radeonsi: remove gfx10 NGG streamout

Unused and unstable. Keep it only for gfx11.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23687>

diff --git a/src/gallium/drivers/radeonsi/gfx10_query.c b/src/gallium/drivers/radeonsi/gfx11_query.c
similarity index 78%
rename from src/gallium/drivers/radeonsi/gfx10_query.c
rename to src/gallium/drivers/radeonsi/gfx11_query.c
index 8040b67..bfcd8e2 100644 (file)
--- a/src/gallium/drivers/radeonsi/gfx10_query.c
+++ b/src/gallium/drivers/radeonsi/gfx11_query.c
@@ -16,19 +16,19 @@ static void emit_shader_query(struct si_context *sctx)
 {
    assert(!list_is_empty(&sctx->shader_query_buffers));
 
-   struct gfx10_sh_query_buffer *qbuf =
-      list_last_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
-   qbuf->head += sizeof(struct gfx10_sh_query_buffer_mem);
+   struct gfx11_sh_query_buffer *qbuf =
+      list_last_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
+   qbuf->head += sizeof(struct gfx11_sh_query_buffer_mem);
 }
 
-static void gfx10_release_query_buffers(struct si_context *sctx,
-                                        struct gfx10_sh_query_buffer *first,
-                                        struct gfx10_sh_query_buffer *last)
+static void gfx11_release_query_buffers(struct si_context *sctx,
+                                        struct gfx11_sh_query_buffer *first,
+                                        struct gfx11_sh_query_buffer *last)
 {
    while (first) {
-      struct gfx10_sh_query_buffer *qbuf = first;
+      struct gfx11_sh_query_buffer *qbuf = first;
       if (first != last)
-         first = list_entry(qbuf->list.next, struct gfx10_sh_query_buffer, list);
+         first = list_entry(qbuf->list.next, struct gfx11_sh_query_buffer, list);
       else
          first = NULL;
 
@@ -47,19 +47,19 @@ static void gfx10_release_query_buffers(struct si_context *sctx,
    }
 }
 
-static bool gfx10_alloc_query_buffer(struct si_context *sctx)
+static bool gfx11_alloc_query_buffer(struct si_context *sctx)
 {
    if (si_is_atom_dirty(sctx, &sctx->atoms.s.shader_query))
       return true;
 
-   struct gfx10_sh_query_buffer *qbuf = NULL;
+   struct gfx11_sh_query_buffer *qbuf = NULL;
 
    if (!list_is_empty(&sctx->shader_query_buffers)) {
-      qbuf = list_last_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
-      if (qbuf->head + sizeof(struct gfx10_sh_query_buffer_mem) <= qbuf->buf->b.b.width0)
+      qbuf = list_last_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
+      if (qbuf->head + sizeof(struct gfx11_sh_query_buffer_mem) <= qbuf->buf->b.b.width0)
          goto success;
 
-      qbuf = list_first_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
+      qbuf = list_first_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
       if (!qbuf->refcount &&
           !si_cs_is_buffer_referenced(sctx, qbuf->buf->buf, RADEON_USAGE_READWRITE) &&
           sctx->ws->buffer_wait(sctx->ws, qbuf->buf->buf, 0, RADEON_USAGE_READWRITE)) {
@@ -71,13 +71,13 @@ static bool gfx10_alloc_query_buffer(struct si_context *sctx)
    }
 
    if (!qbuf) {
-      qbuf = CALLOC_STRUCT(gfx10_sh_query_buffer);
+      qbuf = CALLOC_STRUCT(gfx11_sh_query_buffer);
       if (unlikely(!qbuf))
          return false;
 
       struct si_screen *screen = sctx->screen;
       unsigned buf_size =
-         MAX2(sizeof(struct gfx10_sh_query_buffer_mem), screen->info.min_alloc_size);
+         MAX2(sizeof(struct gfx11_sh_query_buffer_mem), screen->info.min_alloc_size);
       qbuf->buf = si_resource(pipe_buffer_create(&screen->b, 0, PIPE_USAGE_STAGING, buf_size));
       if (unlikely(!qbuf->buf)) {
          FREE(qbuf);
@@ -94,7 +94,7 @@ static bool gfx10_alloc_query_buffer(struct si_context *sctx)
                                             PIPE_MAP_WRITE | PIPE_MAP_UNSYNCHRONIZED);
    assert(results);
 
