anv/pipeline: Roll compute_urb_partition into emit_urb_setup

Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Topi Pohjolainen <[email protected]>

3 files changed, 138 insertions, 156 deletions

diff --git a/src/intel/vulkan/anv_pipeline.c b/src/intel/vulkan/anv_pipeline.c
index 9780201ca33..f96fe2298fc 100644
--- a/src/intel/vulkan/anv_pipeline.c
+++ b/src/intel/vulkan/anv_pipeline.c
@@ -805,147 +805,6 @@ anv_pipeline_compile_cs(struct anv_pipeline *pipeline,
    return VK_SUCCESS;
 }
 
-void
-anv_compute_urb_partition(struct anv_pipeline *pipeline)
-{
-   const struct gen_device_info *devinfo = &pipeline->device->info;
-
-   bool vs_present = pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT;
-   unsigned vs_size = vs_present ?
-      get_vs_prog_data(pipeline)->base.urb_entry_size : 1;
-   unsigned vs_entry_size_bytes = vs_size * 64;
-   bool gs_present = pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT;
-   unsigned gs_size = gs_present ?
-      get_gs_prog_data(pipeline)->base.urb_entry_size : 1;
-   unsigned gs_entry_size_bytes = gs_size * 64;
-
-   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
-    *
-    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
-    *     Allocation Size is less than 9 512-bit URB entries.
-    *
-    * Similar text exists for GS.
-    */
-   unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
-   unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
-
-   /* URB allocations must be done in 8k chunks. */
-   unsigned chunk_size_bytes = 8192;
-
-   /* Determine the size of the URB in chunks. */
-   unsigned urb_chunks = pipeline->urb.total_size * 1024 / chunk_size_bytes;
-
-   /* Reserve space for push constants */
-   unsigned push_constant_kb;
-   if (pipeline->device->info.gen >= 8)
-      push_constant_kb = 32;
-   else if (pipeline->device->info.is_haswell)
-      push_constant_kb = pipeline->device->info.gt == 3 ? 32 : 16;
-   else
-      push_constant_kb = 16;
-
-   unsigned push_constant_bytes = push_constant_kb * 1024;
-   unsigned push_constant_chunks =
-      push_constant_bytes / chunk_size_bytes;
-
-   /* Initially, assign each stage the minimum amount of URB space it needs,
-    * and make a note of how much additional space it "wants" (the amount of
-    * additional space it could actually make use of).
-    */
-
-   /* VS has a lower limit on the number of URB entries */
-   unsigned vs_chunks =
-      ALIGN(devinfo->urb.min_vs_entries * vs_entry_size_bytes,
-            chunk_size_bytes) / chunk_size_bytes;
-   unsigned vs_wants =
-      ALIGN(devinfo->urb.max_vs_entries * vs_entry_size_bytes,
-            chunk_size_bytes) / chunk_size_bytes - vs_chunks;
-
-   unsigned gs_chunks = 0;
-   unsigned gs_wants = 0;
-   if (gs_present) {
-      /* There are two constraints on the minimum amount of URB space we can
-       * allocate:
-       *
-       * (1) We need room for at least 2 URB entries, since we always operate
-       * the GS in DUAL_OBJECT mode.
-       *
-       * (2) We can't allocate less than nr_gs_entries_granularity.
-       */
-      gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
-                        chunk_size_bytes) / chunk_size_bytes;
-      gs_wants =
-         ALIGN(devinfo->urb.max_gs_entries * gs_entry_size_bytes,
-               chunk_size_bytes) / chunk_size_bytes - gs_chunks;
-   }
-
-   /* There should always be enough URB space to satisfy the minimum
-    * requirements of each stage.
-    */
-   unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
-   assert(total_needs <= urb_chunks);
-
-   /* Mete out remaining space (if any) in proportion to "wants". */
-   unsigned total_wants = vs_wants + gs_wants;
-   unsigned remaining_space = urb_chunks - total_needs;
-   if (remaining_space > total_wants)
-      remaining_space = total_wants;
-   if (remaining_space > 0) {
-      unsigned vs_additional = (unsigned)
-         round(vs_wants * (((double) remaining_space) / total_wants));
-      vs_chunks += vs_additional;
-      remaining_space -= vs_additional;
-      gs_chunks += remaining_space;
-   }
-
-   /* Sanity check that we haven't over-allocated. */
-   assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
-
-   /* Finally, compute the number of entries that can fit in the space
-    * allocated to each stage.
-    */
-   unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
-   unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
-
-   /* Since we rounded up when computing *_wants, this may be slightly more
-    * than the maximum allowed amount, so correct for that.
-    */
-   nr_vs_entries = MIN2(nr_vs_entries, devinfo->urb.max_vs_entries);
-   nr_gs_entries = MIN2(nr_gs_entries, devinfo->urb.max_gs_entries);
-
-   /* Ensure that we program a multiple of the granularity. */
-   nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
-   nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
-
-   /* Finally, sanity check to make sure we have at least the minimum number
-    * of entries needed for each stage.
-    */
-   assert(nr_vs_entries >= devinfo->urb.min_vs_entries);
-   if (gs_present)
-      assert(nr_gs_entries >= 2);
-
-   /* Lay out the URB in the following order:
-    * - push constants
-    * - VS
-    * - GS
-    */
-   pipeline->urb.start[MESA_SHADER_VERTEX] = push_constant_chunks;
-   pipeline->urb.size[MESA_SHADER_VERTEX] = vs_size;
-   pipeline->urb.entries[MESA_SHADER_VERTEX] = nr_vs_entries;
-
-   pipeline->urb.start[MESA_SHADER_GEOMETRY] = push_constant_chunks + vs_chunks;
-   pipeline->urb.size[MESA_SHADER_GEOMETRY] = gs_size;
-   pipeline->urb.entries[MESA_SHADER_GEOMETRY] = nr_gs_entries;
-
-   pipeline->urb.start[MESA_SHADER_TESS_CTRL] = push_constant_chunks;
-   pipeline->urb.size[MESA_SHADER_TESS_CTRL] = 1;
-   pipeline->urb.entries[MESA_SHADER_TESS_CTRL] = 0;
-
-   pipeline->urb.start[MESA_SHADER_TESS_EVAL] = push_constant_chunks;
-   pipeline->urb.size[MESA_SHADER_TESS_EVAL] = 1;
-   pipeline->urb.entries[MESA_SHADER_TESS_EVAL] = 0;
-}
-
 /**
  * Copy pipeline state not marked as dynamic.
  * Dynamic state is pipeline state which hasn't been provided at pipeline
@@ -1244,7 +1103,6 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
    }
 
