nvk: Break the guts of CmdDispatch into a helper

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

diff --git a/src/nouveau/vulkan/nvk_cmd_dispatch.c b/src/nouveau/vulkan/nvk_cmd_dispatch.c
index ac3f086..27fd443 100644 (file)
--- a/src/nouveau/vulkan/nvk_cmd_dispatch.c
+++ b/src/nouveau/vulkan/nvk_cmd_dispatch.c
@@ -30,8 +30,8 @@ nvk_cmd_buffer_begin_compute(struct nvk_cmd_buffer *cmd,
 { }
 
 static void
-gv100_compute_setup_launch_desc(uint32_t *qmd,
-                                uint32_t x, uint32_t y, uint32_t z)
+qmd_set_dispatch_size(UNUSED struct nvk_device *dev, uint32_t *qmd,
+                      uint32_t x, uint32_t y, uint32_t z)
 {
    NVC3C0_QMDV02_02_VAL_SET(qmd, CTA_RASTER_WIDTH, x);
    NVC3C0_QMDV02_02_VAL_SET(qmd, CTA_RASTER_HEIGHT, y);
@@ -56,13 +56,9 @@ nvk_cmd_bind_compute_pipeline(struct nvk_cmd_buffer *cmd,
    cmd->state.cs.pipeline = pipeline;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-nvk_CmdDispatch(VkCommandBuffer commandBuffer,
-                uint32_t groupCountX,
-                uint32_t groupCountY,
-                uint32_t groupCountZ)
+static uint64_t
+nvk_flush_compute_state(struct nvk_cmd_buffer *cmd)
 {
-   VK_FROM_HANDLE(nvk_cmd_buffer, cmd, commandBuffer);
    const struct nvk_compute_pipeline *pipeline = cmd->state.cs.pipeline;
    const struct nvk_shader *shader =
       &pipeline->base.shaders[MESA_SHADER_COMPUTE];
@@ -74,35 +70,55 @@ nvk_CmdDispatch(VkCommandBuffer commandBuffer,
    desc->root.cs.block_size[0] = shader->cp.block_size[0];
    desc->root.cs.block_size[1] = shader->cp.block_size[1];
    desc->root.cs.block_size[2] = shader->cp.block_size[2];
-   desc->root.cs.grid_size[0] = groupCountX;
-   desc->root.cs.grid_size[1] = groupCountY;
-   desc->root.cs.grid_size[2] = groupCountZ;
 
-   uint64_t root_table_addr;
+   uint64_t root_desc_addr;
    result = nvk_cmd_buffer_upload_data(cmd, &desc->root, sizeof(desc->root),
                                        NVK_MIN_UBO_ALIGNMENT,
-                                       &root_table_addr);
+                                       &root_desc_addr);
    if (unlikely(result != VK_SUCCESS)) {
       vk_command_buffer_set_error(&cmd->vk, result);
-      return;
+      return 0;
    }
 
    uint32_t qmd[128];
    memset(qmd, 0, sizeof(qmd));
    memcpy(qmd, pipeline->qmd_template, sizeof(pipeline->qmd_template));
 
-   gv100_compute_setup_launch_desc(qmd, groupCountX, groupCountY, groupCountZ);
+   qmd_set_dispatch_size(nvk_cmd_buffer_device(cmd), qmd,
+                         desc->root.cs.grid_size[0],
+                         desc->root.cs.grid_size[1],
+                         desc->root.cs.grid_size[2]);
 
-   gp100_cp_launch_desc_set_cb(qmd, 0, sizeof(desc->root), root_table_addr);
-   gp100_cp_launch_desc_set_cb(qmd, 1, sizeof(desc->root), root_table_addr);
+   gp100_cp_launch_desc_set_cb(qmd, 0, sizeof(desc->root), root_desc_addr);
+   gp100_cp_launch_desc_set_cb(qmd, 1, sizeof(desc->root), root_desc_addr);
 
    uint64_t qmd_addr;
    result = nvk_cmd_buffer_upload_data(cmd, qmd, sizeof(qmd), 256, &qmd_addr);
    if (unlikely(result != VK_SUCCESS)) {
       vk_command_buffer_set_error(&cmd->vk, result);
-      return;
+      return 0;
    }
 
+   return qmd_addr;
+}
+
+VKAPI_ATTR void VKAPI_CALL
+nvk_CmdDispatch(VkCommandBuffer commandBuffer,
+                uint32_t groupCountX,
+                uint32_t groupCountY,
+                uint32_t groupCountZ)
+{
+   VK_FROM_HANDLE(nvk_cmd_buffer, cmd, commandBuffer);
+   struct nvk_descriptor_state *desc = &cmd->state.cs.descriptors;
+
+   desc->root.cs.grid_size[0] = groupCountX;
+   desc->root.cs.grid_size[1] = groupCountY;
+   desc->root.cs.grid_size[2] = groupCountZ;
+
+   uint64_t qmd_addr = nvk_flush_compute_state(cmd);
+   if (unlikely(qmd_addr == 0))
+      return;
+
    struct nv_push *p = nvk_cmd_buffer_push(cmd, 6);
 
    P_MTHD(p, NVA0C0, INVALIDATE_SHADER_CACHES_NO_WFI);