OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nir/opt_large_constants: Handle small float arrays

Browse Source 
Handles small arrays of integer, positive floats.

RADV fossils:

Totals from 65 (0.05% of 131205) affected shaders:
Instrs: 30001 -> 29936 (-0.22%); split: -0.39%, +0.18%
CodeSize: 165676 -> 164996 (-0.41%); split: -0.53%, +0.12%
Latency: 126873 -> 127178 (+0.24%); split: -0.29%, +0.53%
InvThroughput: 26640 -> 26895 (+0.96%); split: -0.48%, +1.44%
VClause: 425 -> 371 (-12.71%)
SClause: 982 -> 981 (-0.10%); split: -0.92%, +0.81%
Copies: 2072 -> 1939 (-6.42%); split: -6.52%, +0.10%
PreVGPRs: 1553 -> 1537 (-1.03%)

Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9000>
This commit is contained in:
Konstantin Seurer
2023-07-17 15:08:43 +02:00
committed by Marge Bot
parent e38522608f
commit 3aa3eb8ddd
1 changed files with 28 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/compiler/nir/nir_opt_large_constants.c
View File

@@ -113,6 +113,7 @@ write_const_values(void *dst, const nir_const_value *src,
struct small_constant { struct small_constant {
uint64_t data; uint64_t data;
uint32_t bit_size; uint32_t bit_size;
bool is_float;
uint32_t bit_stride; uint32_t bit_stride;
}; };
@@ -251,9 +252,26 @@ get_small_constant(struct var_info *info, glsl_type_size_align_func size_align)
nir_const_value values[64]; nir_const_value values[64];
read_const_values(values, info->constant_data, array_len, bit_size); read_const_values(values, info->constant_data, array_len, bit_size);
bool is_float = true;
if (bit_size < 16) {
is_float = false;
} else {
for (unsigned i = 0; i < array_len; i++) {
/* See if it's an easily convertible float.
* TODO: Compute greatest common divisor to support non-integer floats.
* TODO: Compute min value and add it to the result of
* build_small_constant_load for handling negative floats.
*/
uint64_t u = nir_const_value_as_float(values[i], bit_size);
nir_const_value fc = nir_const_value_for_float(u, bit_size);
is_float &= !memcmp(&fc, &values[i], bit_size / 8);
}
}
uint32_t used_bits = 0; uint32_t used_bits = 0;
for (unsigned i = 0; i < array_len; i++) { for (unsigned i = 0; i < array_len; i++) {
uint64_t u64_elem = nir_const_value_as_uint(values[i], bit_size); uint64_t u64_elem = is_float ? nir_const_value_as_float(values[i], bit_size)
: nir_const_value_as_uint(values[i], bit_size);
if (!u64_elem) if (!u64_elem)
continue; continue;
@@ -272,13 +290,16 @@ get_small_constant(struct var_info *info, glsl_type_size_align_func size_align)
info->is_small = true; info->is_small = true;
for (unsigned i = 0; i < array_len; i++) { for (unsigned i = 0; i < array_len; i++) {
uint64_t u64_elem = nir_const_value_as_uint(values[i], bit_size); uint64_t u64_elem = is_float ? nir_const_value_as_float(values[i], bit_size)
: nir_const_value_as_uint(values[i], bit_size);
info->small_constant.data |= u64_elem << (i * used_bits); info->small_constant.data |= u64_elem << (i * used_bits);
} }
/* Limit bit_size >= 32 to avoid unnecessary conversions. */ /* Limit bit_size >= 32 to avoid unnecessary conversions. */
info->small_constant.bit_size = info->small_constant.bit_size =
MAX2(util_next_power_of_two(used_bits * array_len), 32); MAX2(util_next_power_of_two(used_bits * array_len), 32);
info->small_constant.is_float = is_float;
info->small_constant.bit_stride = used_bits; info->small_constant.bit_stride = used_bits;
} }
@@ -303,8 +324,11 @@ build_small_constant_load(nir_builder *b, nir_deref_instr *deref,
/* Booleans are special-cased to be 32-bit */ /* Booleans are special-cased to be 32-bit */
assert(glsl_type_is_boolean(deref->type)); assert(glsl_type_is_boolean(deref->type));
ret = nir_ine_imm(b, ret, 0); ret = nir_ine_imm(b, ret, 0);
} else if (bit_size != constant->bit_size) { } else {
ret = nir_u2uN(b, ret, bit_size); if (constant->is_float)
ret = nir_u2fN(b, ret, bit_size);
else if (bit_size != constant->bit_size)
ret = nir_u2uN(b, ret, bit_size);
} }
return ret; return ret;