-   for (unsigned i = 0, e = qbuf->buf->b.b.width0 / sizeof(struct gfx10_sh_query_buffer_mem); i < e;
+   for (unsigned i = 0, e = qbuf->buf->b.b.width0 / sizeof(struct gfx11_sh_query_buffer_mem); i < e;
         ++i) {
       for (unsigned j = 0; j < 16; ++j)
          results[32 * i + j] = (uint64_t)1 << 63;
@@ -109,7 +109,7 @@ success:;
    struct pipe_shader_buffer sbuf;
    sbuf.buffer = &qbuf->buf->b.b;
    sbuf.buffer_offset = qbuf->head;
-   sbuf.buffer_size = sizeof(struct gfx10_sh_query_buffer_mem);
+   sbuf.buffer_size = sizeof(struct gfx11_sh_query_buffer_mem);
    si_set_internal_shader_buffer(sctx, SI_GS_QUERY_BUF, &sbuf);
    SET_FIELD(sctx->current_gs_state, GS_STATE_STREAMOUT_QUERY_ENABLED, 1);
 
@@ -117,24 +117,24 @@ success:;
    return true;
 }
 
-static void gfx10_sh_query_destroy(struct si_context *sctx, struct si_query *rquery)
+static void gfx11_sh_query_destroy(struct si_context *sctx, struct si_query *rquery)
 {
-   struct gfx10_sh_query *query = (struct gfx10_sh_query *)rquery;
-   gfx10_release_query_buffers(sctx, query->first, query->last);
+   struct gfx11_sh_query *query = (struct gfx11_sh_query *)rquery;
+   gfx11_release_query_buffers(sctx, query->first, query->last);
    FREE(query);
 }
 
-static bool gfx10_sh_query_begin(struct si_context *sctx, struct si_query *rquery)
+static bool gfx11_sh_query_begin(struct si_context *sctx, struct si_query *rquery)
 {
-   struct gfx10_sh_query *query = (struct gfx10_sh_query *)rquery;
+   struct gfx11_sh_query *query = (struct gfx11_sh_query *)rquery;
 
-   gfx10_release_query_buffers(sctx, query->first, query->last);
+   gfx11_release_query_buffers(sctx, query->first, query->last);
    query->first = query->last = NULL;
 
-   if (unlikely(!gfx10_alloc_query_buffer(sctx)))
+   if (unlikely(!gfx11_alloc_query_buffer(sctx)))
       return false;
 
-   query->first = list_last_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
+   query->first = list_last_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
    query->first_begin = query->first->head;
 
    sctx->num_active_shader_queries++;
@@ -143,21 +143,21 @@ static bool gfx10_sh_query_begin(struct si_context *sctx, struct si_query *rquer
    return true;
 }
 
-static bool gfx10_sh_query_end(struct si_context *sctx, struct si_query *rquery)
+static bool gfx11_sh_query_end(struct si_context *sctx, struct si_query *rquery)
 {
-   struct gfx10_sh_query *query = (struct gfx10_sh_query *)rquery;
+   struct gfx11_sh_query *query = (struct gfx11_sh_query *)rquery;
 
    if (unlikely(!query->first))
       return false; /* earlier out of memory error */
 
-   query->last = list_last_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
+   query->last = list_last_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
    query->last_end = query->last->head;
 
    /* Signal the fence of the previous chunk */
    if (query->last_end != 0) {
       uint64_t fence_va = query->last->buf->gpu_address;
-      fence_va += query->last_end - sizeof(struct gfx10_sh_query_buffer_mem);
-      fence_va += offsetof(struct gfx10_sh_query_buffer_mem, fence);
+      fence_va += query->last_end - sizeof(struct gfx11_sh_query_buffer_mem);
+      fence_va += offsetof(struct gfx11_sh_query_buffer_mem, fence);
       si_cp_release_mem(sctx, &sctx->gfx_cs, V_028A90_BOTTOM_OF_PIPE_TS, 0, EOP_DST_SEL_MEM,
                         EOP_INT_SEL_NONE, EOP_DATA_SEL_VALUE_32BIT, query->last->buf, fence_va,
                         0xffffffff, PIPE_QUERY_GPU_FINISHED);
@@ -178,8 +178,8 @@ static bool gfx10_sh_query_end(struct si_context *sctx, struct si_query *rquery)
    return true;
 }
 
