panvk: Add support for moving constants to the FAU

This commit sets up support for constants that should be pushed into
FAU entries after the ones we already push (sysvals, push constants).

The next commit will make the compiler actually find and report the
constants that should be pushed.

Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36872>

src/panfrost/util/pan_ir.h

@@ -120,6 +120,15 @@ struct pan_compile_inputs {
     */
    uint32_t fixed_varying_mask;
 
+   /* Settings to move constants into the FAU. */
+   struct {
+      uint32_t *values;
+      /* In multiples of 32bit. */
+      uint32_t max_amount;
+      /* In multiples of 32bit. */
+      uint32_t offset;
+   } fau_consts;
+
    union {
       struct {
          uint32_t rt_conv[8];
@@ -204,6 +213,9 @@ struct pan_shader_info {
    /* Bit mask of preloaded registers */
    uint64_t preload;
 
+   uint32_t fau_consts_count;
+   uint32_t fau_consts[128];
+
    union {
       struct {
          bool reads_frag_coord;

src/panfrost/vulkan/csf/panvk_vX_cmd_draw.c

@@ -2567,8 +2567,7 @@ panvk_cmd_draw_indirect(struct panvk_cmd_buffer *cmdbuf,
    struct cs_index draw_id = cs_scratch_reg32(b, 7);
    struct cs_index vs_fau_addr = cs_scratch_reg64(b, 8);
    struct cs_index tracing_scratch_regs = cs_scratch_reg_tuple(b, 10, 4);
-   uint32_t vs_fau_count = BITSET_COUNT(vs->fau.used_sysvals) +
-                           BITSET_COUNT(vs->fau.used_push_consts);
+   uint32_t vs_fau_count = vs->fau.total_count;
 
    if (draw->indirect.count_buffer_dev_addr) {
       cs_move32_to(b, max_draw_count, draw->indirect.draw_count);

src/panfrost/vulkan/panvk_vX_cmd_push_constant.c

@@ -70,6 +70,11 @@ panvk_per_arch(cmd_prepare_push_uniforms)(
 
       BITSET_FOREACH_SET(w, shader->fau.used_push_consts, MAX_PUSH_CONST_FAUS)
          faus[fau++] = push_consts[w];
+
+      for (uint32_t i = 0; i < shader->info.fau_consts_count; i += 2) {
+         faus[fau++] = (uint64_t)shader->info.fau_consts[i + 1] << 32 |
+                       shader->info.fau_consts[i];
+      }
    }
 
    *push_ptr = push_uniforms.gpu;

src/panfrost/vulkan/panvk_vX_shader.c

@@ -1,4 +1,5 @@
 /*
+ * Copyright © 2025 Arm Ltd.
  * Copyright © 2021 Collabora Ltd.
  *
  * Derived from tu_shader.c which is:
@@ -62,6 +63,8 @@
 #include "vk_shader.h"
 #include "vk_util.h"
 
+#define FAU_WORD_COUNT 64
+
 struct panvk_lower_sysvals_context {
    struct panvk_shader_variant *shader;
    const struct vk_graphics_pipeline_state *state;
@@ -710,10 +713,11 @@ lower_load_push_consts(nir_shader *nir, struct panvk_shader_variant *shader)
     * needed in the blend shader. */
    shader->fau.sysval_count = BITSET_COUNT(shader->fau.used_sysvals);
    /* 32 FAUs (256 bytes) are reserved for API push constants */
-   assert(shader->fau.sysval_count <= 64 - 32 && "too many sysval FAUs");
+   assert(shader->fau.sysval_count <= FAU_WORD_COUNT - 32 &&
+          "too many sysval FAUs");
    shader->fau.total_count =
       shader->fau.sysval_count + BITSET_COUNT(shader->fau.used_push_consts);
-   assert(shader->fau.total_count <= 64 &&
+   assert(shader->fau.total_count <= FAU_WORD_COUNT &&
           "asking for more FAUs than the hardware has to offer");
 
    if (!progress)
@@ -977,6 +981,11 @@ panvk_compile_nir(struct panvk_device *dev, nir_shader *nir,
    util_dynarray_init(&binary, NULL);
    pan_shader_compile(nir, compile_input, &binary, &shader->info);
 
+   /* Propagate potential additional FAU values into the panvk info struct. */
+   /* FAU consts are pushed as 32bit values, but total_count is for 64bit
+    * ones. */
+   shader->fau.total_count += DIV_ROUND_UP(shader->info.fau_consts_count, 2);
+
    void *bin_ptr = util_dynarray_element(&binary, uint8_t, 0);
    unsigned bin_size = util_dynarray_num_elements(&binary, uint8_t);
 
@@ -1355,6 +1364,13 @@ panvk_compile_shader(struct panvk_device *dev,
                          info->robustness, noperspective_varyings, state,
                          &inputs, variant);
 
+         /* Allow the remaining FAU space to be filled with constants. */
+         input_variants[v].fau_consts.max_amount =
+            2 * (FAU_WORD_COUNT - variant->fau.total_count);
+         input_variants[v].fau_consts.offset = variant->fau.total_count * 2;
+         input_variants[v].fau_consts.values = &variant->info.fau_consts[0];
+         assert(input_variants[v].fau_consts.max_amount <= ARRAY_SIZE(variant->info.fau_consts));
+
          variant->own_bin = true;
 
          result = panvk_compile_nir(dev, nir_variants[v], info->flags,
@@ -1391,6 +1407,12 @@ panvk_compile_shader(struct panvk_device *dev,
          inputs.valhall.use_ld_var_buf = panvk_use_ld_var_buf(variant);
 #endif
 
+      /* Allow the remaining FAU space to be filled with constants. */
+      inputs.fau_consts.max_amount = 2 * (FAU_WORD_COUNT - variant->fau.total_count);
+      inputs.fau_consts.offset = variant->fau.total_count * 2;
+      inputs.fau_consts.values = &variant->info.fau_consts[0];
+      assert(inputs.fau_consts.max_amount <= ARRAY_SIZE(variant->info.fau_consts));
+
       result = panvk_compile_nir(dev, nir, info->flags, &inputs, variant);
 
       /* We need to update info.push.count because it's used to initialize the