Metal: Minimum per-vertex stride, 3D texture size + Transform feedback GPUCapabilities expansion.

- Adding in compatibility paths to support minimum per-vertex strides for vertex formats. OpenGL supports a minimum stride of 1 byte, in Metal, this minimum stride is 4 bytes. Meaing a vertex format must be atleast 4-bytes in size.

- Replacing transform feedback compile-time check to conditional look-up, given TF is supported on macOS with Metal.

- 3D texture size safety check added as a general capability, rather than being in the gl backend only. Also required for Metal.

Authored by Apple: Michael Parkin-White

Ref T96261

Reviewed By: fclem

Maniphest Tasks: T96261

Differential Revision: https://developer.blender.org/D14510

18 files changed, 413 insertions, 240 deletions

diff --git a/source/blender/draw/engines/eevee/eevee_volumes.c b/source/blender/draw/engines/eevee/eevee_volumes.c
index 533e71b9b32..2d96cffb4ba 100644
--- a/source/blender/draw/engines/eevee/eevee_volumes.c
+++ b/source/blender/draw/engines/eevee/eevee_volumes.c
@@ -30,6 +30,7 @@
 #include "DEG_depsgraph_query.h"
 
 #include "GPU_capabilities.h"
+#include "GPU_context.h"
 #include "GPU_material.h"
 #include "GPU_texture.h"
 #include "eevee_private.h"
@@ -82,6 +83,13 @@ void EEVEE_volumes_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata)
   tex_size[1] = (int)ceilf(fmaxf(1.0f, viewport_size[1] / (float)tile_size));
   tex_size[2] = max_ii(scene_eval->eevee.volumetric_samples, 1);
 
+  /* Clamp 3D texture size based on device maximum. */
+  int maxSize = GPU_max_texture_3d_size();
+  BLI_assert(tex_size[0] <= maxSize);
+  tex_size[0] = tex_size[0] > maxSize ? maxSize : tex_size[0];
+  tex_size[1] = tex_size[1] > maxSize ? maxSize : tex_size[1];
+  tex_size[2] = tex_size[2] > maxSize ? maxSize : tex_size[2];
+
   common_data->vol_coord_scale[0] = viewport_size[0] / (float)(tile_size * tex_size[0]);
   common_data->vol_coord_scale[1] = viewport_size[1] / (float)(tile_size * tex_size[1]);
   common_data->vol_coord_scale[2] = 1.0f / viewport_size[0];
diff --git a/source/blender/draw/intern/draw_cache.c b/source/blender/draw/intern/draw_cache.c
index 4ff5745fc86..6537490c06c 100644
--- a/source/blender/draw/intern/draw_cache.c
+++ b/source/blender/draw/intern/draw_cache.c
@@ -826,7 +826,8 @@ GPUBatch *DRW_gpencil_dummy_buffer_get(void)
 {
   if (SHC.drw_gpencil_dummy_quad == NULL) {
     GPUVertFormat format = {0};
-    GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_U8, 1, GPU_FETCH_INT);
+    /* NOTE: Use GPU_COMP_U32 to satisfy minimum 4-byte vertex stride for Metal backend. */
+    GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_U32, 1, GPU_FETCH_INT);
     GPUVertBuf *vbo = GPU_vertbuf_create_with_format(&format);
     GPU_vertbuf_data_alloc(vbo, 4);
 
diff --git a/source/blender/draw/intern/draw_cache_impl_curves.cc b/source/blender/draw/intern/draw_cache_impl_curves.cc
index 4f0072ec657..3bca17d9c56 100644
--- a/source/blender/draw/intern/draw_cache_impl_curves.cc
+++ b/source/blender/draw/intern/draw_cache_impl_curves.cc
@@ -269,7 +269,8 @@ static void curves_batch_cache_ensure_procedural_pos(const Curves &curves,
     GPU_vertformat_attr_add(&format, "posTime", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
     GPU_vertformat_alias_add(&format, "pos");
 
-    cache.proc_point_buf = GPU_vertbuf_create_with_format(&format);
+    cache.proc_point_buf = GPU_vertbuf_create_with_format_ex(
+        &format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache.proc_point_buf, cache.point_len);
 
     MutableSpan posTime_data{
@@ -279,7 +280,8 @@ static void curves_batch_cache_ensure_procedural_pos(const Curves &curves,
     GPUVertFormat length_format = {0};
     GPU_vertformat_attr_add(&length_format, "hairLength", GPU_COMP_F32, 1, GPU_FETCH_FLOAT);
 
-    cache.proc_length_buf = GPU_vertbuf_create_with_format(&length_format);
+    cache.proc_length_buf = GPU_vertbuf_create_with_format_ex(
+        &length_format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache.proc_length_buf, cache.strands_len);
 
     MutableSpan hairLength_data{
@@ -319,8 +321,8 @@ static void curves_batch_cache_ensure_procedural_final_attr(CurvesEvalCache &cac
                                                             const char *name)
 {
   CurvesEvalFinalCache &final_cache = cache.final[subdiv];
-  final_cache.attributes_buf[index] = GPU_vertbuf_create_with_format_ex(format,
-                                                                        GPU_USAGE_DEVICE_ONLY);
+  final_cache.attributes_buf[index] = GPU_vertbuf_create_with_format_ex(
+      format, GPU_USAGE_DEVICE_ONLY | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
 
   /* Create a destination buffer for the transform feedback. Sized appropriately */
   /* Those are points! not line segments. */
@@ -351,7 +353,8 @@ static void curves_batch_ensure_attribute(const Curves &curves,
   /* All attributes use vec4, see comment below. */
   GPU_vertformat_attr_add(&format, sampler_name, GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-  cache.proc_attributes_buf[index] = GPU_vertbuf_create_with_format(&format);
+  cache.proc_attributes_buf[index] = GPU_vertbuf_create_with_format_ex(
+      &format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
   GPUVertBuf *attr_vbo = cache.proc_attributes_buf[index];
 
   GPU_vertbuf_data_alloc(attr_vbo,
@@ -416,11 +419,13 @@ static void curves_batch_cache_ensure_procedural_strand_data(Curves &curves,
   uint seg_id = GPU_vertformat_attr_add(&format_seg, "data", GPU_COMP_U16, 1, GPU_FETCH_INT);
 
   /* Curve Data. */
-  cache.proc_strand_buf = GPU_vertbuf_create_with_format(&format_data);
+  cache.proc_strand_buf = GPU_vertbuf_create_with_format_ex(
+      &format_data, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
   GPU_vertbuf_data_alloc(cache.proc_strand_buf, cache.strands_len);
   GPU_vertbuf_attr_get_raw_data(cache.proc_strand_buf, data_id, &data_step);
 
-  cache.proc_strand_seg_buf = GPU_vertbuf_create_with_format(&format_seg);
+  cache.proc_strand_seg_buf = GPU_vertbuf_create_with_format_ex(
+      &format_seg, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
   GPU_vertbuf_data_alloc(cache.proc_strand_seg_buf, cache.strands_len);
   GPU_vertbuf_attr_get_raw_data(cache.proc_strand_seg_buf, seg_id, &seg_step);
 
@@ -441,7 +446,8 @@ static void curves_batch_cache_ensure_procedural_final_points(CurvesEvalCache &c
   GPUVertFormat format = {0};
   GPU_vertformat_attr_add(&format, "pos", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-  cache.final[subdiv].proc_buf = GPU_vertbuf_create_with_format_ex(&format, GPU_USAGE_DEVICE_ONLY);
+  cache.final[subdiv].proc_buf = GPU_vertbuf_create_with_format_ex(
+      &format, GPU_USAGE_DEVICE_ONLY | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
 