-static void gfx10_sh_query_add_result(struct gfx10_sh_query *query,
-                                      struct gfx10_sh_query_buffer_mem *qmem,
+static void gfx11_sh_query_add_result(struct gfx11_sh_query *query,
+                                      struct gfx11_sh_query_buffer_mem *qmem,
                                       union pipe_query_result *result)
 {
    static const uint64_t mask = ((uint64_t)1 << 63) - 1;
@@ -212,10 +212,10 @@ static void gfx10_sh_query_add_result(struct gfx10_sh_query *query,
    }
 }
 
-static bool gfx10_sh_query_get_result(struct si_context *sctx, struct si_query *rquery, bool wait,
+static bool gfx11_sh_query_get_result(struct si_context *sctx, struct si_query *rquery, bool wait,
                                       union pipe_query_result *result)
 {
-   struct gfx10_sh_query *query = (struct gfx10_sh_query *)rquery;
+   struct gfx11_sh_query *query = (struct gfx11_sh_query *)rquery;
 
    util_query_clear_result(result, query->b.type);
 
@@ -223,8 +223,8 @@ static bool gfx10_sh_query_get_result(struct si_context *sctx, struct si_query *
       return false; /* earlier out of memory error */
    assert(query->last);
 
-   for (struct gfx10_sh_query_buffer *qbuf = query->last;;
-        qbuf = list_entry(qbuf->list.prev, struct gfx10_sh_query_buffer, list)) {
+   for (struct gfx11_sh_query_buffer *qbuf = query->last;;
+        qbuf = list_entry(qbuf->list.prev, struct gfx11_sh_query_buffer, list)) {
       unsigned usage = PIPE_MAP_READ | (wait ? 0 : PIPE_MAP_DONTBLOCK);
       void *map;
 
@@ -244,10 +244,10 @@ static bool gfx10_sh_query_get_result(struct si_context *sctx, struct si_query *
          results_end = query->last_end;
 
       while (results_begin != results_end) {
-         struct gfx10_sh_query_buffer_mem *qmem = map + results_begin;
+         struct gfx11_sh_query_buffer_mem *qmem = map + results_begin;
          results_begin += sizeof(*qmem);
 
-         gfx10_sh_query_add_result(query, qmem, result);
+         gfx11_sh_query_add_result(query, qmem, result);
       }
 
       if (qbuf == query->first)
@@ -257,19 +257,19 @@ static bool gfx10_sh_query_get_result(struct si_context *sctx, struct si_query *
    return true;
 }
 
-static void gfx10_sh_query_get_result_resource(struct si_context *sctx, struct si_query *rquery,
+static void gfx11_sh_query_get_result_resource(struct si_context *sctx, struct si_query *rquery,
                                                enum pipe_query_flags flags,
                                                enum pipe_query_value_type result_type,
                                                int index, struct pipe_resource *resource,
                                                unsigned offset)
 {
-   struct gfx10_sh_query *query = (struct gfx10_sh_query *)rquery;
+   struct gfx11_sh_query *query = (struct gfx11_sh_query *)rquery;
    struct si_qbo_state saved_state = {};
    struct pipe_resource *tmp_buffer = NULL;
    unsigned tmp_buffer_offset = 0;
 
    if (!sctx->sh_query_result_shader) {
-      sctx->sh_query_result_shader = gfx10_create_sh_query_result_cs(sctx);
+      sctx->sh_query_result_shader = gfx11_create_sh_query_result_cs(sctx);
       if (!sctx->sh_query_result_shader)
          return;
    }
@@ -345,7 +345,7 @@ static void gfx10_sh_query_get_result_resource(struct si_context *sctx, struct s
    grid.grid[1] = 1;
    grid.grid[2] = 1;
 
-   struct gfx10_sh_query_buffer *qbuf = query->first;
+   struct gfx11_sh_query_buffer *qbuf = query->first;
    for (;;) {
       unsigned begin = qbuf == query->first ? query->first_begin : 0;
       unsigned end = qbuf == query->last ? query->last_end : qbuf->buf->b.b.width0;
@@ -356,7 +356,7 @@ static void gfx10_sh_query_get_result_resource(struct si_context *sctx, struct s
       ssbo[0].buffer_offset = begin;
       ssbo[0].buffer_size = end - begin;
 
