[Mesa-dev] [PATCH 4/4] panfrost/midgard: Vectorize I/O

[Alyssa Rosenzweig]

This uses the new mesa/st functionality for NIR I/O vectorization, which
eliminates a number of corner cases (resulting in assorted dEQP
failures and regressions) and should improve performance substantial due
to lessened pressure on the load/store pipe.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig at collabora.com>
---
 .../drivers/panfrost/ci/expected-failures.txt |  3 ---
 .../panfrost/midgard/midgard_compile.c        | 20 +++++++++++++++----
 .../panfrost/midgard/midgard_compile.h        |  2 ++
 3 files changed, 18 insertions(+), 7 deletions(-)

diff --git a/src/gallium/drivers/panfrost/ci/expected-failures.txt b/src/gallium/drivers/panfrost/ci/expected-failures.txt
index 1f4e95f25ce..12f3c48047b 100644
--- a/src/gallium/drivers/panfrost/ci/expected-failures.txt
+++ b/src/gallium/drivers/panfrost/ci/expected-failures.txt
@@ -304,9 +304,6 @@ dEQP-GLES2.functional.polygon_offset.fixed16_render_with_units
 dEQP-GLES2.functional.polygon_offset.fixed16_result_depth_clamp
 dEQP-GLES2.functional.rasterization.limits.points
 dEQP-GLES2.functional.shaders.builtin_variable.fragcoord_w
-dEQP-GLES2.functional.shaders.conversions.scalar_to_matrix.bool_to_mat3_vertex
-dEQP-GLES2.functional.shaders.conversions.scalar_to_matrix.float_to_mat3_vertex
-dEQP-GLES2.functional.shaders.conversions.scalar_to_matrix.int_to_mat3_vertex
 dEQP-GLES2.functional.shaders.functions.control_flow.return_after_continue_vertex
 dEQP-GLES2.functional.shaders.functions.control_flow.return_after_loop_sequence_vertex
 dEQP-GLES2.functional.shaders.functions.control_flow.return_in_loop_if_fragment
diff --git a/src/gallium/drivers/panfrost/midgard/midgard_compile.c b/src/gallium/drivers/panfrost/midgard/midgard_compile.c
index f1f38fee5d3..92bfe51ce85 100644
--- a/src/gallium/drivers/panfrost/midgard/midgard_compile.c
+++ b/src/gallium/drivers/panfrost/midgard/midgard_compile.c
@@ -1238,6 +1238,12 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
                         midgard_instruction ins = v_fmov(reg, blank_alu_src, SSA_FIXED_REGISTER(26));
                         emit_mir_instruction(ctx, ins);
 
+                        /* We should have been vectorized. That also lets us
+                         * ignore the mask. because the mask component on
+                         * st_vary is (as far as I can tell) ignored [the blob
+                         * sets it to zero] */
+                        assert(nir_intrinsic_component(instr) == 0);
+
                         midgard_instruction st = m_st_vary_32(SSA_FIXED_REGISTER(0), offset);
                         st.load_store.unknown = 0x1E9E; /* XXX: What is this? */
                         emit_mir_instruction(ctx, st);
@@ -2314,16 +2320,22 @@ midgard_compile_shader_nir(nir_shader *nir, midgard_program *program, bool is_bl
         struct exec_list *varyings =
                 ctx->stage == MESA_SHADER_VERTEX ? &nir->outputs : &nir->inputs;
 
+        unsigned max_varying = 0;
         nir_foreach_variable(var, varyings) {
                 unsigned loc = var->data.driver_location;
                 unsigned sz = glsl_type_size(var->type, FALSE);
 
-                for (int c = 0; c < sz; ++c) {
-                        program->varyings[loc + c] = var->data.location;
+                for (int c = loc; c < (loc + sz); ++c) {
+                        program->varyings[c] = var->data.location;
+                        max_varying = MAX2(max_varying, c);
                 }
         }
 
-        /* Lower gl_Position pre-optimisation */
+        /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
+         * (so we don't accidentally duplicate the epilogue since mesa/st has
+         * messed with our I/O quite a bit already) */
+
+        NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 
         if (ctx->stage == MESA_SHADER_VERTEX)
                 NIR_PASS_V(nir, nir_lower_viewport_transform);
@@ -2356,7 +2368,7 @@ midgard_compile_shader_nir(nir_shader *nir, midgard_program *program, bool is_bl
         memcpy(program->sysvals, ctx->sysvals, sizeof(ctx->sysvals[0]) * ctx->sysval_count);
 
         program->attribute_count = (ctx->stage == MESA_SHADER_VERTEX) ? nir->num_inputs : 0;
-        program->varying_count = (ctx->stage == MESA_SHADER_VERTEX) ? nir->num_outputs : ((ctx->stage == MESA_SHADER_FRAGMENT) ? nir->num_inputs : 0);
+        program->varying_count = max_varying + 1; /* Fencepost off-by-one */
 
         nir_foreach_function(func, nir) {
                 if (!func->impl)
diff --git a/src/gallium/drivers/panfrost/midgard/midgard_compile.h b/src/gallium/drivers/panfrost/midgard/midgard_compile.h
index 2ee0dfbb1ae..b21f054c031 100644
--- a/src/gallium/drivers/panfrost/midgard/midgard_compile.h
+++ b/src/gallium/drivers/panfrost/midgard/midgard_compile.h
@@ -109,6 +109,8 @@ static const nir_shader_compiler_options midgard_nir_options = {
         .lower_extract_word = true,
 
         .lower_doubles_options = nir_lower_dmod,
+
+        .vectorize_io = true,
 };
 
 #endif
-- 
2.20.1