   /* Create a destination buffer for the transform feedback. Sized appropriately */
   /* Those are points! not line segments. */
diff --git a/source/blender/draw/intern/draw_cache_impl_particles.c b/source/blender/draw/intern/draw_cache_impl_particles.c
index 02afbab6899..4fdc46ea18b 100644
--- a/source/blender/draw/intern/draw_cache_impl_particles.c
+++ b/source/blender/draw/intern/draw_cache_impl_particles.c
@@ -32,6 +32,8 @@
 #include "ED_particle.h"
 
 #include "GPU_batch.h"
+#include "GPU_capabilities.h"
+#include "GPU_context.h"
 #include "GPU_material.h"
 
 #include "DEG_depsgraph_query.h"
@@ -808,7 +810,10 @@ static void particle_batch_cache_ensure_procedural_final_points(ParticleHairCach
   GPUVertFormat format = {0};
   GPU_vertformat_attr_add(&format, "pos", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-  cache->final[subdiv].proc_buf = GPU_vertbuf_create_with_format(&format);
+  /* Transform feedback buffer only needs to be resident in device memory. */
+  GPUUsageType type = GPU_transform_feedback_support() ? GPU_USAGE_DEVICE_ONLY : GPU_USAGE_STATIC;
+  cache->final[subdiv].proc_buf = GPU_vertbuf_create_with_format_ex(
+      &format, type | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
 
   /* Create a destination buffer for the transform feedback. Sized appropriately */
   /* Those are points! not line segments. */
@@ -873,17 +878,20 @@ static void particle_batch_cache_ensure_procedural_strand_data(PTCacheEdit *edit
   memset(cache->uv_layer_names, 0, sizeof(cache->uv_layer_names));
 
   /* Strand Data */
-  cache->proc_strand_buf = GPU_vertbuf_create_with_format(&format_data);
+  cache->proc_strand_buf = GPU_vertbuf_create_with_format_ex(
+      &format_data, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
   GPU_vertbuf_data_alloc(cache->proc_strand_buf, cache->strands_len);
   GPU_vertbuf_attr_get_raw_data(cache->proc_strand_buf, data_id, &data_step);
 
-  cache->proc_strand_seg_buf = GPU_vertbuf_create_with_format(&format_seg);
+  cache->proc_strand_seg_buf = GPU_vertbuf_create_with_format_ex(
+      &format_seg, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
   GPU_vertbuf_data_alloc(cache->proc_strand_seg_buf, cache->strands_len);
   GPU_vertbuf_attr_get_raw_data(cache->proc_strand_seg_buf, seg_id, &seg_step);
 
   /* UV layers */
   for (int i = 0; i < cache->num_uv_layers; i++) {
-    cache->proc_uv_buf[i] = GPU_vertbuf_create_with_format(&format_uv);
+    cache->proc_uv_buf[i] = GPU_vertbuf_create_with_format_ex(
+        &format_uv, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache->proc_uv_buf[i], cache->strands_len);
     GPU_vertbuf_attr_get_raw_data(cache->proc_uv_buf[i], uv_id, &uv_step[i]);
 
@@ -913,7 +921,8 @@ static void particle_batch_cache_ensure_procedural_strand_data(PTCacheEdit *edit
 
   /* Vertex colors */
   for (int i = 0; i < cache->num_col_layers; i++) {
-    cache->proc_col_buf[i] = GPU_vertbuf_create_with_format(&format_col);
+    cache->proc_col_buf[i] = GPU_vertbuf_create_with_format_ex(
+        &format_col, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache->proc_col_buf[i], cache->strands_len);
     GPU_vertbuf_attr_get_raw_data(cache->proc_col_buf[i], col_id, &col_step[i]);
 
@@ -1059,8 +1068,9 @@ static void particle_batch_cache_ensure_procedural_indices(PTCacheEdit *edit,
   static GPUVertFormat format = {0};
   GPU_vertformat_clear(&format);
 
-  /* initialize vertex format */
-  GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_U8, 1, GPU_FETCH_INT_TO_FLOAT_UNIT);
+  /* NOTE: initialize vertex format. Using GPU_COMP_U32 to satisfy Metal's 4-byte minimum
+   * stride requirement. */
+  GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_U32, 1, GPU_FETCH_INT_TO_FLOAT_UNIT);
 
   GPUVertBuf *vbo = GPU_vertbuf_create_with_format(&format);
   GPU_vertbuf_data_alloc(vbo, 1);
@@ -1101,7 +1111,8 @@ static void particle_batch_cache_ensure_procedural_pos(PTCacheEdit *edit,
     uint pos_id = GPU_vertformat_attr_add(
         &pos_format, "posTime", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-    cache->proc_point_buf = GPU_vertbuf_create_with_format(&pos_format);
+    cache->proc_point_buf = GPU_vertbuf_create_with_format_ex(
+        &pos_format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache->proc_point_buf, cache->point_len);
 
     GPUVertBufRaw pos_step;
@@ -1111,7 +1122,8 @@ static void particle_batch_cache_ensure_procedural_pos(PTCacheEdit *edit,
     uint length_id = GPU_vertformat_attr_add(
         &length_format, "hairLength", GPU_COMP_F32, 1, GPU_FETCH_FLOAT);
 
-    cache->proc_length_buf = GPU_vertbuf_create_with_format(&length_format);
+    cache->proc_length_buf = GPU_vertbuf_create_with_format_ex(
+        &length_format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
     GPU_vertbuf_data_alloc(cache->proc_length_buf, cache->strands_len);
 
     GPUVertBufRaw length_step;
diff --git a/source/blender/draw/intern/draw_curves.cc b/source/blender/draw/intern/draw_curves.cc
index 233af08c363..9c4181b0161 100644
--- a/source/blender/draw/intern/draw_curves.cc
+++ b/source/blender/draw/intern/draw_curves.cc
@@ -33,25 +33,17 @@
 #include "draw_manager.h"
 #include "draw_shader.h"
 
-#ifndef __APPLE__
-#  define USE_TRANSFORM_FEEDBACK
-#  define USE_COMPUTE_SHADERS
-#endif
-
 BLI_INLINE eParticleRefineShaderType drw_curves_shader_type_get()
 {
-#ifdef USE_COMPUTE_SHADERS
   if (GPU_compute_shader_support() && GPU_shader_storage_buffer_objects_support()) {
     return PART_REFINE_SHADER_COMPUTE;
   }
-#endif
-#ifdef USE_TRANSFORM_FEEDBACK
-  return PART_REFINE_SHADER_TRANSFORM_FEEDBACK;
-#endif
+  if (GPU_transform_feedback_support()) {
+    return PART_REFINE_SHADER_TRANSFORM_FEEDBACK;
+  }
   return PART_REFINE_SHADER_TRANSFORM_FEEDBACK_WORKAROUND;
 }
 
-#ifndef USE_TRANSFORM_FEEDBACK
 struct CurvesEvalCall {
   struct CurvesEvalCall *next;
   GPUVertBuf *vbo;
@@ -63,7 +55,6 @@ static CurvesEvalCall *g_tf_calls = nullptr;
 static int g_tf_id_offset;
 static int g_tf_target_width;
 static int g_tf_target_height;
-#endif
 
 static GPUVertBuf *g_dummy_vbo = nullptr;
 static GPUTexture *g_dummy_texture = nullptr;
@@ -106,18 +97,20 @@ void DRW_curves_init(DRWData *drw_data)
   CurvesUniformBufPool *pool = drw_data->curves_ubos;
   pool->reset();
 