-      consts.result_count = (end - begin) / sizeof(struct gfx10_sh_query_buffer_mem);
+      consts.result_count = (end - begin) / sizeof(struct gfx11_sh_query_buffer_mem);
       consts.chain = 0;
       if (qbuf != query->first)
          consts.chain |= 1;
@@ -379,8 +379,8 @@ static void gfx10_sh_query_get_result_resource(struct si_context *sctx, struct s
           * serialized in the CP.
           */
          va = qbuf->buf->gpu_address;
-         va += end - sizeof(struct gfx10_sh_query_buffer_mem);
-         va += offsetof(struct gfx10_sh_query_buffer_mem, fence);
+         va += end - sizeof(struct gfx11_sh_query_buffer_mem);
+         va += offsetof(struct gfx11_sh_query_buffer_mem, fence);
 
          si_cp_wait_mem(sctx, &sctx->gfx_cs, va, 0x00000001, 0x00000001, 0);
       }
@@ -393,49 +393,49 @@ static void gfx10_sh_query_get_result_resource(struct si_context *sctx, struct s
 
       if (qbuf == query->last)
          break;
-      qbuf = list_entry(qbuf->list.next, struct gfx10_sh_query_buffer, list);
+      qbuf = list_entry(qbuf->list.next, struct gfx11_sh_query_buffer, list);
    }
 
    si_restore_qbo_state(sctx, &saved_state);
    pipe_resource_reference(&tmp_buffer, NULL);
 }
 
-static const struct si_query_ops gfx10_sh_query_ops = {
-   .destroy = gfx10_sh_query_destroy,
-   .begin = gfx10_sh_query_begin,
-   .end = gfx10_sh_query_end,
-   .get_result = gfx10_sh_query_get_result,
-   .get_result_resource = gfx10_sh_query_get_result_resource,
+static const struct si_query_ops gfx11_sh_query_ops = {
+   .destroy = gfx11_sh_query_destroy,
+   .begin = gfx11_sh_query_begin,
+   .end = gfx11_sh_query_end,
+   .get_result = gfx11_sh_query_get_result,
+   .get_result_resource = gfx11_sh_query_get_result_resource,
 };
 
-struct pipe_query *gfx10_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type,
+struct pipe_query *gfx11_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type,
                                          unsigned index)
 {
-   struct gfx10_sh_query *query = CALLOC_STRUCT(gfx10_sh_query);
+   struct gfx11_sh_query *query = CALLOC_STRUCT(gfx11_sh_query);
    if (unlikely(!query))
       return NULL;
 
-   query->b.ops = &gfx10_sh_query_ops;
+   query->b.ops = &gfx11_sh_query_ops;
    query->b.type = query_type;
    query->stream = index;
 
    return (struct pipe_query *)query;
 }
 
-void gfx10_init_query(struct si_context *sctx)
+void gfx11_init_query(struct si_context *sctx)
 {
    list_inithead(&sctx->shader_query_buffers);
    sctx->atoms.s.shader_query.emit = emit_shader_query;
 }
 
-void gfx10_destroy_query(struct si_context *sctx)
+void gfx11_destroy_query(struct si_context *sctx)
 {
    if (!sctx->shader_query_buffers.next)
       return;
 
    while (!list_is_empty(&sctx->shader_query_buffers)) {
-      struct gfx10_sh_query_buffer *qbuf =
-         list_first_entry(&sctx->shader_query_buffers, struct gfx10_sh_query_buffer, list);
+      struct gfx11_sh_query_buffer *qbuf =
+         list_first_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
       list_del(&qbuf->list);
 
       assert(!qbuf->refcount);

diff --git a/src/gallium/drivers/radeonsi/meson.build b/src/gallium/drivers/radeonsi/meson.build
index a057027..1d80ee3 100644 (file)
--- a/src/gallium/drivers/radeonsi/meson.build
+++ b/src/gallium/drivers/radeonsi/meson.build
@@ -20,8 +20,8 @@
 
 files_libradeonsi = files(
   'driinfo_radeonsi.h',
-  'gfx10_query.c',
   'gfx10_shader_ngg.c',
+  'gfx11_query.c',
   'si_blit.c',
   'si_buffer.c',
   'si_build_pm4.h',

diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index 75db358..8aa969d 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -85,12 +85,8 @@ void si_flush_gfx_cs(struct si_context *ctx, unsigned flags, struct pipe_fence_h
           * idle when we leave the IB, otherwise another process
           * might overwrite it while our shaders are busy.
           */
-         if (sscreen->use_ngg_streamout) {
-            if (ctx->gfx_level >= GFX11)
-               wait_flags |= SI_CONTEXT_VS_PARTIAL_FLUSH;
-            else
-               wait_flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
-         }
+         if (ctx->gfx_level >= GFX11)
+            wait_flags |= SI_CONTEXT_VS_PARTIAL_FLUSH;
       }
    }
 
