mirrors/linux
1
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git synced 2026-02-16 17:45:01 -05:00
Code Issues Packages Projects Releases Wiki Activity

drm/xe: Waste fewer instructions in emit_wa_job()

Browse Source 
I was debugging some unrelated issue and noticed the current code was
very verbose. We can improve it easily by using the more common batch
buffer building pattern.

Before:
                bb->cs[bb->len++] = MI_LOAD_REGISTER_REG | MI_LRR_DST_CS_MMIO;
     c4d:       41 8b 56 10             mov    0x10(%r14),%edx
     c51:       49 8b 4e 08             mov    0x8(%r14),%rcx
     c55:       8d 72 01                lea    0x1(%rdx),%esi
     c58:       41 89 76 10             mov    %esi,0x10(%r14)
     c5c:       c7 04 91 01 00 08 15    movl   $0x15080001,(%rcx,%rdx,4)
                        bb->cs[bb->len++] = entry->reg.addr;
     c63:       8b 08                   mov    (%rax),%ecx
     c65:       41 8b 56 10             mov    0x10(%r14),%edx
     c69:       49 8b 76 08             mov    0x8(%r14),%rsi
     c6d:       81 e1 ff ff 3f 00       and    $0x3fffff,%ecx
     c73:       8d 7a 01                lea    0x1(%rdx),%edi
     c76:       41 89 7e 10             mov    %edi,0x10(%r14)
     c7a:       89 0c 96                mov    %ecx,(%rsi,%rdx,4)
 ..etc..

After:
                *cs++ = MI_LOAD_REGISTER_REG | MI_LRR_DST_CS_MMIO;
     c52:       41 c7 04 24 01 00 08    movl   $0x15080001,(%r12)
     c59:       15
                        *cs++ = entry->reg.addr;
     c5a:       8b 10                   mov    (%rax),%edx
 ..etc..

Resulting in the following binary change:

	add/remove: 0/0 grow/shrink: 0/2 up/down: 0/-348 (-348)
	Function                                     old     new   delta
	xe_gt_record_default_lrcs.cold               304     296      -8
	xe_gt_record_default_lrcs                   2200    1860    -340
	Total: Before=13554, After=13206, chg -2.57%

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20250710-lrc-refactors-v2-7-a5e2ca03f6bd@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
This commit is contained in:
Tvrtko Ursulin
2025-07-10 13:33:52 -07:00
committed by Lucas De Marchi
parent f4b538245f
commit aded26ccaa
3 changed files with 46 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
drivers/gpu/drm/xe/xe_gt.c
View File