-#if defined(USE_TRANSFORM_FEEDBACK) || defined(USE_COMPUTE_SHADERS)
-  g_tf_pass = DRW_pass_create("Update Curves Pass", (DRWState)0);
-#else
-  g_tf_pass = DRW_pass_create("Update Curves Pass", DRW_STATE_WRITE_COLOR);
-#endif
+  if (GPU_transform_feedback_support() || GPU_compute_shader_support()) {
+    g_tf_pass = DRW_pass_create("Update Curves Pass", (DRWState)0);
+  }
+  else {
+    g_tf_pass = DRW_pass_create("Update Curves Pass", DRW_STATE_WRITE_COLOR);
+  }
 
   if (g_dummy_vbo == nullptr) {
     /* initialize vertex format */
     GPUVertFormat format = {0};
     uint dummy_id = GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-    g_dummy_vbo = GPU_vertbuf_create_with_format(&format);
+    g_dummy_vbo = GPU_vertbuf_create_with_format_ex(
+        &format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
 
     const float vert[4] = {0.0f, 0.0f, 0.0f, 0.0f};
     GPU_vertbuf_data_alloc(g_dummy_vbo, 1);
@@ -201,21 +194,24 @@ static void drw_curves_cache_update_transform_feedback(CurvesEvalCache *cache,
 {
   GPUShader *tf_shader = curves_eval_shader_get(CURVES_EVAL_CATMULL_ROM);
 
-#ifdef USE_TRANSFORM_FEEDBACK
-  DRWShadingGroup *tf_shgrp = DRW_shgroup_transform_feedback_create(tf_shader, g_tf_pass, vbo);
-#else
-  DRWShadingGroup *tf_shgrp = DRW_shgroup_create(tf_shader, g_tf_pass);
-
-  CurvesEvalCall *pr_call = MEM_new<CurvesEvalCall>(__func__);
-  pr_call->next = g_tf_calls;
-  pr_call->vbo = vbo;
-  pr_call->shgrp = tf_shgrp;
-  pr_call->vert_len = final_points_len;
-  g_tf_calls = pr_call;
-  DRW_shgroup_uniform_int(tf_shgrp, "targetHeight", &g_tf_target_height, 1);
-  DRW_shgroup_uniform_int(tf_shgrp, "targetWidth", &g_tf_target_width, 1);
-  DRW_shgroup_uniform_int(tf_shgrp, "idOffset", &g_tf_id_offset, 1);
-#endif
+  DRWShadingGroup *tf_shgrp = nullptr;
+  if (GPU_transform_feedback_support()) {
+    tf_shgrp = DRW_shgroup_transform_feedback_create(tf_shader, g_tf_pass, vbo);
+  }
+  else {
+    tf_shgrp = DRW_shgroup_create(tf_shader, g_tf_pass);
+
+    CurvesEvalCall *pr_call = MEM_new<CurvesEvalCall>(__func__);
+    pr_call->next = g_tf_calls;
+    pr_call->vbo = vbo;
+    pr_call->shgrp = tf_shgrp;
+    pr_call->vert_len = final_points_len;
+    g_tf_calls = pr_call;
+    DRW_shgroup_uniform_int(tf_shgrp, "targetHeight", &g_tf_target_height, 1);
+    DRW_shgroup_uniform_int(tf_shgrp, "targetWidth", &g_tf_target_width, 1);
+    DRW_shgroup_uniform_int(tf_shgrp, "idOffset", &g_tf_id_offset, 1);
+  }
+  BLI_assert(tf_shgrp != nullptr);
 
   drw_curves_cache_shgrp_attach_resources(tf_shgrp, cache, tex, subdiv);
   DRW_shgroup_call_procedural_points(tf_shgrp, nullptr, final_points_len);
@@ -411,82 +407,118 @@ void DRW_curves_update()
   /* Update legacy hair too, to avoid verbosity in callers. */
   DRW_hair_update();
 
-#ifndef USE_TRANSFORM_FEEDBACK
-  /**
-   * Workaround to transform feedback not working on mac.
-   * On some system it crashes (see T58489) and on some other it renders garbage (see T60171).
-   *
-   * So instead of using transform feedback we render to a texture,
-   * read back the result to system memory and re-upload as VBO data.
-   * It is really not ideal performance wise, but it is the simplest
-   * and the most local workaround that still uses the power of the GPU.
-   */
-
-  if (g_tf_calls == nullptr) {
-    return;
-  }
+  if (!GPU_transform_feedback_support()) {
+    /**
+     * Workaround to transform feedback not working on mac.
+     * On some system it crashes (see T58489) and on some other it renders garbage (see T60171).
+     *
+     * So instead of using transform feedback we render to a texture,
+     * read back the result to system memory and re-upload as VBO data.
+     * It is really not ideal performance wise, but it is the simplest
+     * and the most local workaround that still uses the power of the GPU.
+     */
+
+    if (g_tf_calls == nullptr) {
+      return;
+    }
 
-  /* Search ideal buffer size. */
-  uint max_size = 0;
-  for (CurvesEvalCall *pr_call = g_tf_calls; pr_call; pr_call = pr_call->next) {
-    max_size = max_ii(max_size, pr_call->vert_len);
-  }
+    /* Search ideal buffer size. */
+    uint max_size = 0;
+    for (CurvesEvalCall *pr_call = g_tf_calls; pr_call; pr_call = pr_call->next) {
+      max_size = max_ii(max_size, pr_call->vert_len);
+    }
+
+    /* Create target Texture / Frame-buffer */
+    /* Don't use max size as it can be really heavy and fail.
+     * Do chunks of maximum 2048 * 2048 hair points. */
+    int width = 2048;
+    int height = min_ii(width, 1 + max_size / width);
+    GPUTexture *tex = DRW_texture_pool_query_2d(
+        width, height, GPU_RGBA32F, (DrawEngineType *)DRW_curves_update);
+    g_tf_target_height = height;
+    g_tf_target_width = width;
+
+    GPUFrameBuffer *fb = nullptr;
+    GPU_framebuffer_ensure_config(&fb,
+                                  {
+                                      GPU_ATTACHMENT_NONE,
+                                      GPU_ATTACHMENT_TEXTURE(tex),
+                                  });
+
+    float *data = static_cast<float *>(
+        MEM_mallocN(sizeof(float[4]) * width * height, "tf fallback buffer"));
+
+    GPU_framebuffer_bind(fb);
+    while (g_tf_calls != nullptr) {
+      CurvesEvalCall *pr_call = g_tf_calls;
+      g_tf_calls = g_tf_calls->next;
+
+      g_tf_id_offset = 0;
+      while (pr_call->vert_len > 0) {
+        int max_read_px_len = min_ii(width * height, pr_call->vert_len);
+
+        DRW_draw_pass_subset(g_tf_pass, pr_call->shgrp, pr_call->shgrp);
+        /* Read back result to main memory. */
+        GPU_framebuffer_read_color(fb, 0, 0, width, height, 4, 0, GPU_DATA_FLOAT, data);
+        /* Upload back to VBO. */
+        GPU_vertbuf_use(pr_call->vbo);
+        GPU_vertbuf_update_sub(pr_call->vbo,
+                               sizeof(float[4]) * g_tf_id_offset,
+                               sizeof(float[4]) * max_read_px_len,
+                               data);
+
+        g_tf_id_offset += max_read_px_len;
+        pr_call->vert_len -= max_read_px_len;
+      }
 