@@ -194,8 +190,6 @@ static void si_begin_gfx_cs_debug(struct si_context *ctx)
 
 static void si_add_gds_to_buffer_list(struct si_context *sctx)
 {
-   if (sctx->screen->gds)
-      sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->screen->gds, RADEON_USAGE_READWRITE, 0);
    if (sctx->screen->gds_oa)
       sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->screen->gds_oa, RADEON_USAGE_READWRITE, 0);
 }
@@ -204,25 +198,16 @@ void si_allocate_gds(struct si_context *sctx)
 {
    struct radeon_winsys *ws = sctx->ws;
 
-   assert(sctx->screen->use_ngg_streamout);
+   assert(sctx->gfx_level >= GFX11);
 
    if (sctx->screen->gds_oa)
       return;
 
-   assert(!sctx->screen->gds && !sctx->screen->gds_oa);
-
-   /* Gfx11 only uses GDS OA, not GDS memory.
-    * Gfx10 needs 256B (64 dw) of GDS, otherwise streamout hangs.
-    */
+   /* Gfx11 only uses GDS OA, not GDS memory. */
    simple_mtx_lock(&sctx->screen->gds_mutex);
    if (!sctx->screen->gds_oa) {
       sctx->screen->gds_oa = ws->buffer_create(ws, 1, 1, RADEON_DOMAIN_OA, RADEON_FLAG_DRIVER_INTERNAL);
       assert(sctx->screen->gds_oa);
-
-      if (sctx->gfx_level < GFX11) {
-         sctx->screen->gds = ws->buffer_create(ws, 256, 4, RADEON_DOMAIN_GDS, RADEON_FLAG_DRIVER_INTERNAL);
-         assert(sctx->screen->gds);
-      }
    }
    simple_mtx_unlock(&sctx->screen->gds_mutex);
 
@@ -514,7 +499,7 @@ void si_begin_new_gfx_cs(struct si_context *ctx, bool first_cs)
          si_mark_atom_dirty(ctx, &ctx->atoms.s.dpbb_state);
       si_mark_atom_dirty(ctx, &ctx->atoms.s.stencil_ref);
       si_mark_atom_dirty(ctx, &ctx->atoms.s.spi_map);
-      if (!ctx->screen->use_ngg_streamout)
+      if (ctx->gfx_level < GFX11)
          si_mark_atom_dirty(ctx, &ctx->atoms.s.streamout_enable);
       /* CLEAR_STATE disables all window rectangles. */
       if (!has_clear_state || ctx->num_window_rectangles > 0)

diff --git a/src/gallium/drivers/radeonsi/si_nir_lower_abi.c b/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
index 4841a37..342b4cb 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
+++ b/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
@@ -497,8 +497,8 @@ static bool lower_intrinsic(nir_builder *b, nir_instr *instr, struct lower_abi_s
 
       unsigned stream = nir_intrinsic_stream_id(intrin);
       unsigned offset = intrin->intrinsic == nir_intrinsic_atomic_add_gen_prim_count_amd ?
-         offsetof(struct gfx10_sh_query_buffer_mem, stream[stream].generated_primitives) :
-         offsetof(struct gfx10_sh_query_buffer_mem, stream[stream].emitted_primitives);
+         offsetof(struct gfx11_sh_query_buffer_mem, stream[stream].generated_primitives) :
+         offsetof(struct gfx11_sh_query_buffer_mem, stream[stream].emitted_primitives);
 
       nir_ssa_def *prim_count = intrin->src[0].ssa;
       nir_ssbo_atomic(b, 32, buf, nir_imm_int(b, offset), prim_count,

