Mesa (main): pan/mdg: Delete dedicated fdot2 lowering

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 8b70e7491a0adfcf201db69f3e83ca61b554b6c8
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=8b70e7491a0adfcf201db69f3e83ca61b554b6c8

Author: Alyssa Rosenzweig <alyssa at collabora.com>
Date:   Fri Feb  4 18:13:55 2022 -0500

pan/mdg: Delete dedicated fdot2 lowering

It's just lower_alu_to_scalar

total instructions in shared programs: 72542 -> 72528 (-0.02%)
instructions in affected programs: 673 -> 659 (-2.08%)
helped: 4
HURT: 1
helped stats (abs) min: 1.0 max: 11.0 x̄: 3.75 x̃: 1
helped stats (rel) min: 0.28% max: 6.79% x̄: 3.07% x̃: 2.60%
HURT stats (abs)   min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 3.03% max: 3.03% x̄: 3.03% x̃: 3.03%
95% mean confidence interval for instructions value: -8.65 3.05
95% mean confidence interval for instructions %-change: -6.32% 2.62%
Inconclusive result (value mean confidence interval includes 0).

total bundles in shared programs: 32051 -> 32036 (-0.05%)
bundles in affected programs: 207 -> 192 (-7.25%)
helped: 3
HURT: 0
helped stats (abs) min: 1.0 max: 10.0 x̄: 5.00 x̃: 4
helped stats (rel) min: 3.28% max: 13.89% x̄: 8.29% x̃: 7.69%

total quadwords in shared programs: 56496 -> 56487 (-0.02%)
quadwords in affected programs: 422 -> 413 (-2.13%)
helped: 2
HURT: 0

total registers in shared programs: 5106 -> 5104 (-0.04%)
registers in affected programs: 8 -> 6 (-25.00%)
helped: 1
HURT: 0

Signed-off-by: Alyssa Rosenzweig <alyssa at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14888>

---

 src/panfrost/midgard/midgard_compile.c | 45 +++-------------------------------
 1 file changed, 4 insertions(+), 41 deletions(-)

diff --git a/src/panfrost/midgard/midgard_compile.c b/src/panfrost/midgard/midgard_compile.c
index a8abdd98e0f..a255044e674 100644
--- a/src/panfrost/midgard/midgard_compile.c
+++ b/src/panfrost/midgard/midgard_compile.c
@@ -214,43 +214,6 @@ glsl_type_size(const struct glsl_type *type, bool bindless)
         return glsl_count_attribute_slots(type, false);
 }
 
-/* Lower fdot2 to a vector multiplication followed by channel addition  */
-static bool
-midgard_nir_lower_fdot2_instr(nir_builder *b, nir_instr *instr, void *data)
-{
-        if (instr->type != nir_instr_type_alu)
-                return false;
-
-        nir_alu_instr *alu = nir_instr_as_alu(instr);
-        if (alu->op != nir_op_fdot2)
-                return false;
-
-        b->cursor = nir_before_instr(&alu->instr);
-
-        nir_ssa_def *src0 = nir_ssa_for_alu_src(b, alu, 0);
-        nir_ssa_def *src1 = nir_ssa_for_alu_src(b, alu, 1);
-
-        nir_ssa_def *product = nir_fmul(b, src0, src1);
-
-        nir_ssa_def *sum = nir_fadd(b,
-                                    nir_channel(b, product, 0),
-                                    nir_channel(b, product, 1));
-
-        /* Replace the fdot2 with this sum */
-        nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, sum);
-
-        return true;
-}
-
-static bool
-midgard_nir_lower_fdot2(nir_shader *shader)
-{
-        return nir_shader_instructions_pass(shader,
-                                            midgard_nir_lower_fdot2_instr,
-                                            nir_metadata_block_index | nir_metadata_dominance,
-                                            NULL);
-}
-
 static bool
 midgard_nir_lower_global_load_instr(nir_builder *b, nir_instr *instr, void *data)
 {
@@ -320,7 +283,7 @@ midgard_nir_lower_global_load(nir_shader *shader)
 }
 
 static bool
-mdg_is_64(const nir_instr *instr, const void *_unused)
+mdg_should_scalarize(const nir_instr *instr, const void *_unused)
 {
         const nir_alu_instr *alu = nir_instr_as_alu(instr);
 
@@ -328,6 +291,7 @@ mdg_is_64(const nir_instr *instr, const void *_unused)
                 return true;
 
         switch (alu->op) {
+        case nir_op_fdot2:
         case nir_op_umul_high:
         case nir_op_imul_high:
                 return true;
@@ -384,8 +348,6 @@ optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
 
         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 
-        /* Must lower fdot2 after tex is lowered */
-        NIR_PASS(progress, nir, midgard_nir_lower_fdot2);
 
         /* T720 is broken. */
 
@@ -399,6 +361,7 @@ optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
         NIR_PASS(progress, nir, pan_lower_sample_pos);
 
         NIR_PASS(progress, nir, midgard_nir_lower_algebraic_early);
+        NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
 
         do {
                 progress = false;
@@ -443,7 +406,7 @@ optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
                          midgard_vectorize_filter, NULL);
         } while (progress);
 
-        NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_is_64, NULL);
+        NIR_PASS_V(nir, nir_lower_alu_to_scalar, mdg_should_scalarize, NULL);
 
         /* Run after opts so it can hit more */
         if (!is_blend)