-  /* Create target Texture / Frame-buffer */
-  /* Don't use max size as it can be really heavy and fail.
-   * Do chunks of maximum 2048 * 2048 hair points. */
-  int width = 2048;
-  int height = min_ii(width, 1 + max_size / width);
-  GPUTexture *tex = DRW_texture_pool_query_2d(
-      width, height, GPU_RGBA32F, (DrawEngineType *)DRW_curves_update);
-  g_tf_target_height = height;
-  g_tf_target_width = width;
-
-  GPUFrameBuffer *fb = nullptr;
-  GPU_framebuffer_ensure_config(&fb,
-                                {
-                                    GPU_ATTACHMENT_NONE,
-                                    GPU_ATTACHMENT_TEXTURE(tex),
-                                });
-
-  float *data = static_cast<float *>(
-      MEM_mallocN(sizeof(float[4]) * width * height, "tf fallback buffer"));
-
-  GPU_framebuffer_bind(fb);
-  while (g_tf_calls != nullptr) {
-    CurvesEvalCall *pr_call = g_tf_calls;
-    g_tf_calls = g_tf_calls->next;
-
-    g_tf_id_offset = 0;
-    while (pr_call->vert_len > 0) {
-      int max_read_px_len = min_ii(width * height, pr_call->vert_len);
-
-      DRW_draw_pass_subset(g_tf_pass, pr_call->shgrp, pr_call->shgrp);
-      /* Read back result to main memory. */
-      GPU_framebuffer_read_color(fb, 0, 0, width, height, 4, 0, GPU_DATA_FLOAT, data);
-      /* Upload back to VBO. */
-      GPU_vertbuf_use(pr_call->vbo);
-      GPU_vertbuf_update_sub(pr_call->vbo,
-                             sizeof(float[4]) * g_tf_id_offset,
-                             sizeof(float[4]) * max_read_px_len,
-                             data);
-
-      g_tf_id_offset += max_read_px_len;
-      pr_call->vert_len -= max_read_px_len;
+      MEM_freeN(pr_call);
     }
 
-    MEM_freeN(pr_call);
+    MEM_freeN(data);
+    GPU_framebuffer_free(fb);
   }
+  else {
+    /* Note(Metal): If compute is not supported, bind a temporary framebuffer to avoid
+     * side-effects from rendering in the active buffer.
+     * We also need to guarantee that a Framebuffer is active to perform any rendering work,
+     * even if there is no output */
+    GPUFrameBuffer *temp_fb = nullptr;
+    GPUFrameBuffer *prev_fb = nullptr;
+    if (GPU_type_matches_ex(GPU_DEVICE_ANY, GPU_OS_MAC, GPU_DRIVER_ANY, GPU_BACKEND_METAL)) {
+      if (!GPU_compute_shader_support()) {
+        prev_fb = GPU_framebuffer_active_get();
+        char errorOut[256];
+        /* if the framebuffer is invalid we need a dummy framebuffer to be bound. */
+        if (!GPU_framebuffer_check_valid(prev_fb, errorOut)) {
+          int width = 64;
+          int height = 64;
+          GPUTexture *tex = DRW_texture_pool_query_2d(
+              width, height, GPU_DEPTH_COMPONENT32F, (DrawEngineType *)DRW_hair_update);
+          g_tf_target_height = height;
+          g_tf_target_width = width;
+
+          GPU_framebuffer_ensure_config(&temp_fb, {GPU_ATTACHMENT_TEXTURE(tex)});
+
+          GPU_framebuffer_bind(temp_fb);
+        }
+      }
+    }
 
-  MEM_freeN(data);
-  GPU_framebuffer_free(fb);
-#else
-  /* Just render the pass when using compute shaders or transform feedback. */
-  DRW_draw_pass(g_tf_pass);
-  if (drw_curves_shader_type_get() == PART_REFINE_SHADER_COMPUTE) {
-    GPU_memory_barrier(GPU_BARRIER_SHADER_STORAGE);
+    /* Just render the pass when using compute shaders or transform feedback. */
+    DRW_draw_pass(g_tf_pass);
+    if (drw_curves_shader_type_get() == PART_REFINE_SHADER_COMPUTE) {
+      GPU_memory_barrier(GPU_BARRIER_SHADER_STORAGE);
+    }
+
+    /* Release temporary framebuffer. */
+    if (temp_fb != nullptr) {
+      GPU_framebuffer_free(temp_fb);
+    }
+    /* Rebind existing framebuffer */
+    if (prev_fb != nullptr) {
+      GPU_framebuffer_bind(prev_fb);
+    }
   }
-#endif
 }
 
 void DRW_curves_free()
diff --git a/source/blender/draw/intern/draw_hair.cc b/source/blender/draw/intern/draw_hair.cc
index dc791314333..69f123b95f3 100644
--- a/source/blender/draw/intern/draw_hair.cc
+++ b/source/blender/draw/intern/draw_hair.cc
@@ -22,6 +22,7 @@
 #include "GPU_batch.h"
 #include "GPU_capabilities.h"
 #include "GPU_compute.h"
+#include "GPU_context.h"
 #include "GPU_material.h"
 #include "GPU_shader.h"
 #include "GPU_texture.h"
@@ -33,25 +34,17 @@
 #include "draw_shader.h"
 #include "draw_shader_shared.h"
 
-#ifndef __APPLE__
-#  define USE_TRANSFORM_FEEDBACK
-#  define USE_COMPUTE_SHADERS
-#endif
-
 BLI_INLINE eParticleRefineShaderType drw_hair_shader_type_get()
 {
-#ifdef USE_COMPUTE_SHADERS
   if (GPU_compute_shader_support() && GPU_shader_storage_buffer_objects_support()) {
     return PART_REFINE_SHADER_COMPUTE;
   }
-#endif
-#ifdef USE_TRANSFORM_FEEDBACK
-  return PART_REFINE_SHADER_TRANSFORM_FEEDBACK;
-#endif
+  if (GPU_transform_feedback_support()) {
+    return PART_REFINE_SHADER_TRANSFORM_FEEDBACK;
+  }
   return PART_REFINE_SHADER_TRANSFORM_FEEDBACK_WORKAROUND;
 }
 
-#ifndef USE_TRANSFORM_FEEDBACK
 struct ParticleRefineCall {
   struct ParticleRefineCall *next;
   GPUVertBuf *vbo;
@@ -63,7 +56,6 @@ static ParticleRefineCall *g_tf_calls = nullptr;
 static int g_tf_id_offset;
 static int g_tf_target_width;
 static int g_tf_target_height;
-#endif
 
 static GPUVertBuf *g_dummy_vbo = nullptr;
 static GPUTexture *g_dummy_texture = nullptr;
@@ -77,18 +69,20 @@ static GPUShader *hair_refine_shader_get(ParticleRefineShader refinement)
 
 void DRW_hair_init(void)
 {
-#if defined(USE_TRANSFORM_FEEDBACK) || defined(USE_COMPUTE_SHADERS)
-  g_tf_pass = DRW_pass_create("Update Hair Pass", DRW_STATE_NO_DRAW);
-#else
-  g_tf_pass = DRW_pass_create("Update Hair Pass", DRW_STATE_WRITE_COLOR);
-#endif
+  if (GPU_transform_feedback_support() || GPU_compute_shader_support()) {
+    g_tf_pass = DRW_pass_create("Update Hair Pass", DRW_STATE_NO_DRAW);
+  }
+  else {
+    g_tf_pass = DRW_pass_create("Update Hair Pass", DRW_STATE_WRITE_COLOR);
+  }
 
   if (g_dummy_vbo == nullptr) {
     /* initialize vertex format */
     GPUVertFormat format = {0};
     uint dummy_id = GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_F32, 4, GPU_FETCH_FLOAT);
 
-    g_dummy_vbo = GPU_vertbuf_create_with_format(&format);
+    g_dummy_vbo = GPU_vertbuf_create_with_format_ex(
+        &format, GPU_USAGE_STATIC | GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
 
     const float vert[4] = {0.0f, 0.0f, 0.0f, 0.0f};
     GPU_vertbuf_data_alloc(g_dummy_vbo, 1);
@@ -146,22 +140,25 @@ static void drw_hair_particle_cache_update_transform_feedback(ParticleHairCache
   if (final_points_len > 0) {
     GPUShader *tf_shader = hair_refine_shader_get(PART_REFINE_CATMULL_ROM);
 