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 1580a5e..80695b9 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -192,7 +192,7 @@ static void si_destroy_context(struct pipe_context *context)
    si_release_all_descriptors(sctx);
 
    if (sctx->gfx_level >= GFX10 && sctx->has_graphics)
-      gfx10_destroy_query(sctx);
+      gfx11_destroy_query(sctx);
 
    if (sctx->sqtt) {
       struct si_screen *sscreen = sctx->screen;
@@ -634,7 +634,7 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen, unsign
    /* Initialize graphics-only context functions. */
    if (sctx->has_graphics) {
       if (sctx->gfx_level >= GFX10)
-         gfx10_init_query(sctx);
+         gfx11_init_query(sctx);
       si_init_msaa_functions(sctx);
       si_init_shader_functions(sctx);
       si_init_state_functions(sctx);
@@ -993,7 +993,6 @@ static void si_destroy_screen(struct pipe_screen *pscreen)
    simple_mtx_destroy(&sscreen->gpu_load_mutex);
    simple_mtx_destroy(&sscreen->gds_mutex);
 
-   radeon_bo_reference(sscreen->ws, &sscreen->gds, NULL);
    radeon_bo_reference(sscreen->ws, &sscreen->gds_oa, NULL);
 
    slab_destroy_parent(&sscreen->pool_transfers);
@@ -1308,7 +1307,6 @@ static struct pipe_screen *radeonsi_screen_create_impl(struct radeon_winsys *ws,
 
    if (sscreen->info.gfx_level >= GFX11) {
       sscreen->use_ngg = true;
-      sscreen->use_ngg_streamout = true;
       /* TODO: Disable for now. Investigate if it helps.  */
       sscreen->use_ngg_culling = (sscreen->debug_flags & DBG(ALWAYS_NGG_CULLING_ALL)) &&
                                  !(sscreen->debug_flags & DBG(NO_NGG_CULLING));
@@ -1317,7 +1315,6 @@ static struct pipe_screen *radeonsi_screen_create_impl(struct radeon_winsys *ws,
                          sscreen->info.gfx_level >= GFX10 &&
                          (sscreen->info.family != CHIP_NAVI14 ||
                           sscreen->info.is_pro_graphics);
-      sscreen->use_ngg_streamout = false;
       sscreen->use_ngg_culling = sscreen->use_ngg &&
                                  sscreen->info.max_render_backends >= 2 &&
                                  !(sscreen->debug_flags & DBG(NO_NGG_CULLING));

diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index f2d0b21..34902f9 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -564,7 +564,6 @@ struct si_screen {
    bool dpbb_allowed;
    bool use_ngg;
    bool use_ngg_culling;
-   bool use_ngg_streamout;
    bool allow_dcc_msaa_clear_to_reg_for_bpp[5]; /* indexed by log2(Bpp) */
    bool always_allow_dcc_stores;
 
@@ -690,7 +689,6 @@ struct si_screen {
 
    /* NGG streamout. */
    simple_mtx_t gds_mutex;
-   struct pb_buffer *gds;
    struct pb_buffer *gds_oa;
 };
 
@@ -1597,11 +1595,11 @@ void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx);
 void *si_clear_12bytes_buffer_shader(struct pipe_context *ctx);
 void *si_create_fmask_expand_cs(struct pipe_context *ctx, unsigned num_samples, bool is_array);
 void *si_create_query_result_cs(struct si_context *sctx);
-void *gfx10_create_sh_query_result_cs(struct si_context *sctx);
+void *gfx11_create_sh_query_result_cs(struct si_context *sctx);
 
-/* gfx10_query.c */
-void gfx10_init_query(struct si_context *sctx);
-void gfx10_destroy_query(struct si_context *sctx);
+/* gfx11_query.c */
+void gfx11_init_query(struct si_context *sctx);
+void gfx11_destroy_query(struct si_context *sctx);
 
 /* si_test_image_copy_region.c */
 void si_test_image_copy_region(struct si_screen *sscreen);

diff --git a/src/gallium/drivers/radeonsi/si_query.c b/src/gallium/drivers/radeonsi/si_query.c
index 9893d69..70339c7 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_query.c
+++ b/src/gallium/drivers/radeonsi/si_query.c
@@ -1050,10 +1050,10 @@ static void si_emit_query_predication(struct si_context *ctx)
    flag_wait = ctx->render_cond_mode == PIPE_RENDER_COND_WAIT ||
                ctx->render_cond_mode == PIPE_RENDER_COND_BY_REGION_WAIT;
 
