Mesa (main): panvk: So more nir_lower_tex before descriptor lowering

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 4050697a8f4620109388c02cd8e7437f21e12dc4
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=4050697a8f4620109388c02cd8e7437f21e12dc4

Author: Jason Ekstrand <jason.ekstrand at collabora.com>
Date:   Thu May 12 09:05:34 2022 -0500

panvk: So more nir_lower_tex before descriptor lowering

Some texture lowering generates more txs which means it needs to happen
before we lower descriptors because descriptor lowering is where txs is
actually handled in panvk.

Reviewed-by: Boris Brezillon <boris.brezillon at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16483>

---

 src/panfrost/ci/deqp-panfrost-g52-vk.toml |  1 +
 src/panfrost/vulkan/panvk_vX_shader.c     | 29 ++++++++++++++++++++++++++---
 2 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/src/panfrost/ci/deqp-panfrost-g52-vk.toml b/src/panfrost/ci/deqp-panfrost-g52-vk.toml
index c0044b140ce..ea7c50bfb6b 100644
--- a/src/panfrost/ci/deqp-panfrost-g52-vk.toml
+++ b/src/panfrost/ci/deqp-panfrost-g52-vk.toml
@@ -45,4 +45,5 @@ include = [
     "dEQP-VK.spirv_assembly.instruction.compute.shader_default_output.*",
     "dEQP-VK.spirv_assembly.instruction.compute.workgroup_memory.*",
     "dEQP-VK.ssbo.layout.single_basic_type.*",
+    "dEQP-VK.texture.explicit_lod.*.derivatives.*",
 ]
diff --git a/src/panfrost/vulkan/panvk_vX_shader.c b/src/panfrost/vulkan/panvk_vX_shader.c
index 8b6e1f5dda8..203c884c44a 100644
--- a/src/panfrost/vulkan/panvk_vX_shader.c
+++ b/src/panfrost/vulkan/panvk_vX_shader.c
@@ -396,12 +396,35 @@ panvk_per_arch(shader_create)(struct panvk_device *dev,
       panvk_lower_blend(pdev, nir, &inputs, blend_state, static_blend_constants);
    }
 
-   /* We need to lower nir_texop_txs with LOD before we lower descriptor
-    * access because nir_texop_txs gets turned into a descriptor UBO read
-    * and a bit of math by the descriptor lowering code.
+   /* Do texture lowering here.  Yes, it's a duplication of the texture
+    * lowering in bifrost_compile.  However, we need to lower texture stuff
+    * now, before we call panvk_per_arch(nir_lower_descriptors)() because some
+    * of the texture lowering generates nir_texop_txs which we handle as part
+    * of descriptor lowering.
+    *
+    * TODO: We really should be doing this in common code, not dpulicated in
+    * panvk.  In order to do that, we need to rework the panfrost compile
+    * flow to look more like the Intel flow:
+    *
+    *  1. Compile SPIR-V to NIR and maybe do a tiny bit of lowering that needs
+    *     to be done really early.
+    *
+    *  2. bi_preprocess_nir: Does common lowering and runs the optimization
+    *     loop.  Nothing here should be API-specific.
+    *
+    *  3. Do additional lowering in panvk
+    *
+    *  4. bi_postprocess_nir: Does final lowering and runs the optimization
+    *     loop again.  This can happen as part of the final compile.
+    *
+    * This would give us a better place to do panvk-specific lowering.
     */
    nir_lower_tex_options lower_tex_options = {
       .lower_txs_lod = true,
+      .lower_txp = ~0,
+      .lower_tg4_broadcom_swizzle = true,
+      .lower_txd = true,
+      .lower_invalid_implicit_lod = true,
    };
    NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);