-#ifdef USE_TRANSFORM_FEEDBACK
-    DRWShadingGroup *tf_shgrp = DRW_shgroup_transform_feedback_create(
-        tf_shader, g_tf_pass, cache->final[subdiv].proc_buf);
-#else
-    DRWShadingGroup *tf_shgrp = DRW_shgroup_create(tf_shader, g_tf_pass);
-
-    ParticleRefineCall *pr_call = (ParticleRefineCall *)MEM_mallocN(sizeof(*pr_call), __func__);
-    pr_call->next = g_tf_calls;
-    pr_call->vbo = cache->final[subdiv].proc_buf;
-    pr_call->shgrp = tf_shgrp;
-    pr_call->vert_len = final_points_len;
-    g_tf_calls = pr_call;
-    DRW_shgroup_uniform_int(tf_shgrp, "targetHeight", &g_tf_target_height, 1);
-    DRW_shgroup_uniform_int(tf_shgrp, "targetWidth", &g_tf_target_width, 1);
-    DRW_shgroup_uniform_int(tf_shgrp, "idOffset", &g_tf_id_offset, 1);
-#endif
+    DRWShadingGroup *tf_shgrp = nullptr;
+    if (GPU_transform_feedback_support()) {
+      tf_shgrp = DRW_shgroup_transform_feedback_create(
+          tf_shader, g_tf_pass, cache->final[subdiv].proc_buf);
+    }
+    else {
+      tf_shgrp = DRW_shgroup_create(tf_shader, g_tf_pass);
+
+      ParticleRefineCall *pr_call = (ParticleRefineCall *)MEM_mallocN(sizeof(*pr_call), __func__);
+      pr_call->next = g_tf_calls;
+      pr_call->vbo = cache->final[subdiv].proc_buf;
+      pr_call->shgrp = tf_shgrp;
+      pr_call->vert_len = final_points_len;
+      g_tf_calls = pr_call;
+      DRW_shgroup_uniform_int(tf_shgrp, "targetHeight", &g_tf_target_height, 1);
+      DRW_shgroup_uniform_int(tf_shgrp, "targetWidth", &g_tf_target_width, 1);
+      DRW_shgroup_uniform_int(tf_shgrp, "idOffset", &g_tf_id_offset, 1);
+    }
+    BLI_assert(tf_shgrp != nullptr);
 
     drw_hair_particle_cache_shgrp_attach_resources(tf_shgrp, cache, subdiv);
     DRW_shgroup_call_procedural_points(tf_shgrp, nullptr, final_points_len);
@@ -306,81 +303,117 @@ DRWShadingGroup *DRW_shgroup_hair_create_sub(Object *object,
 
 void DRW_hair_update()
 {
-#ifndef USE_TRANSFORM_FEEDBACK
-  /**
-   * Workaround to transform feedback not working on mac.
-   * On some system it crashes (see T58489) and on some other it renders garbage (see T60171).
-   *
-   * So instead of using transform feedback we render to a texture,
-   * read back the result to system memory and re-upload as VBO data.
-   * It is really not ideal performance wise, but it is the simplest
-   * and the most local workaround that still uses the power of the GPU.
-   */
-
-  if (g_tf_calls == nullptr) {
-    return;
-  }
+  if (!GPU_transform_feedback_support()) {
+    /**
+     * Workaround to transform feedback not working on mac.
+     * On some system it crashes (see T58489) and on some other it renders garbage (see T60171).
+     *
+     * So instead of using transform feedback we render to a texture,
+     * read back the result to system memory and re-upload as VBO data.
+     * It is really not ideal performance wise, but it is the simplest
+     * and the most local workaround that still uses the power of the GPU.
+     */
+
+    if (g_tf_calls == nullptr) {
+      return;
+    }
 
-  /* Search ideal buffer size. */
-  uint max_size = 0;
-  for (ParticleRefineCall *pr_call = g_tf_calls; pr_call; pr_call = pr_call->next) {
-    max_size = max_ii(max_size, pr_call->vert_len);
-  }
+    /* Search ideal buffer size. */
+    uint max_size = 0;
+    for (ParticleRefineCall *pr_call = g_tf_calls; pr_call; pr_call = pr_call->next) {
+      max_size = max_ii(max_size, pr_call->vert_len);
+    }
+
+    /* Create target Texture / Frame-buffer */
+    /* Don't use max size as it can be really heavy and fail.
+     * Do chunks of maximum 2048 * 2048 hair points. */
+    int width = 2048;
+    int height = min_ii(width, 1 + max_size / width);
+    GPUTexture *tex = DRW_texture_pool_query_2d(
+        width, height, GPU_RGBA32F, (DrawEngineType *)DRW_hair_update);
+    g_tf_target_height = height;
+    g_tf_target_width = width;
+
+    GPUFrameBuffer *fb = nullptr;
+    GPU_framebuffer_ensure_config(&fb,
+                                  {
+                                      GPU_ATTACHMENT_NONE,
+                                      GPU_ATTACHMENT_TEXTURE(tex),
+                                  });
+
+    float *data = (float *)MEM_mallocN(sizeof(float[4]) * width * height, "tf fallback buffer");
+
+    GPU_framebuffer_bind(fb);
+    while (g_tf_calls != nullptr) {
+      ParticleRefineCall *pr_call = g_tf_calls;
+      g_tf_calls = g_tf_calls->next;
+
+      g_tf_id_offset = 0;
+      while (pr_call->vert_len > 0) {
+        int max_read_px_len = min_ii(width * height, pr_call->vert_len);
+
+        DRW_draw_pass_subset(g_tf_pass, pr_call->shgrp, pr_call->shgrp);
+        /* Read back result to main memory. */
+        GPU_framebuffer_read_color(fb, 0, 0, width, height, 4, 0, GPU_DATA_FLOAT, data);
+        /* Upload back to VBO. */
+        GPU_vertbuf_use(pr_call->vbo);
+        GPU_vertbuf_update_sub(pr_call->vbo,
+                               sizeof(float[4]) * g_tf_id_offset,
+                               sizeof(float[4]) * max_read_px_len,
+                               data);
+
+        g_tf_id_offset += max_read_px_len;
+        pr_call->vert_len -= max_read_px_len;
+      }
 
-  /* Create target Texture / Frame-buffer */
-  /* Don't use max size as it can be really heavy and fail.
-   * Do chunks of maximum 2048 * 2048 hair points. */
-  int width = 2048;
-  int height = min_ii(width, 1 + max_size / width);
-  GPUTexture *tex = DRW_texture_pool_query_2d(
-      width, height, GPU_RGBA32F, (DrawEngineType *)DRW_hair_update);
-  g_tf_target_height = height;
-  g_tf_target_width = width;
-
-  GPUFrameBuffer *fb = nullptr;
-  GPU_framebuffer_ensure_config(&fb,
-                                {
-                                    GPU_ATTACHMENT_NONE,
-                                    GPU_ATTACHMENT_TEXTURE(tex),
-                                });
-
-  float *data = (float *)MEM_mallocN(sizeof(float[4]) * width * height, "tf fallback buffer");
-
-  GPU_framebuffer_bind(fb);
-  while (g_tf_calls != nullptr) {
-    ParticleRefineCall *pr_call = g_tf_calls;
-    g_tf_calls = g_tf_calls->next;
-
-    g_tf_id_offset = 0;
-    while (pr_call->vert_len > 0) {
-      int max_read_px_len = min_ii(width * height, pr_call->vert_len);
-
-      DRW_draw_pass_subset(g_tf_pass, pr_call->shgrp, pr_call->shgrp);
-      /* Read back result to main memory. */
-      GPU_framebuffer_read_color(fb, 0, 0, width, height, 4, 0, GPU_DATA_FLOAT, data);
-      /* Upload back to VBO. */
-      GPU_vertbuf_use(pr_call->vbo);
-      GPU_vertbuf_update_sub(pr_call->vbo,
-                             sizeof(float[4]) * g_tf_id_offset,
-                             sizeof(float[4]) * max_read_px_len,
-                             data);
-
-      g_tf_id_offset += max_read_px_len;
-      pr_call->vert_len -= max_read_px_len;
+      MEM_freeN(pr_call);
     }
 