-   if (ctx->screen->use_ngg_streamout && (query->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE ||
-                                          query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE)) {
-      struct gfx10_sh_query *gfx10_query = (struct gfx10_sh_query *)query;
-      struct gfx10_sh_query_buffer *qbuf, *first, *last;
+   if (ctx->gfx_level >= GFX11 && (query->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE ||
+                                   query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE)) {
+      struct gfx11_sh_query *gfx10_query = (struct gfx11_sh_query *)query;
+      struct gfx11_sh_query_buffer *qbuf, *first, *last;
 
       op = PRED_OP(PREDICATION_OP_PRIMCOUNT);
 
@@ -1071,7 +1071,7 @@ static void si_emit_query_predication(struct si_context *ctx)
       while (first) {
          qbuf = first;
          if (first != last)
-            first = list_entry(qbuf->list.next, struct gfx10_sh_query_buffer, list);
+            first = list_entry(qbuf->list.next, struct gfx11_sh_query_buffer, list);
          else
             first = NULL;
 
@@ -1082,7 +1082,7 @@ static void si_emit_query_predication(struct si_context *ctx)
          unsigned begin = qbuf == gfx10_query->first ? gfx10_query->first_begin : 0;
          unsigned end = qbuf == gfx10_query->last ? gfx10_query->last_end : qbuf->buf->b.b.width0;
 
-         unsigned count = (end - begin) / sizeof(struct gfx10_sh_query_buffer_mem);
+         unsigned count = (end - begin) / sizeof(struct gfx11_sh_query_buffer_mem);
          do {
             if (gfx10_query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE) {
                for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream) {
@@ -1096,7 +1096,7 @@ static void si_emit_query_predication(struct si_context *ctx)
                op |= PREDICATION_CONTINUE;
             }
 
-            results_base += sizeof(struct gfx10_sh_query_buffer_mem);
+            results_base += sizeof(struct gfx11_sh_query_buffer_mem);
          } while (count--);
       }
    } else {
@@ -1178,12 +1178,12 @@ static struct pipe_query *si_create_query(struct pipe_context *ctx, unsigned que
        (query_type >= PIPE_QUERY_DRIVER_SPECIFIC))
       return si_query_sw_create(query_type);
 
-   if (sscreen->use_ngg_streamout &&
+   if (sscreen->info.gfx_level >= GFX11 &&
        (query_type == PIPE_QUERY_PRIMITIVES_EMITTED ||
         query_type == PIPE_QUERY_PRIMITIVES_GENERATED || query_type == PIPE_QUERY_SO_STATISTICS ||
         query_type == PIPE_QUERY_SO_OVERFLOW_PREDICATE ||
         query_type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE))
-      return gfx10_sh_query_create(sscreen, query_type, index);
+      return gfx11_sh_query_create(sscreen, query_type, index);
 
    return si_query_hw_create(sscreen, query_type, index);
 }

diff --git a/src/gallium/drivers/radeonsi/si_query.h b/src/gallium/drivers/radeonsi/si_query.h
index 2e1e175..e51bf4a 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_query.h
+++ b/src/gallium/drivers/radeonsi/si_query.h
@@ -221,7 +221,7 @@ void si_query_hw_resume(struct si_context *sctx, struct si_query *query);
  * point into the ring, allowing an arbitrary number of queries to be active
  * without additional GPU cost.
  */
-struct gfx10_sh_query_buffer {
+struct gfx11_sh_query_buffer {
    struct list_head list;
    struct si_resource *buf;
    unsigned refcount;
@@ -237,7 +237,7 @@ struct gfx10_sh_query_buffer {
  * SET_PREDICATION packet, which also means that we're setting the high bit
  * of all those values unconditionally.
  */
-struct gfx10_sh_query_buffer_mem {
+struct gfx11_sh_query_buffer_mem {
    struct {
       uint64_t generated_primitives_start_dummy;
       uint64_t emitted_primitives_start_dummy;
@@ -248,18 +248,18 @@ struct gfx10_sh_query_buffer_mem {
    uint32_t pad[31];
 };
 
-struct gfx10_sh_query {
+struct gfx11_sh_query {
    struct si_query b;
 