    anv_pipeline_setup_l3_config(pipeline, false);
-   anv_compute_urb_partition(pipeline);
 
    const VkPipelineVertexInputStateCreateInfo *vi_info =
       pCreateInfo->pVertexInputState;
diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index f32806ca299..eb91d4b3143 100644
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -1521,9 +1521,6 @@ struct anv_pipeline {
    struct anv_shader_bin *                      shaders[MESA_SHADER_STAGES];
 
    struct {
-      uint32_t                                  start[MESA_SHADER_GEOMETRY + 1];
-      uint32_t                                  size[MESA_SHADER_GEOMETRY + 1];
-      uint32_t                                  entries[MESA_SHADER_GEOMETRY + 1];
       const struct gen_l3_config *              l3_config;
       uint32_t                                  total_size;
    } urb;
@@ -1638,9 +1635,6 @@ anv_get_isl_format(const struct gen_device_info *devinfo, VkFormat vk_format,
 }
 
 void
-anv_compute_urb_partition(struct anv_pipeline *pipeline);
-
-void
 anv_pipeline_setup_l3_config(struct anv_pipeline *pipeline, bool needs_slm);
 
 /**
diff --git a/src/intel/vulkan/genX_pipeline_util.h b/src/intel/vulkan/genX_pipeline_util.h
index 6518fae7988..b23a4212fcf 100644
--- a/src/intel/vulkan/genX_pipeline_util.h
+++ b/src/intel/vulkan/genX_pipeline_util.h
@@ -190,9 +190,123 @@ emit_vertex_input(struct anv_pipeline *pipeline,
 static inline void
 emit_urb_setup(struct anv_pipeline *pipeline)
 {
-#if GEN_GEN == 7 && !GEN_IS_HASWELL
    struct anv_device *device = pipeline->device;
 