-    MEM_freeN(pr_call);
+    MEM_freeN(data);
+    GPU_framebuffer_free(fb);
   }
+  else {
+    /* Note(Metal): If compute is not supported, bind a temporary framebuffer to avoid
+     * side-effects from rendering in the active buffer.
+     * We also need to guarantee that a Framebuffer is active to perform any rendering work,
+     * even if there is no output */
+    GPUFrameBuffer *temp_fb = nullptr;
+    GPUFrameBuffer *prev_fb = nullptr;
+    if (GPU_type_matches_ex(GPU_DEVICE_ANY, GPU_OS_MAC, GPU_DRIVER_ANY, GPU_BACKEND_METAL)) {
+      if (!GPU_compute_shader_support()) {
+        prev_fb = GPU_framebuffer_active_get();
+        char errorOut[256];
+        /* if the framebuffer is invalid we need a dummy framebuffer to be bound. */
+        if (!GPU_framebuffer_check_valid(prev_fb, errorOut)) {
+          int width = 64;
+          int height = 64;
+          GPUTexture *tex = DRW_texture_pool_query_2d(
+              width, height, GPU_DEPTH_COMPONENT32F, (DrawEngineType *)DRW_hair_update);
+          g_tf_target_height = height;
+          g_tf_target_width = width;
+
+          GPU_framebuffer_ensure_config(&temp_fb, {GPU_ATTACHMENT_TEXTURE(tex)});
+
+          GPU_framebuffer_bind(temp_fb);
+        }
+      }
+    }
 
-  MEM_freeN(data);
-  GPU_framebuffer_free(fb);
-#else
-  /* Just render the pass when using compute shaders or transform feedback. */
-  DRW_draw_pass(g_tf_pass);
-  if (drw_hair_shader_type_get() == PART_REFINE_SHADER_COMPUTE) {
-    GPU_memory_barrier(GPU_BARRIER_SHADER_STORAGE);
+    /* Just render the pass when using compute shaders or transform feedback. */
+    DRW_draw_pass(g_tf_pass);
+    if (drw_hair_shader_type_get() == PART_REFINE_SHADER_COMPUTE) {
+      GPU_memory_barrier(GPU_BARRIER_SHADER_STORAGE);
+    }
+
+    /* Release temporary framebuffer. */
+    if (temp_fb != nullptr) {
+      GPU_framebuffer_free(temp_fb);
+    }
+    /* Rebind existing framebuffer */
+    if (prev_fb != nullptr) {
+      GPU_framebuffer_bind(prev_fb);
+    }
   }
-#endif
 }
 
 void DRW_hair_free(void)
diff --git a/source/blender/draw/intern/mesh_extractors/extract_mesh_vbo_edge_fac.cc b/source/blender/draw/intern/mesh_extractors/extract_mesh_vbo_edge_fac.cc
index eb6e800023a..50c37f6397c 100644
--- a/source/blender/draw/intern/mesh_extractors/extract_mesh_vbo_edge_fac.cc
+++ b/source/blender/draw/intern/mesh_extractors/extract_mesh_vbo_edge_fac.cc
@@ -174,7 +174,7 @@ static void extract_edge_fac_finish(const MeshRenderData *mr,
   GPUVertBuf *vbo = static_cast<GPUVertBuf *>(buf);
   MeshExtract_EdgeFac_Data *data = static_cast<MeshExtract_EdgeFac_Data *>(_data);
 
-  if (GPU_crappy_amd_driver()) {
+  if (GPU_crappy_amd_driver() || GPU_minimum_per_vertex_stride() > 1) {
     /* Some AMD drivers strangely crash with VBO's with a one byte format.
      * To workaround we reinitialize the VBO with another format and convert
      * all bytes to floats. */
@@ -206,7 +206,7 @@ static GPUVertFormat *get_subdiv_edge_fac_format()
 {
   static GPUVertFormat format = {0};
   if (format.attr_len == 0) {
-    if (GPU_crappy_amd_driver()) {
+    if (GPU_crappy_amd_driver() || GPU_minimum_per_vertex_stride() > 1) {
       GPU_vertformat_attr_add(&format, "wd", GPU_COMP_F32, 1, GPU_FETCH_FLOAT);
     }
     else {
@@ -268,7 +268,7 @@ static void extract_edge_fac_loose_geom_subdiv(const DRWSubdivCache *subdiv_cach
 
   uint offset = subdiv_cache->num_subdiv_loops;
   for (int i = 0; i < loose_geom.edge_len; i++) {
-    if (GPU_crappy_amd_driver()) {
+    if (GPU_crappy_amd_driver() || GPU_minimum_per_vertex_stride() > 1) {
       float loose_edge_fac[2] = {1.0f, 1.0f};
       GPU_vertbuf_update_sub(vbo, offset * sizeof(float), sizeof(loose_edge_fac), loose_edge_fac);
     }
diff --git a/source/blender/gpu/GPU_capabilities.h b/source/blender/gpu/GPU_capabilities.h
index aa01f446b9b..61c60f336e1 100644
--- a/source/blender/gpu/GPU_capabilities.h
+++ b/source/blender/gpu/GPU_capabilities.h
@@ -16,6 +16,7 @@ extern "C" {
 #endif
 
 int GPU_max_texture_size(void);
+int GPU_max_texture_3d_size(void);
 int GPU_max_texture_layers(void);
 int GPU_max_textures(void);
 int GPU_max_textures_vert(void);
@@ -31,6 +32,7 @@ int GPU_max_vertex_attribs(void);
 int GPU_max_varying_floats(void);
 int GPU_max_shader_storage_buffer_bindings(void);
 int GPU_max_compute_shader_storage_blocks(void);
+int GPU_max_samplers(void);
 
 int GPU_extensions_len(void);
 const char *GPU_extension_get(int i);
@@ -57,6 +59,9 @@ void GPU_mem_stats_get(int *totalmem, int *freemem);
  */
 bool GPU_stereo_quadbuffer_support(void);
 
+int GPU_minimum_per_vertex_stride(void);
+bool GPU_transform_feedback_support(void);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/source/blender/gpu/GPU_vertex_buffer.h b/source/blender/gpu/GPU_vertex_buffer.h
index 722ef878271..d3c1bd8145d 100644
--- a/source/blender/gpu/GPU_vertex_buffer.h
+++ b/source/blender/gpu/GPU_vertex_buffer.h
@@ -40,12 +40,20 @@ extern "C" {
 
 typedef enum {
   /* can be extended to support more types */
-  GPU_USAGE_STREAM,
-  GPU_USAGE_STATIC, /* do not keep data in memory */
-  GPU_USAGE_DYNAMIC,
-  GPU_USAGE_DEVICE_ONLY, /* Do not do host->device data transfers. */
+  GPU_USAGE_STREAM = 0,
+  GPU_USAGE_STATIC = 1, /* do not keep data in memory */
+  GPU_USAGE_DYNAMIC = 2,
+  GPU_USAGE_DEVICE_ONLY = 3, /* Do not do host->device data transfers. */
+
+  /** Extended usage flags. */
+  /* Flag for vertex buffers used for textures. Skips additional padding/compaction to ensure
+   * format matches the texture exactly. Can be masked with other properties, and is stripped
+   * during VertBuf::init. */
+  GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY = 1 << 3,
 } GPUUsageType;
 
