nvk: Fill out sample locations on Maxwell B+

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

src/nouveau/vulkan/nvk_cmd_buffer.c

@@ -51,6 +51,8 @@ nvk_create_cmd_buffer(struct vk_command_pool *vk_pool,
 
    cmd->vk.dynamic_graphics_state.vi =
       &cmd->state.gfx._dynamic_vi;
+   cmd->vk.dynamic_graphics_state.ms.sample_locations =
+      &cmd->state.gfx._dynamic_sl;
 
    list_inithead(&cmd->bos);
    util_dynarray_init(&cmd->pushes, NULL);

src/nouveau/vulkan/nvk_cmd_buffer.h

@@ -83,6 +83,7 @@ struct nvk_graphics_state {
 
    /* Needed by vk_command_buffer::dynamic_graphics_state */
    struct vk_vertex_input_state _dynamic_vi;
+   struct vk_sample_locations_state _dynamic_sl;
 };
 
 struct nvk_compute_state {

src/nouveau/vulkan/nvk_cmd_draw.c

@@ -9,6 +9,7 @@
 #include "nvk_pipeline.h"
 
 #include "nil_format.h"
+#include "util/bitpack_helpers.h"
 #include "vulkan/runtime/vk_render_pass.h"
 #include "vulkan/util/vk_format.h"
 
@@ -909,6 +910,69 @@ nvk_flush_rs_state(struct nvk_cmd_buffer *cmd)
    }
 }
 
+static VkSampleLocationEXT
+vk_sample_location(const struct vk_sample_locations_state *sl,
+                   uint32_t x, uint32_t y, uint32_t s)
+{
+   x = x % sl->grid_size.width;
+   y = y % sl->grid_size.height;
+
+   return sl->locations[(x + y * sl->grid_size.width) * sl->per_pixel + s];
+}
+
+struct nvk_sample_location {
+   uint8_t x_u4:4;
+   uint8_t y_u4:4;
+};
+
+static struct nvk_sample_location
+vk_to_nvk_sample_location(VkSampleLocationEXT loc)
+{
+   return (struct nvk_sample_location) {
+      .x_u4 = util_bitpack_ufixed(loc.x, 0, 3, 4),
+      .y_u4 = util_bitpack_ufixed(loc.y, 0, 3, 4),
+   };
+}
+
+static void
+nvk_flush_ms_state(struct nvk_cmd_buffer *cmd)
+{
+   const struct vk_dynamic_graphics_state *dyn =
+      &cmd->vk.dynamic_graphics_state;
+
+   if (BITSET_TEST(dyn->dirty, MESA_VK_DYNAMIC_MS_SAMPLE_LOCATIONS)) {
+      const struct vk_sample_locations_state *sl = dyn->ms.sample_locations;
+
+      if (nvk_cmd_buffer_3d_cls(cmd) >= MAXWELL_B) {
+         struct nvk_sample_location loc[16];
+         for (uint32_t n = 0; n < ARRAY_SIZE(loc); n++) {
+            const uint32_t s = n % sl->per_pixel;
+            const uint32_t px = n / sl->per_pixel;
+            const uint32_t x = px % 2;
+            const uint32_t y = px / 2;
+
+            loc[n] = vk_to_nvk_sample_location(vk_sample_location(sl, x, y, s));
+         }
+
+         struct nv_push *p = nvk_cmd_buffer_push(cmd, 5);
+
+         P_MTHD(p, NVB197, SET_ANTI_ALIAS_SAMPLE_POSITIONS(0));
+         for (uint32_t i = 0; i < 4; i++) {
+            P_NVB197_SET_ANTI_ALIAS_SAMPLE_POSITIONS(p, i, {
+               .x0 = loc[i * 4 + 0].x_u4,
+               .y0 = loc[i * 4 + 0].y_u4,
+               .x1 = loc[i * 4 + 1].x_u4,
+               .y1 = loc[i * 4 + 1].y_u4,
+               .x2 = loc[i * 4 + 2].x_u4,
+               .y2 = loc[i * 4 + 2].y_u4,
+               .x3 = loc[i * 4 + 3].x_u4,
+               .y3 = loc[i * 4 + 3].y_u4,
+            });
+         }
+      }
+   }
+}
+
 static uint32_t
 vk_to_nv9097_compare_op(VkCompareOp vk_op)
 {
@@ -1085,8 +1149,8 @@ nvk_flush_dynamic_state(struct nvk_cmd_buffer *cmd)
    nvk_flush_rs_state(cmd);
 