@@ -194,6 +194,7 @@ static int emit_wa_job(struct xe_gt *gt, struct xe_exec_queue *q)
unsigned long idx;
struct xe_bb *bb;
size_t bb_len = 0;
u32 *cs;
/* count RMW registers as those will be handled separately */
xa_for_each(&sr->xa, idx, entry) {
@@ -222,13 +223,15 @@ static int emit_wa_job(struct xe_gt *gt, struct xe_exec_queue *q)
if (IS_ERR(bb))
return PTR_ERR(bb);
cs = bb->cs;
if (count) {
/*
* Emit single LRI with all non RMW regs: 1 leading dw + 2dw per
* reg + 1
*/
bb->cs[bb->len++] = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(count);
*cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(count);
xa_for_each(&sr->xa, idx, entry) {
struct xe_reg reg = entry->reg;
@@ -243,8 +246,8 @@ static int emit_wa_job(struct xe_gt *gt, struct xe_exec_queue *q)
val |= entry->set_bits;
bb->cs[bb->len++] = reg.addr;
bb->cs[bb->len++] = val;
*cs++ = reg.addr;
*cs++ = val;
xe_gt_dbg(gt, "REG[0x%x] = 0x%08x", reg.addr, val);
}
}
@@ -256,46 +259,49 @@ static int emit_wa_job(struct xe_gt *gt, struct xe_exec_queue *q)
if (entry->reg.masked || entry->clr_bits == ~0)
continue;
bb->cs[bb->len++] = MI_LOAD_REGISTER_REG | MI_LRR_DST_CS_MMIO;
bb->cs[bb->len++] = entry->reg.addr;
bb->cs[bb->len++] = CS_GPR_REG(0, 0).addr;
*cs++ = MI_LOAD_REGISTER_REG | MI_LRR_DST_CS_MMIO;
*cs++ = entry->reg.addr;
*cs++ = CS_GPR_REG(0, 0).addr;
bb->cs[bb->len++] = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(2) |
MI_LRI_LRM_CS_MMIO;
bb->cs[bb->len++] = CS_GPR_REG(0, 1).addr;
bb->cs[bb->len++] = entry->clr_bits;
bb->cs[bb->len++] = CS_GPR_REG(0, 2).addr;
bb->cs[bb->len++] = entry->set_bits;
*cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(2) |
MI_LRI_LRM_CS_MMIO;
*cs++ = CS_GPR_REG(0, 1).addr;
*cs++ = entry->clr_bits;
*cs++ = CS_GPR_REG(0, 2).addr;
*cs++ = entry->set_bits;
bb->cs[bb->len++] = MI_MATH(8);
bb->cs[bb->len++] = CS_ALU_INSTR_LOAD(SRCA, REG0);
bb->cs[bb->len++] = CS_ALU_INSTR_LOADINV(SRCB, REG1);
bb->cs[bb->len++] = CS_ALU_INSTR_AND;
bb->cs[bb->len++] = CS_ALU_INSTR_STORE(REG0, ACCU);
bb->cs[bb->len++] = CS_ALU_INSTR_LOAD(SRCA, REG0);
bb->cs[bb->len++] = CS_ALU_INSTR_LOAD(SRCB, REG2);
bb->cs[bb->len++] = CS_ALU_INSTR_OR;
bb->cs[bb->len++] = CS_ALU_INSTR_STORE(REG0, ACCU);
*cs++ = MI_MATH(8);
*cs++ = CS_ALU_INSTR_LOAD(SRCA, REG0);
*cs++ = CS_ALU_INSTR_LOADINV(SRCB, REG1);
*cs++ = CS_ALU_INSTR_AND;
*cs++ = CS_ALU_INSTR_STORE(REG0, ACCU);
*cs++ = CS_ALU_INSTR_LOAD(SRCA, REG0);
*cs++ = CS_ALU_INSTR_LOAD(SRCB, REG2);
*cs++ = CS_ALU_INSTR_OR;
*cs++ = CS_ALU_INSTR_STORE(REG0, ACCU);
bb->cs[bb->len++] = MI_LOAD_REGISTER_REG | MI_LRR_SRC_CS_MMIO;
bb->cs[bb->len++] = CS_GPR_REG(0, 0).addr;
bb->cs[bb->len++] = entry->reg.addr;
*cs++ = MI_LOAD_REGISTER_REG | MI_LRR_SRC_CS_MMIO;
*cs++ = CS_GPR_REG(0, 0).addr;
*cs++ = entry->reg.addr;
xe_gt_dbg(gt, "REG[%#x] = ~%#x|%#x\n",
entry->reg.addr, entry->clr_bits, entry->set_bits);
}
/* reset used GPR */
bb->cs[bb->len++] = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(3) | MI_LRI_LRM_CS_MMIO;
bb->cs[bb->len++] = CS_GPR_REG(0, 0).addr;
bb->cs[bb->len++] = 0;
bb->cs[bb->len++] = CS_GPR_REG(0, 1).addr;
bb->cs[bb->len++] = 0;
bb->cs[bb->len++] = CS_GPR_REG(0, 2).addr;
bb->cs[bb->len++] = 0;
*cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(3) |
MI_LRI_LRM_CS_MMIO;
*cs++ = CS_GPR_REG(0, 0).addr;
*cs++ = 0;
*cs++ = CS_GPR_REG(0, 1).addr;
*cs++ = 0;
*cs++ = CS_GPR_REG(0, 2).addr;
*cs++ = 0;
}
xe_lrc_emit_hwe_state_instructions(q, bb);
cs = xe_lrc_emit_hwe_state_instructions(q, cs);
bb->len = cs - bb->cs;
ret = emit_job_sync(q, bb, HZ);

12
drivers/gpu/drm/xe/xe_lrc.c
View File

@@ -1888,7 +1888,7 @@ static const struct instr_state xe_hpg_svg_state[] = {
{ .instr = CMD_3DSTATE_DRAWING_RECTANGLE, .num_dw = 4 },
};
void xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, struct xe_bb *bb)
u32 *xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, u32 *cs)
{
struct xe_gt *gt = q->hwe->gt;
struct xe_device *xe = gt_to_xe(gt);
@@ -1923,7 +1923,7 @@ void xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, struct xe_bb *b
if (!state_table) {
xe_gt_dbg(gt, "No non-register state to emit on graphics ver %d.%02d\n",
GRAPHICS_VER(xe), GRAPHICS_VERx100(xe) % 100);
return;
return cs;
}
for (int i = 0; i < state_table_size; i++) {
@@ -1946,12 +1946,14 @@ void xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, struct xe_bb *b
instr == CMD_3DSTATE_DRAWING_RECTANGLE)
instr = CMD_3DSTATE_DRAWING_RECTANGLE_FAST;
bb->cs[bb->len] = instr;
*cs = instr;
if (!is_single_dw)
bb->cs[bb->len] |= (num_dw - 2);
*cs |= (num_dw - 2);
bb->len += num_dw;
cs += num_dw;
}
return cs;
}
struct xe_lrc_snapshot *xe_lrc_snapshot_capture(struct xe_lrc *lrc)

2
drivers/gpu/drm/xe/xe_lrc.h
View File

@@ -112,7 +112,7 @@ void xe_lrc_dump_default(struct drm_printer *p,
struct xe_gt *gt,
enum xe_engine_class);
void xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, struct xe_bb *bb);
u32 *xe_lrc_emit_hwe_state_instructions(struct xe_exec_queue *q, u32 *cs);
struct xe_lrc_snapshot *xe_lrc_snapshot_capture(struct xe_lrc *lrc);
void xe_lrc_snapshot_capture_delayed(struct xe_lrc_snapshot *snapshot);