+ENUM_OPERATORS(GPUUsageType, GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY);
+
 /** Opaque type hiding blender::gpu::VertBuf. */
 typedef struct GPUVertBuf GPUVertBuf;
 
diff --git a/source/blender/gpu/intern/gpu_capabilities.cc b/source/blender/gpu/intern/gpu_capabilities.cc
index 73f94ecfb1b..e584b757a05 100644
--- a/source/blender/gpu/intern/gpu_capabilities.cc
+++ b/source/blender/gpu/intern/gpu_capabilities.cc
@@ -33,6 +33,11 @@ int GPU_max_texture_size()
   return GCaps.max_texture_size;
 }
 
+int GPU_max_texture_3d_size(void)
+{
+  return GCaps.max_texture_3d_size;
+}
+
 int GPU_texture_size_with_limit(int res)
 {
   int size = GPU_max_texture_size();
@@ -115,6 +120,11 @@ const char *GPU_extension_get(int i)
   return GCaps.extension_get ? GCaps.extension_get(i) : "\0";
 }
 
+int GPU_max_samplers()
+{
+  return GCaps.max_samplers;
+}
+
 bool GPU_mip_render_workaround()
 {
   return GCaps.mip_render_workaround;
@@ -176,6 +186,16 @@ int GPU_max_compute_shader_storage_blocks()
   return GCaps.max_compute_shader_storage_blocks;
 }
 
+int GPU_minimum_per_vertex_stride(void)
+{
+  return GCaps.minimum_per_vertex_stride;
+}
+
+bool GPU_transform_feedback_support(void)
+{
+  return GCaps.transform_feedback_support;
+}
+
 /** \} */
 
 /* -------------------------------------------------------------------- */
diff --git a/source/blender/gpu/intern/gpu_texture.cc b/source/blender/gpu/intern/gpu_texture.cc
index 65daa416cae..bec8b8a0df3 100644
--- a/source/blender/gpu/intern/gpu_texture.cc
+++ b/source/blender/gpu/intern/gpu_texture.cc
@@ -360,6 +360,13 @@ GPUTexture *GPU_texture_create_compressed_2d(
 
 GPUTexture *GPU_texture_create_from_vertbuf(const char *name, GPUVertBuf *vert)
 {
+#ifndef NDEBUG
+  /* Vertex buffers used for texture buffers must be flagged with:
+   * GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY. */
+  BLI_assert_msg(unwrap(vert)->extended_usage_ & GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY,
+                 "Vertex Buffers used for textures should have usage flag "
+                 "GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY.");
+#endif
   eGPUTextureFormat tex_format = to_texture_format(GPU_vertbuf_get_format(vert));
   Texture *tex = GPUBackend::get()->texture_alloc(name);
 
diff --git a/source/blender/gpu/intern/gpu_vertex_buffer.cc b/source/blender/gpu/intern/gpu_vertex_buffer.cc
index 0dbd565291b..a441cfe2fb8 100644
--- a/source/blender/gpu/intern/gpu_vertex_buffer.cc
+++ b/source/blender/gpu/intern/gpu_vertex_buffer.cc
@@ -40,10 +40,21 @@ VertBuf::~VertBuf()
 
 void VertBuf::init(const GPUVertFormat *format, GPUUsageType usage)
 {
-  usage_ = usage;
+  /* Strip extended usage flags. */
+  usage_ = usage & ~GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY;
+#ifndef NDEBUG
+  /* Store extended usage. */
+  extended_usage_ = usage;
+#endif
   flag = GPU_VERTBUF_DATA_DIRTY;
   GPU_vertformat_copy(&this->format, format);
-  if (!format->packed) {
+  /* Avoid packing vertex formats which are used for texture buffers.
+   * These cases use singular types and do not need packing. They must
+   * also not have increased alignment padding to the minimum per-vertex stride. */
+  if (usage & GPU_USAGE_FLAG_BUFFER_TEXTURE_ONLY) {
+    VertexFormat_texture_buffer_pack(&this->format);
+  }
+  if (!this->format.packed) {
     VertexFormat_pack(&this->format);
   }
   flag |= GPU_VERTBUF_INIT;
@@ -62,6 +73,10 @@ VertBuf *VertBuf::duplicate()
   *dst = *this;
   /* Almost full copy... */
   dst->handle_refcount_ = 1;
+  /* Metadata. */
+#ifndef NDEBUG
+  dst->extended_usage_ = extended_usage_;
+#endif
   /* Duplicate all needed implementation specifics data. */
   this->duplicate_data(dst);
   return dst;
@@ -192,6 +207,7 @@ void GPU_vertbuf_data_len_set(GPUVertBuf *verts_, uint v_len)
 void GPU_vertbuf_attr_set(GPUVertBuf *verts_, uint a_idx, uint v_idx, const void *data)
 {
   VertBuf *verts = unwrap(verts_);
+  BLI_assert(verts->get_usage_type() != GPU_USAGE_DEVICE_ONLY);
   const GPUVertFormat *format = &verts->format;
   const GPUVertAttr *a = &format->attrs[a_idx];
   BLI_assert(v_idx < verts->vertex_alloc);
@@ -215,6 +231,7 @@ void GPU_vertbuf_attr_fill(GPUVertBuf *verts_, uint a_idx, const void *data)
 void GPU_vertbuf_vert_set(GPUVertBuf *verts_, uint v_idx, const void *data)
 {
   VertBuf *verts = unwrap(verts_);
+  BLI_assert(verts->get_usage_type() != GPU_USAGE_DEVICE_ONLY);
   const GPUVertFormat *format = &verts->format;
   BLI_assert(v_idx < verts->vertex_alloc);
   BLI_assert(verts->data != nullptr);
@@ -225,6 +242,7 @@ void GPU_vertbuf_vert_set(GPUVertBuf *verts_, uint v_idx, const void *data)
 void GPU_vertbuf_attr_fill_stride(GPUVertBuf *verts_, uint a_idx, uint stride, const void *data)
 {
   VertBuf *verts = unwrap(verts_);
+  BLI_assert(verts->get_usage_type() != GPU_USAGE_DEVICE_ONLY);
   const GPUVertFormat *format = &verts->format;
   const GPUVertAttr *a = &format->attrs[a_idx];
   BLI_assert(a_idx < format->attr_len);
diff --git a/source/blender/gpu/intern/gpu_vertex_buffer_private.hh b/source/blender/gpu/intern/gpu_vertex_buffer_private.hh
index a7920bacaec..f20f6caf6de 100644
--- a/source/blender/gpu/intern/gpu_vertex_buffer_private.hh
+++ b/source/blender/gpu/intern/gpu_vertex_buffer_private.hh
@@ -31,6 +31,11 @@ class VertBuf {
   /** NULL indicates data in VRAM (unmapped) */
   uchar *data = nullptr;
 
+#ifndef NDEBUG
+  /** Usage including extended usage flags. */
+  GPUUsageType extended_usage_ = GPU_USAGE_STATIC;
+#endif
+
  protected:
   /** Usage hint for GL optimization. */
   GPUUsageType usage_ = GPU_USAGE_STATIC;
@@ -83,6 +88,11 @@ class VertBuf {
     }
   }
 