    /* MESA_VK_DYNAMIC_FSR */
-   /* MESA_VK_DYNAMIC_MS_SAMPLE_LOCATIONS */
 
+   nvk_flush_ms_state(cmd);
    nvk_flush_ds_state(cmd);
    nvk_flush_cb_state(cmd);
 

src/nouveau/vulkan/nvk_cmd_meta.c

@@ -46,6 +46,7 @@ nvk_device_finish_meta(struct nvk_device *dev)
 
 struct nvk_meta_save {
    struct vk_vertex_input_state _dynamic_vi;
+   struct vk_sample_locations_state _dynamic_sl;
    struct vk_dynamic_graphics_state dynamic;
    struct nvk_graphics_pipeline *pipeline;
    struct nvk_addr_range vb0;
@@ -61,6 +62,7 @@ nvk_meta_begin(struct nvk_cmd_buffer *cmd,
 {
    save->dynamic = cmd->vk.dynamic_graphics_state;
    save->_dynamic_vi = cmd->state.gfx._dynamic_vi;
+   save->_dynamic_sl = cmd->state.gfx._dynamic_sl;
 
    save->pipeline = cmd->state.gfx.pipeline;
    save->vb0 = cmd->state.gfx.vb0;
@@ -105,8 +107,10 @@ nvk_meta_end(struct nvk_cmd_buffer *cmd,
 
    /* Restore the dynamic state */
    assert(save->dynamic.vi == &cmd->state.gfx._dynamic_vi);
+   assert(save->dynamic.ms.sample_locations == &cmd->state.gfx._dynamic_sl);
    cmd->vk.dynamic_graphics_state = save->dynamic;
    cmd->state.gfx._dynamic_vi = save->_dynamic_vi;
+   cmd->state.gfx._dynamic_sl = save->_dynamic_sl;
    memcpy(cmd->vk.dynamic_graphics_state.dirty,
           cmd->vk.dynamic_graphics_state.set,
           sizeof(cmd->vk.dynamic_graphics_state.set));

src/nouveau/vulkan/nvk_graphics_pipeline.c

@@ -355,6 +355,7 @@ nvk_graphics_pipeline_create(struct nvk_device *device,
    pipeline->push_dw_count = nv_push_dw_count(&push);
 
    pipeline->dynamic.vi = &pipeline->_dynamic_vi;
+   pipeline->dynamic.ms.sample_locations = &pipeline->_dynamic_sl;
    vk_dynamic_graphics_state_fill(&pipeline->dynamic, &state);
 
    *pPipeline = nvk_pipeline_to_handle(&pipeline->base);

src/nouveau/vulkan/nvk_physical_device.c

@@ -16,6 +16,7 @@
 #include "cla0c0.h"
 #include "cla1c0.h"
 #include "clb0c0.h"
+#include "clb197.h"
 #include "clb1c0.h"
 #include "clc0c0.h"
 #include "clc1c0.h"
@@ -355,7 +356,7 @@ vk_icdGetPhysicalDeviceProcAddr(VkInstance _instance, const char *pName)
 }
 
 static void
-nvk_get_device_extensions(const struct nvk_physical_device *device,
+nvk_get_device_extensions(const struct nvk_physical_device *pdev,
                           struct vk_device_extension_table *ext)
 {
    *ext = (struct vk_device_extension_table) {
@@ -390,6 +391,7 @@ nvk_get_device_extensions(const struct nvk_physical_device *device,
       .EXT_inline_uniform_block = true,
       .EXT_pci_bus_info = true,
       .EXT_private_data = true,
+      .EXT_sample_locations = pdev->info.cls_eng3d >= MAXWELL_B,
       .EXT_vertex_input_dynamic_state = true,
    };
 }

src/nouveau/vulkan/nvk_pipeline.h

@@ -54,6 +54,7 @@ struct nvk_graphics_pipeline {
    uint32_t push_dw_count;
 
    struct vk_vertex_input_state _dynamic_vi;
+   struct vk_sample_locations_state _dynamic_sl;
    struct vk_dynamic_graphics_state dynamic;
 };