fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-15 05:38:11 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

nir/coopmat: rename the box split variables.
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
When we add workgroup support we have to add inner box splits as well
so name the outer splits correctly.

Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41500>
This commit is contained in:
David Airlie 2026-05-12 18:37:09 +10:00 • committed by Marge Bot
parent 6062bcde56
commit 6a7a46ac21
1 changed files with 36 additions and 36 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

72
src/compiler/nir/nir_lower_cooperative_matrix.c
View file

@ -11,8 +11,8 @@
* Lower flexible size cooperative matrix operations down to operations at the supported granularity.
*/
struct split_box {
unsigned num_row_splits;
unsigned num_col_splits;
unsigned outer_rows;
unsigned outer_cols;
};
struct split_mat {
@ -29,7 +29,7 @@ struct split_info {
static unsigned get_num_splits_box(const struct split_box *box)
{
return box->num_row_splits * box->num_col_splits;
return box->outer_rows * box->outer_cols;
}
static unsigned get_num_splits(struct split_mat *split)
@ -134,21 +134,21 @@ split_desc(struct glsl_cmat_description *desc, struct split_info *info,
struct split_box *box)
{
unsigned split_rows = 0, split_cols = 0;
box->num_col_splits = 1;
box->num_row_splits = 1;
box->outer_cols = 1;
box->outer_rows = 1;
get_lower_sizes(*desc, info->m_gran, info->n_gran, info->k_gran, &split_rows, &split_cols);
if (split_rows) {
box->num_row_splits = desc->rows / split_rows;
box->outer_rows = desc->rows / split_rows;
desc->rows = split_rows;
}
if (split_cols) {
box->num_col_splits = desc->cols / split_cols;
box->outer_cols = desc->cols / split_cols;
desc->cols = split_cols;
}
if (box->num_row_splits == 1 && box->num_col_splits == 1)
if (box->outer_rows == 1 && box->outer_cols == 1)
return false;
return true;
}
@ -358,9 +358,9 @@ split_cmat_transpose(nir_builder *b,
unsigned dst_splits = get_num_splits(dst_split);
for (unsigned i = 0; i < dst_splits; i++) {
int r = i / dst_split->box.num_row_splits;
int c = i % dst_split->box.num_row_splits;
int out_idx = c * dst_split->box.num_col_splits + r;
int r = i / dst_split->box.outer_rows;
int c = i % dst_split->box.outer_rows;
int out_idx = c * dst_split->box.outer_cols + r;
nir_deref_instr *dst_deref = recreate_derefs(b, &intr->src[0], dst_split->split_vars[out_idx]);
nir_deref_instr *src_deref = recreate_derefs(b, &intr->src[1], src_split->split_vars[i]);
b->cursor = nir_before_instr(instr);
@ -511,22 +511,22 @@ split_cmat_muladd(nir_builder *b,
if (result_split) {
assert(c_split);
m_splits = result_split->box.num_row_splits;
n_splits = result_split->box.num_col_splits;
m_splits = result_split->box.outer_rows;
n_splits = result_split->box.outer_cols;
assert(c_split->box.num_row_splits == m_splits);
assert(c_split->box.num_col_splits == n_splits);
assert(c_split->box.outer_rows == m_splits);
assert(c_split->box.outer_cols == n_splits);
if (a_split)
assert(a_split->box.num_row_splits == m_splits);
assert(a_split->box.outer_rows == m_splits);
if (b_split)
assert(b_split->box.num_col_splits == n_splits);
assert(b_split->box.outer_cols == n_splits);
}
if (a_split && a_split->box.num_col_splits > 1) {
if (a_split && a_split->box.outer_cols > 1) {
assert(b_split);
assert(b_split->box.num_row_splits == a_split->box.num_col_splits);
k_splits = a_split->box.num_col_splits;
assert(b_split->box.outer_rows == a_split->box.outer_cols);
k_splits = a_split->box.outer_cols;
}
for (unsigned m = 0; m < m_splits; m++) {
@ -641,14 +641,14 @@ split_cmat_call_reduce(nir_builder *b,
call_reduce_finish(b, call, reduce, &temp_derefs[0]->def, &temp_derefs[0]->def, &second_deref->def);
}
} else if (reduce & NIR_CMAT_REDUCE_ROW) {
for (unsigned i = 1; i < src_split->box.num_col_splits; i++) {