+  GPUUsageType get_usage_type() const
+  {
+    return usage_;
+  }
+
   virtual void update_sub(uint start, uint len, const void *data) = 0;
   virtual const void *read() const = 0;
   virtual void *unmap(const void *mapped_data) const = 0;
diff --git a/source/blender/gpu/intern/gpu_vertex_format.cc b/source/blender/gpu/intern/gpu_vertex_format.cc
index c16c06c1421..5c21bc313eb 100644
--- a/source/blender/gpu/intern/gpu_vertex_format.cc
+++ b/source/blender/gpu/intern/gpu_vertex_format.cc
@@ -8,6 +8,8 @@
  */
 
 #include "GPU_vertex_format.h"
+#include "GPU_capabilities.h"
+
 #include "gpu_shader_create_info.hh"
 #include "gpu_shader_private.hh"
 #include "gpu_vertex_format_private.h"
@@ -68,7 +70,7 @@ static uint attr_size(const GPUVertAttr *a)
   return a->comp_len * comp_size(static_cast<GPUVertCompType>(a->comp_type));
 }
 
-static uint attr_align(const GPUVertAttr *a)
+static uint attr_align(const GPUVertAttr *a, uint minimum_stride)
 {
   if (a->comp_type == GPU_COMP_I10) {
     return 4; /* always packed as 10_10_10_2 */
@@ -78,7 +80,10 @@ static uint attr_align(const GPUVertAttr *a)
     return 4 * c; /* AMD HW can't fetch these well, so pad it out (other vendors too?) */
   }
 
-  return c; /* most fetches are ok if components are naturally aligned */
+  /* Most fetches are ok if components are naturally aligned.
+   * However, in Metal,the minimum supported per-vertex stride is 4,
+   * so we must query the GPU and pad out the size accordingly. */
+  return max_ii(minimum_stride, c);
 }
 
 uint vertex_buffer_size(const GPUVertFormat *format, uint vertex_len)
@@ -308,7 +313,7 @@ static void show_pack(uint a_idx, uint size, uint pad)
 }
 #endif
 
-void VertexFormat_pack(GPUVertFormat *format)
+static void VertexFormat_pack_impl(GPUVertFormat *format, uint minimum_stride)
 {
   GPUVertAttr *a0 = &format->attrs[0];
   a0->offset = 0;
@@ -320,7 +325,7 @@ void VertexFormat_pack(GPUVertFormat *format)
 
   for (uint a_idx = 1; a_idx < format->attr_len; a_idx++) {
     GPUVertAttr *a = &format->attrs[a_idx];
-    uint mid_padding = padding(offset, attr_align(a));
+    uint mid_padding = padding(offset, attr_align(a, minimum_stride));
     offset += mid_padding;
     a->offset = offset;
     offset += a->size;
@@ -330,7 +335,7 @@ void VertexFormat_pack(GPUVertFormat *format)
 #endif
   }
 
-  uint end_padding = padding(offset, attr_align(a0));
+  uint end_padding = padding(offset, attr_align(a0, minimum_stride));
 
 #if PACK_DEBUG
   show_pack(0, 0, end_padding);
diff --git a/source/blender/gpu/intern/gpu_vertex_format_private.h b/source/blender/gpu/intern/gpu_vertex_format_private.h
index 0f8a869f6df..430008b4cb9 100644
--- a/source/blender/gpu/intern/gpu_vertex_format_private.h
+++ b/source/blender/gpu/intern/gpu_vertex_format_private.h
@@ -16,6 +16,7 @@ extern "C" {
 struct GPUVertFormat;
 
 void VertexFormat_pack(struct GPUVertFormat *format);
+void VertexFormat_texture_buffer_pack(struct GPUVertFormat *format);
 uint padding(uint offset, uint alignment);
 uint vertex_buffer_size(const struct GPUVertFormat *format, uint vertex_len);
 
diff --git a/source/blender/gpu/opengl/gl_backend.cc b/source/blender/gpu/opengl/gl_backend.cc
index 24ca8c25bc0..2375e78d9f1 100644
--- a/source/blender/gpu/opengl/gl_backend.cc
+++ b/source/blender/gpu/opengl/gl_backend.cc
@@ -386,6 +386,11 @@ static void detect_workarounds()
     }
   }
 
+  /* Disable TF on macOS. */
+  if (GPU_type_matches(GPU_DEVICE_ANY, GPU_OS_MAC, GPU_DRIVER_ANY)) {
+    GCaps.transform_feedback_support = false;
+  }
+
   /* Some Intel drivers have issues with using mips as frame-buffer targets if
    * GL_TEXTURE_MAX_LEVEL is higher than the target MIP.
    * Only check at the end after all other workarounds because this uses the drawing code.
@@ -431,7 +436,6 @@ static void detect_workarounds()
 /** Internal capabilities. */
 
 GLint GLContext::max_cubemap_size = 0;
-GLint GLContext::max_texture_3d_size = 0;
 GLint GLContext::max_ubo_binds = 0;
 GLint GLContext::max_ubo_size = 0;
 GLint GLContext::max_ssbo_binds = 0;
@@ -499,6 +503,8 @@ void GLBackend::capabilities_init()
   GCaps.shader_draw_parameters_support = epoxy_has_gl_extension("GL_ARB_shader_draw_parameters");
   GCaps.compute_shader_support = epoxy_has_gl_extension("GL_ARB_compute_shader") &&
                                  epoxy_gl_version() >= 43;
+  GCaps.max_samplers = GCaps.max_textures;
+
   if (GCaps.compute_shader_support) {
     glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, 0, &GCaps.max_work_group_count[0]);
     glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, 1, &GCaps.max_work_group_count[1]);
@@ -512,8 +518,10 @@ void GLBackend::capabilities_init()
   }
   GCaps.shader_storage_buffer_objects_support = epoxy_has_gl_extension(
       "GL_ARB_shader_storage_buffer_object");
+  GCaps.transform_feedback_support = true;
+
   /* GL specific capabilities. */
-  glGetIntegerv(GL_MAX_3D_TEXTURE_SIZE, &GLContext::max_texture_3d_size);
+  glGetIntegerv(GL_MAX_3D_TEXTURE_SIZE, &GCaps.max_texture_3d_size);
   glGetIntegerv(GL_MAX_CUBE_MAP_TEXTURE_SIZE, &GLContext::max_cubemap_size);
   glGetIntegerv(GL_MAX_FRAGMENT_UNIFORM_BLOCKS, &GLContext::max_ubo_binds);
   glGetIntegerv(GL_MAX_UNIFORM_BLOCK_SIZE, &GLContext::max_ubo_size);
diff --git a/source/blender/gpu/opengl/gl_context.hh b/source/blender/gpu/opengl/gl_context.hh
index 2f8c2b762f8..1d413750fd4 100644
--- a/source/blender/gpu/opengl/gl_context.hh
+++ b/source/blender/gpu/opengl/gl_context.hh
@@ -40,7 +40,6 @@ class GLContext : public Context {
   /** Capabilities. */
 
   static GLint max_cubemap_size;
-  static GLint max_texture_3d_size;
   static GLint max_ubo_size;
   static GLint max_ubo_binds;
   static GLint max_ssbo_size;
diff --git a/source/blender/gpu/opengl/gl_texture.cc b/source/blender/gpu/opengl/gl_texture.cc
index 2ce205353a3..02fc7dfdb5f 100644
--- a/source/blender/gpu/opengl/gl_texture.cc
+++ b/source/blender/gpu/opengl/gl_texture.cc
@@ -603,7 +603,7 @@ bool GLTexture::proxy_check(int mip)
 {
   /* Manual validation first, since some implementation have issues with proxy creation. */
   int max_size = GPU_max_texture_size();
-  int max_3d_size = GLContext::max_texture_3d_size;
+  int max_3d_size = GPU_max_texture_3d_size();
   int max_cube_size = GLContext::max_cubemap_size;
   int size[3] = {1, 1, 1};
   this->mip_size_get(mip, size);