radeonsi: increase NGG workgroup size to 256 for VS/TES with streamout and GS

NGG streamout performance is limited by the workgroup size, so make it as
large as possible.

Since this uses si_get_max_workgroup_size() to set the NGG workgroup size,
the side effect is that all GS is also getting an increase to 256, which
is OK.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21403>

diff --git a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
index 6fe8b1d..2880f32 100644 (file)
--- a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
+++ b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
@@ -117,8 +117,9 @@ bool gfx10_ngg_calculate_subgroup_info(struct si_shader *shader)
       gs_sel->screen->info.gfx_level >= GFX11 ? 3 : /* gfx11 requires at least 1 primitive per TG */
       gs_sel->screen->info.gfx_level >= GFX10_3 ? 29 : (24 - 1 + max_verts_per_prim);
    bool max_vert_out_per_gs_instance = false;
-   unsigned max_gsprims_base = gs_sel->screen->ngg_subgroup_size; /* default prim group size clamp */
-   unsigned max_esverts_base = gs_sel->screen->ngg_subgroup_size;
+   unsigned max_gsprims_base, max_esverts_base;
+
+   max_gsprims_base = max_esverts_base = si_get_max_workgroup_size(shader);
 
    if (gs_stage == MESA_SHADER_GEOMETRY) {
       bool force_multi_cycling = false;

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 06b13ba..4dfbb0d 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -1413,8 +1413,6 @@ static struct pipe_screen *radeonsi_screen_create_impl(struct radeon_winsys *ws,
       }
    }
 
-   sscreen->ngg_subgroup_size = 128;
-
    if (sscreen->info.gfx_level >= GFX11) {
       unsigned attr_ring_size = sscreen->info.attribute_ring_size_per_se * sscreen->info.max_se;
       sscreen->attribute_ring = si_aligned_buffer_create(&sscreen->b,

diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 368e3ab..3c3ad6a 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -694,8 +694,6 @@ struct si_screen {
     * We want to minimize the impact on multithreaded Mesa. */
    struct ac_llvm_compiler compiler_lowp[10];
 
-   unsigned ngg_subgroup_size;
-
    struct util_idalloc_mt buffer_ids;
    struct util_vertex_state_cache vertex_state_cache;
 

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 095ed12..ac97f71 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -206,7 +206,11 @@ unsigned si_get_max_workgroup_size(const struct si_shader *shader)
    switch (shader->selector->stage) {
    case MESA_SHADER_VERTEX:
    case MESA_SHADER_TESS_EVAL:
-      return shader->key.ge.as_ngg ? shader->selector->screen->ngg_subgroup_size : 0;
+      /* Use the largest workgroup size for streamout */
+      if (shader->key.ge.as_ngg)
+         return si_shader_uses_streamout(shader) ? 256 : 128;
+      else
+         return 0;
 
    case MESA_SHADER_TESS_CTRL:
       /* Return this so that LLVM doesn't remove s_barrier
@@ -214,7 +218,7 @@ unsigned si_get_max_workgroup_size(const struct si_shader *shader)
       return shader->selector->screen->info.gfx_level >= GFX7 ? 128 : 0;
 
    case MESA_SHADER_GEOMETRY:
-      /* ngg_subgroup_size is only the input size. GS can always generate up to 256 vertices. */
+      /* GS can always generate up to 256 vertices. */
       return shader->selector->screen->info.gfx_level >= GFX9 ? 256 : 0;
 
    case MESA_SHADER_COMPUTE:

diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index c4f7cda..a01944d 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -1058,7 +1058,7 @@ static inline bool gfx10_ngg_writes_user_edgeflags(struct si_shader *shader)
           shader->selector->info.writes_edgeflag;
 }
 
-static inline bool si_shader_uses_streamout(struct si_shader *shader)
+static inline bool si_shader_uses_streamout(const struct si_shader *shader)
 {
    return shader->selector->stage <= MESA_SHADER_GEOMETRY &&
           shader->selector->info.enabled_streamout_buffer_mask &&