __load<64>(val);
v = _mm512_mask_extloadunpacklo_epi32(v, mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
v = _mm512_mask_extloadunpackhi_epi32(v, mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
__store<64>(val, v);
return _mm_countbits_32(uint32_t(mask));
return __packed_load_active((uint32_t *)p, val, mask);
}
static FORCEINLINE int32_t __packed_store_active(int32_t *p, __vec16_i32 val, __vec16_i1 mask)
{
_mm512_mask_extpackstorelo_epi32(p, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
_mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
return _mm_countbits_32(uint32_t(mask));
return __packed_store_active((uint32_t *)p, val, mask);
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *p, __vec16_i32 val, __vec16_i1 mask)
{
return __packed_store_active(p, val, mask);
}
///////////////////////////////////////////////////////////////////////////

View File

@@ -2496,20 +2496,23 @@ static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec8_i32 val,
_mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, 0xFF & mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
return _mm_countbits_32(uint32_t(0xFF & mask));
}
static FORCEINLINE int32_t __packed_store_active2(uint32_t *ptr, __vec4_i32 val,
__vec4_i1 mask) {
return __packed_store_active(ptr, val, mask);
}
static FORCEINLINE int32_t __packed_load_active(int32_t *p, __vec8_i32 *val,
__vec8_i1 mask) {
__vec8_i32 v = __load<64>(val);
v = _mm512_mask_extloadunpacklo_epi32(v, 0xFF & mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
v = _mm512_mask_extloadunpackhi_epi32(v, 0xFF & mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
__store<64>(val, v);
return _mm_countbits_32(uint32_t(0xFF & mask));
return __packed_load_active((uint32_t *)p, val, mask);
}
static FORCEINLINE int32_t __packed_store_active(int32_t *p, __vec8_i32 val,
__vec8_i1 mask) {
_mm512_mask_extpackstorelo_epi32(p, 0xFF & mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
_mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, 0xFF & mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
return _mm_countbits_32(uint32_t(0xFF & mask));
return __packed_store_active((uint32_t *)p, val, mask);
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *ptr, __vec4_i32 val,
__vec4_i1 mask) {
return __packed_store_active(ptr, val, mask);
}
#endif
///////////////////////////////////////////////////////////////////////////

View File

@@ -1878,6 +1878,11 @@ static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val,
return _mm_countbits_32(uint32_t(mask));
}
static FORCEINLINE int32_t __packed_store_active2(uint32_t *p, __vec16_i32 val, __vec16_i1 mask)
{
return __packed_store_active(p, val, mask);
}
///////////////////////////////////////////////////////////////////////////
// prefetch
///////////////////////////////////////////////////////////////////////////

View File

@@ -3798,6 +3798,25 @@ static FORCEINLINE int32_t __packed_store_active(int32_t *ptr, __vec4_i32 val,
return count;
}
static FORCEINLINE int32_t __packed_store_active2(int32_t *ptr, __vec4_i32 val,
__vec4_i1 mask) {
int count = 0;
ptr[count] = _mm_extract_epi32(val.v, 0);
count -= _mm_extract_ps(mask.v, 0);
ptr[count] = _mm_extract_epi32(val.v, 1);
count -= _mm_extract_ps(mask.v, 1);
ptr[count] = _mm_extract_epi32(val.v, 2);
count -= _mm_extract_ps(mask.v, 2);
ptr[count] = _mm_extract_epi32(val.v, 3);
count -= _mm_extract_ps(mask.v, 3);
return count;
}
static FORCEINLINE int32_t __packed_load_active(uint32_t *ptr, __vec4_i32 *val,
__vec4_i1 mask) {
return __packed_load_active((int32_t *)ptr, val, mask);
@@ -3808,6 +3827,11 @@ static FORCEINLINE int32_t __packed_store_active(uint32_t *ptr, __vec4_i32 val,
return __packed_store_active((int32_t *)ptr, val, mask);
}
static FORCEINLINE int32_t __packed_store_active2(uint32_t *ptr, __vec4_i32 val,
__vec4_i1 mask) {
return __packed_store_active2((int32_t *)ptr, val, mask);
}
///////////////////////////////////////////////////////////////////////////
// aos/soa