pan/midgard: Break out one-src read_components

For constant packing, this is interesting to break down further.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3653>
This commit is contained in:
Alyssa Rosenzweig 2020-01-31 08:11:13 -05:00 committed by Marge Bot
parent b74212e701
commit a12fe52cbc
2 changed files with 33 additions and 25 deletions

View file

@ -530,6 +530,7 @@ bool mir_special_index(compiler_context *ctx, unsigned idx);
unsigned mir_use_count(compiler_context *ctx, unsigned value);
bool mir_is_written_before(compiler_context *ctx, midgard_instruction *ins, unsigned node);
uint16_t mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node);
uint16_t mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i);
midgard_reg_mode mir_typesize(midgard_instruction *ins);
midgard_reg_mode mir_srcsize(midgard_instruction *ins, unsigned i);
unsigned mir_bytes_for_mode(midgard_reg_mode mode);

View file

@ -465,6 +465,37 @@ mir_bytemask_of_read_components_single(unsigned *swizzle, unsigned inmask, midga
return mir_to_bytemask(mode, cmask);
}
uint16_t
mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i)
{
uint16_t mask = 0;
/* Branch writeout uses all components */
if (ins->compact_branch && ins->writeout && (i == 0))
return 0xFFFF;
/* Conditional branches read one 32-bit component = 4 bytes (TODO: multi branch??) */
if (ins->compact_branch && ins->branch.conditional && (i == 0))
return 0xF;
/* ALU ops act componentwise so we need to pay attention to
* their mask. Texture/ldst does not so we don't clamp source
* readmasks based on the writemask */
unsigned qmask = (ins->type == TAG_ALU_4) ? ins->mask : ~0;
/* Handle dot products and things */
if (ins->type == TAG_ALU_4 && !ins->compact_branch) {
unsigned props = alu_opcode_props[ins->alu.op].props;
unsigned channel_override = GET_CHANNEL_COUNT(props);
if (channel_override)
qmask = mask_of(channel_override);
}
return mir_bytemask_of_read_components_single(ins->swizzle[i], qmask, mir_srcsize(ins, i));
}
uint16_t
mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node)
{
@ -475,31 +506,7 @@ mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node)
mir_foreach_src(ins, i) {
if (ins->src[i] != node) continue;
/* Branch writeout uses all components */
if (ins->compact_branch && ins->writeout && (i == 0))
return 0xFFFF;
/* Conditional branches read one 32-bit component = 4 bytes (TODO: multi branch??) */
if (ins->compact_branch && ins->branch.conditional && (i == 0))
return 0xF;
/* ALU ops act componentwise so we need to pay attention to
* their mask. Texture/ldst does not so we don't clamp source
* readmasks based on the writemask */
unsigned qmask = (ins->type == TAG_ALU_4) ? ins->mask : ~0;
/* Handle dot products and things */
if (ins->type == TAG_ALU_4 && !ins->compact_branch) {
unsigned props = alu_opcode_props[ins->alu.op].props;
unsigned channel_override = GET_CHANNEL_COUNT(props);
if (channel_override)
qmask = mask_of(channel_override);
}
mask |= mir_bytemask_of_read_components_single(ins->swizzle[i], qmask, mir_srcsize(ins, i));
mask |= mir_bytemask_of_read_components_index(ins, i);
}
return mask;