+   bool vs_present = pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT;
+   unsigned vs_size = vs_present ?
+      get_vs_prog_data(pipeline)->base.urb_entry_size : 1;
+   unsigned vs_entry_size_bytes = vs_size * 64;
+   bool gs_present = pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT;
+   unsigned gs_size = gs_present ?
+      get_gs_prog_data(pipeline)->base.urb_entry_size : 1;
+   unsigned gs_entry_size_bytes = gs_size * 64;
+
+   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
+    *
+    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
+    *     Allocation Size is less than 9 512-bit URB entries.
+    *
+    * Similar text exists for GS.
+    */
+   unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
+   unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
+
+   /* URB allocations must be done in 8k chunks. */
+   unsigned chunk_size_bytes = 8192;
+
+   /* Determine the size of the URB in chunks. */
+   unsigned urb_chunks = pipeline->urb.total_size * 1024 / chunk_size_bytes;
+
+   /* Reserve space for push constants */
+   unsigned push_constant_kb;
+   if (pipeline->device->info.gen >= 8)
+      push_constant_kb = 32;
+   else if (pipeline->device->info.is_haswell)
+      push_constant_kb = pipeline->device->info.gt == 3 ? 32 : 16;
+   else
+      push_constant_kb = 16;
+
+   unsigned push_constant_bytes = push_constant_kb * 1024;
+   unsigned push_constant_chunks =
+      push_constant_bytes / chunk_size_bytes;
+
+   /* Initially, assign each stage the minimum amount of URB space it needs,
+    * and make a note of how much additional space it "wants" (the amount of
+    * additional space it could actually make use of).
+    */
+
+   /* VS has a lower limit on the number of URB entries */
+   unsigned vs_chunks =
+      ALIGN(device->info.urb.min_vs_entries * vs_entry_size_bytes,
+            chunk_size_bytes) / chunk_size_bytes;
+   unsigned vs_wants =
+      ALIGN(device->info.urb.max_vs_entries * vs_entry_size_bytes,
+            chunk_size_bytes) / chunk_size_bytes - vs_chunks;
+
+   unsigned gs_chunks = 0;
+   unsigned gs_wants = 0;
+   if (gs_present) {
+      /* There are two constraints on the minimum amount of URB space we can
+       * allocate:
+       *
+       * (1) We need room for at least 2 URB entries, since we always operate
+       * the GS in DUAL_OBJECT mode.
+       *
+       * (2) We can't allocate less than nr_gs_entries_granularity.
+       */
+      gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
+                        chunk_size_bytes) / chunk_size_bytes;
+      gs_wants =
+         ALIGN(device->info.urb.max_gs_entries * gs_entry_size_bytes,
+               chunk_size_bytes) / chunk_size_bytes - gs_chunks;
+   }
+
+   /* There should always be enough URB space to satisfy the minimum
+    * requirements of each stage.
+    */
+   unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
+   assert(total_needs <= urb_chunks);
+
+   /* Mete out remaining space (if any) in proportion to "wants". */
+   unsigned total_wants = vs_wants + gs_wants;
+   unsigned remaining_space = urb_chunks - total_needs;
+   if (remaining_space > total_wants)
+      remaining_space = total_wants;
+   if (remaining_space > 0) {
+      unsigned vs_additional = (unsigned)
+         round(vs_wants * (((double) remaining_space) / total_wants));
+      vs_chunks += vs_additional;
+      remaining_space -= vs_additional;
+      gs_chunks += remaining_space;
+   }
+
+   /* Sanity check that we haven't over-allocated. */
+   assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
+
+   /* Finally, compute the number of entries that can fit in the space
+    * allocated to each stage.
+    */
+   unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
+   unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
+
+   /* Since we rounded up when computing *_wants, this may be slightly more
+    * than the maximum allowed amount, so correct for that.
+    */
+   nr_vs_entries = MIN2(nr_vs_entries, device->info.urb.max_vs_entries);
+   nr_gs_entries = MIN2(nr_gs_entries, device->info.urb.max_gs_entries);
+
+   /* Ensure that we program a multiple of the granularity. */
+   nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
+   nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
+
+   /* Finally, sanity check to make sure we have at least the minimum number
+    * of entries needed for each stage.
+    */
+   assert(nr_vs_entries >= device->info.urb.min_vs_entries);
+   if (gs_present)
+      assert(nr_gs_entries >= 2);
+
+#if GEN_GEN == 7 && !GEN_IS_HASWELL
    /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
     *
     *    "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth stall
@@ -208,13 +322,29 @@ emit_urb_setup(struct anv_pipeline *pipeline)
    }
 #endif
 
-   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
-      anv_batch_emit(&pipeline->batch, GENX(3DSTATE_URB_VS), urb) {
-         urb._3DCommandSubOpcode       = 48 + i;
-         urb.VSURBStartingAddress      = pipeline->urb.start[i];
-         urb.VSURBEntryAllocationSize  = pipeline->urb.size[i] - 1;
-         urb.VSNumberofURBEntries      = pipeline->urb.entries[i];
-      }
+   /* Lay out the URB in the following order:
+    * - push constants
+    * - VS
+    * - GS
+    */
+   anv_batch_emit(&pipeline->batch, GENX(3DSTATE_URB_VS), urb) {
+      urb.VSURBStartingAddress      = push_constant_chunks;
+      urb.VSURBEntryAllocationSize  = vs_size - 1;
+      urb.VSNumberofURBEntries      = nr_vs_entries;
+   }
+
+   anv_batch_emit(&pipeline->batch, GENX(3DSTATE_URB_HS), urb) {
+      urb.HSURBStartingAddress      = push_constant_chunks;
+   }
+
+   anv_batch_emit(&pipeline->batch, GENX(3DSTATE_URB_DS), urb) {
+      urb.DSURBStartingAddress      = push_constant_chunks;
+   }
+
+   anv_batch_emit(&pipeline->batch, GENX(3DSTATE_URB_GS), urb) {
+      urb.GSURBStartingAddress      = push_constant_chunks + vs_chunks;
+      urb.GSURBEntryAllocationSize  = gs_size - 1;
+      urb.GSNumberofURBEntries      = nr_gs_entries;
    }
 }