[Mesa-stable] [Mesa-dev] [PATCH 2/3] i965/fs: Consider cmod when propagating to inst with different type.
Kenneth Graunke
kenneth at whitecape.org
Tue Mar 3 16:15:38 PST 2015
On Friday, February 27, 2015 11:34:06 AM Matt Turner wrote:
> We can safely propagate the conditional mod to an instruction with a
> different type if the conditional mod does not involve comparing for
> equality with zero (or probably NaN, but ignore that for now).
>
> This is because -0.0 and +0.0 are both test equal to zero, but their
> integer representations do not.
>
> Cc: 10.5 <mesa-stable at lists.freedesktop.org>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89317
> ---
> .../drivers/dri/i965/brw_fs_cmod_propagation.cpp | 8 +++
> .../drivers/dri/i965/test_fs_cmod_propagation.cpp | 67 ++++++++++++++++++++++
> 2 files changed, 75 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_cmod_propagation.cpp b/src/mesa/drivers/dri/i965/brw_fs_cmod_propagation.cpp
> index c6384ab..a92eef6 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_cmod_propagation.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_cmod_propagation.cpp
> @@ -80,6 +80,14 @@ opt_cmod_propagation_local(fs_visitor *v, bblock_t *block)
> scan_inst->dst.reg_offset != inst->src[0].reg_offset)
> break;
>
> + /* Testing equality with zero is different for ints and floats */
> + if (scan_inst->dst.type != inst->dst.type &&
> + (inst->conditional_mod == BRW_CONDITIONAL_Z ||
> + inst->conditional_mod == BRW_CONDITIONAL_NZ ||
> + inst->conditional_mod == BRW_CONDITIONAL_GE ||
> + inst->conditional_mod == BRW_CONDITIONAL_LE))
> + break;
> +
> /* If the instruction generating inst's source also wrote the
> * flag, and inst is doing a simple .nz comparison, then inst
> * is redundant - the appropriate value is already in the flag
> diff --git a/src/mesa/drivers/dri/i965/test_fs_cmod_propagation.cpp b/src/mesa/drivers/dri/i965/test_fs_cmod_propagation.cpp
> index fbe4fd9..0287161 100644
> --- a/src/mesa/drivers/dri/i965/test_fs_cmod_propagation.cpp
> +++ b/src/mesa/drivers/dri/i965/test_fs_cmod_propagation.cpp
> @@ -415,3 +415,70 @@ TEST_F(cmod_propagation_test, movnz)
> EXPECT_EQ(BRW_OPCODE_CMP, instruction(block0, 0)->opcode);
> EXPECT_EQ(BRW_CONDITIONAL_GE, instruction(block0, 0)->conditional_mod);
> }
> +
> +TEST_F(cmod_propagation_test, different_types)
> +{
> + fs_reg dest = v->vgrf(glsl_type::int_type);
> + fs_reg src0 = v->vgrf(glsl_type::int_type);
> + fs_reg src1 = v->vgrf(glsl_type::int_type);
> + fs_reg zero(0.0f);
> + v->emit(BRW_OPCODE_ADD, dest, src0, src1);
> + v->emit(BRW_OPCODE_CMP, v->reg_null_f, retype(dest, BRW_REGISTER_TYPE_F),
> + zero)
> + ->conditional_mod = BRW_CONDITIONAL_G;
> +
> + /* = Before =
> + *
> + * 0: add(8) dest:D src0:D src1:D
> + * 1: cmp.g.f0(8) null:F dest:F 0.0f
> + *
> + * = After =
> + * 0: add.g.f0(8) dest:D src0:D src1:D
> + */
This is wrong - you're adding two integers and doing an integer compare
with 0, rather than a float comparison.
I verified that in C, (-0.0 > 0.0) == false. I then modified your
Piglit test to be:
if (uintBitsToFloat(a + b) > 0.0)
frag_color = vec4(1.0, 0.0, 0.0, 1.0);
else
frag_color = vec4(0.0, 1.0, 0.0, 1.0);
(> instead of ==, swap red/green).
This fails with your code, because it ends up doing:
0x80000000 > 0x0
=> 0x80000000 - 0x0 > 0x0.
=> 0x80000000 > 0
=> True
It passes if you eliminate the conditional_mod checks and just do:
if (scan_inst->dst.type != inst->dst.type)
break;
Assuming you make that change, and drop/fix this unit test, you can put
my R-b on it. Good find :)
> +
> + v->calculate_cfg();
> + bblock_t *block0 = v->cfg->blocks[0];
> +
> + EXPECT_EQ(0, block0->start_ip);
> + EXPECT_EQ(1, block0->end_ip);
> +
> + EXPECT_TRUE(cmod_propagation(v));
> + EXPECT_EQ(0, block0->start_ip);
> + EXPECT_EQ(0, block0->end_ip);
> + EXPECT_EQ(BRW_OPCODE_ADD, instruction(block0, 0)->opcode);
> + EXPECT_EQ(BRW_CONDITIONAL_G, instruction(block0, 0)->conditional_mod);
> +}
> +
> +TEST_F(cmod_propagation_test, different_types_cmod_with_zero)
> +{
> + fs_reg dest = v->vgrf(glsl_type::int_type);
> + fs_reg src0 = v->vgrf(glsl_type::int_type);
> + fs_reg src1 = v->vgrf(glsl_type::int_type);
> + fs_reg zero(0.0f);
> + v->emit(BRW_OPCODE_ADD, dest, src0, src1);
> + v->emit(BRW_OPCODE_CMP, v->reg_null_f, retype(dest, BRW_REGISTER_TYPE_F),
> + zero)
> + ->conditional_mod = BRW_CONDITIONAL_GE;
> +
> + /* = Before =
> + *
> + * 0: add(8) dest:D src0:D src1:D
> + * 1: cmp.ge.f0(8) null:F dest:F 0.0f
> + *
> + * = After =
> + * (no changes)
> + */
> +
> + v->calculate_cfg();
> + bblock_t *block0 = v->cfg->blocks[0];
> +
> + EXPECT_EQ(0, block0->start_ip);
> + EXPECT_EQ(1, block0->end_ip);
> +
> + EXPECT_FALSE(cmod_propagation(v));
> + EXPECT_EQ(0, block0->start_ip);
> + EXPECT_EQ(1, block0->end_ip);
> + EXPECT_EQ(BRW_OPCODE_ADD, instruction(block0, 0)->opcode);
> + EXPECT_EQ(BRW_OPCODE_CMP, instruction(block0, 1)->opcode);
> + EXPECT_EQ(BRW_CONDITIONAL_GE, instruction(block0, 1)->conditional_mod);
> +}
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-stable/attachments/20150303/6bcbf65b/attachment.sig>
More information about the mesa-stable
mailing list