for (unsigned i = 1; i < src_split->box.outer_cols; i++) {
nir_deref_instr *second_deref = temp_derefs[i];
b->cursor = nir_before_instr(instr);
call_reduce_finish(b, call, reduce, &temp_derefs[0]->def, &temp_derefs[0]->def, &second_deref->def);
}
} else if (reduce & NIR_CMAT_REDUCE_COLUMN) {
for (unsigned i = 1; i < src_split->box.num_row_splits; i++) {
nir_deref_instr *second_deref = temp_derefs[i * src_split->box.num_col_splits];
for (unsigned i = 1; i < src_split->box.outer_rows; i++) {
nir_deref_instr *second_deref = temp_derefs[i * src_split->box.outer_cols];
b->cursor = nir_before_instr(instr);
call_reduce_finish(b, call, reduce, &temp_derefs[0]->def, &temp_derefs[0]->def, &second_deref->def);
}
@ -660,16 +660,16 @@ split_cmat_call_reduce(nir_builder *b,
/* at this point temp_derefs should contain all the split reduced src matrices
now to store them */
if (dst_split) {
for (unsigned r = 0; r < dst_split->box.num_row_splits; r++) {
for (unsigned c = 0; c < dst_split->box.num_col_splits; c++) {
int didx = r * dst_split->box.num_col_splits + c;
for (unsigned r = 0; r < dst_split->box.outer_rows; r++) {
for (unsigned c = 0; c < dst_split->box.outer_cols; c++) {
int didx = r * dst_split->box.outer_cols + c;
int idx;
if ((reduce & (NIR_CMAT_REDUCE_ROW | NIR_CMAT_REDUCE_COLUMN)) == (NIR_CMAT_REDUCE_ROW | NIR_CMAT_REDUCE_COLUMN))
idx = 0;
else if (reduce & NIR_CMAT_REDUCE_ROW)
idx = r % (src_split ? src_split->box.num_row_splits : 1);
idx = r % (src_split ? src_split->box.outer_rows : 1);
else if (reduce & NIR_CMAT_REDUCE_COLUMN)
idx = c % (src_split ? src_split->box.num_col_splits : 1);
idx = c % (src_split ? src_split->box.outer_cols : 1);
else
UNREACHABLE("Unknown NIR_CMAT_REDUCE_*");
@ -691,8 +691,8 @@ split_cmat_call_reduce(nir_builder *b,
int rows = 1, cols = 1;
if (dst_split) {
rows = dst_split->box.num_row_splits;
cols = dst_split->box.num_col_splits;
rows = dst_split->box.outer_rows;
cols = dst_split->box.outer_cols;
}
for (unsigned r = 0; r < rows; r++) {
@ -700,8 +700,8 @@ split_cmat_call_reduce(nir_builder *b,
int d_idx = c + r * cols;
int src_top_left_col = c * 2;
int src_top_left_row = r * 2;
int src_top_idx = src_top_left_col + src_top_left_row * src_split->box.num_col_splits;
int src_bottom_idx = src_top_left_col + (src_top_left_row + 1) * src_split->box.num_col_splits;
int src_top_idx = src_top_left_col + src_top_left_row * src_split->box.outer_cols;
int src_bottom_idx = src_top_left_col + (src_top_left_row + 1) * src_split->box.outer_cols;
nir_deref_instr *src0_deref = recreate_derefs(b, &call->params[1], src_split->split_vars[src_top_idx]);
nir_deref_instr *src1_deref = recreate_derefs(b, &call->params[1], src_split->split_vars[src_top_idx + 1]);
nir_deref_instr *src2_deref = recreate_derefs(b, &call->params[1], src_split->split_vars[src_bottom_idx]);
@ -749,8 +749,8 @@ split_cmat_load_store(nir_builder *b,
struct glsl_cmat_description desc = *glsl_get_cmat_description(split->split_vars[i]->type);
unsigned row_offset, col_offset;
row_offset = (i % split->box.num_col_splits) * desc.cols;
col_offset = (i / split->box.num_col_splits) * desc.rows;
row_offset = (i % split->box.outer_cols) * desc.cols;
col_offset = (i / split->box.outer_cols) * desc.rows;
if (layout == GLSL_MATRIX_LAYOUT_ROW_MAJOR)
SWAP(row_offset, col_offset);
@ -793,8 +793,8 @@ split_cmat_call_per_element_op(nir_builder *b,
return false;
for (unsigned i = 0; i < splits; i++) {
int r = i / dst_split->box.num_col_splits;
int c = i % dst_split->box.num_col_splits;
int r = i / dst_split->box.outer_cols;
int c = i % dst_split->box.outer_cols;
nir_deref_instr *dst_deref = recreate_derefs(b, &call->params[0], dst_split->split_vars[i]);
nir_deref_instr *src_deref = recreate_derefs(b, &call->params[3], src_split->split_vars[i]);
struct glsl_cmat_description cmat_desc = *glsl_get_cmat_description(src_split->split_vars[0]->type);