-   struct gfx10_sh_query_buffer *first;
-   struct gfx10_sh_query_buffer *last;
+   struct gfx11_sh_query_buffer *first;
+   struct gfx11_sh_query_buffer *last;
    unsigned first_begin;
    unsigned last_end;
 
    unsigned stream;
 };
 
-struct pipe_query *gfx10_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type,
+struct pipe_query *gfx11_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type,
                                          unsigned index);
 
 /* Performance counters */

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 18b79a2..38d29da 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -142,7 +142,8 @@ static void declare_streamout_params(struct si_shader_args *args, struct si_shad
 {
    struct si_shader_selector *sel = shader->selector;
 
-   if (sel->screen->use_ngg_streamout) {
+   if (shader->selector->screen->info.gfx_level >= GFX11) {
+      /* NGG streamout. */
       if (sel->stage == MESA_SHADER_TESS_EVAL)
          ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
       return;
@@ -1761,7 +1762,7 @@ static void si_lower_ngg(struct si_shader *shader, nir_shader *nir)
       options.passthrough = gfx10_is_ngg_passthrough(shader);
       options.use_edgeflags = gfx10_edgeflags_have_effect(shader);
       options.has_gen_prim_query = options.has_xfb_prim_query =
-         sel->screen->use_ngg_streamout && !sel->info.base.vs.blit_sgprs_amd;
+         sel->screen->info.gfx_level >= GFX11 && !sel->info.base.vs.blit_sgprs_amd;
       options.export_primitive_id = key->ge.mono.u.vs_export_prim_id;
       options.instance_rate_inputs = instance_rate_inputs;
       options.user_clip_plane_enable_mask = clip_plane_enable;
@@ -1772,7 +1773,7 @@ static void si_lower_ngg(struct si_shader *shader, nir_shader *nir)
 
       options.gs_out_vtx_bytes = sel->info.gsvs_vertex_size;
       options.has_gen_prim_query = options.has_xfb_prim_query =
-         sel->screen->use_ngg_streamout;
+         sel->screen->info.gfx_level >= GFX11;
 
       NIR_PASS_V(nir, ac_nir_lower_ngg_gs, &options);
    }

diff --git a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
index 42dc364..b01fdcf 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
+++ b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
@@ -581,7 +581,7 @@ void *si_create_fmask_expand_cs(struct pipe_context *ctx, unsigned num_samples,
  *          2: write next summary buffer
  *  0.w = result_count
  */
-void *gfx10_create_sh_query_result_cs(struct si_context *sctx)
+void *gfx11_create_sh_query_result_cs(struct si_context *sctx)
 {
    /* TEMP[0].x = accumulated result so far
     * TEMP[0].y = result missing

diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.cpp b/src/gallium/drivers/radeonsi/si_state_shaders.cpp
index 4d7e5ba..777aa16 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_state_shaders.cpp
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.cpp
@@ -1681,7 +1681,7 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
    else if (sscreen->info.gfx_level == GFX9)
       rsrc2 |= S_00B12C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
 
-   if (!sscreen->use_ngg_streamout && si_shader_uses_streamout(shader)) {
+   if (si_shader_uses_streamout(shader)) {
       rsrc2 |= S_00B12C_SO_BASE0_EN(!!shader->selector->info.base.xfb_stride[0]) |
                S_00B12C_SO_BASE1_EN(!!shader->selector->info.base.xfb_stride[1]) |
                S_00B12C_SO_BASE2_EN(!!shader->selector->info.base.xfb_stride[2]) |
@@ -2963,7 +2963,7 @@ static void si_init_shader_selector_async(void *job, void *gdata, int thread_ind
 
       if (sel->stage <= MESA_SHADER_GEOMETRY &&
           sscreen->use_ngg && (!sel->info.enabled_streamout_buffer_mask ||
-                               sscreen->use_ngg_streamout) &&
+                               sscreen->info.gfx_level >= GFX11) &&
           ((sel->stage == MESA_SHADER_VERTEX && !shader->key.ge.as_ls) ||
            sel->stage == MESA_SHADER_TESS_EVAL || sel->stage == MESA_SHADER_GEOMETRY))
          shader->key.ge.as_ngg = 1;
@@ -3258,7 +3258,7 @@ static void si_update_streamout_state(struct si_context *sctx)
    sctx->streamout.stride_in_dw = shader_with_so->info.base.xfb_stride;
 
    /* GDS must be allocated when any GDS instructions are used, otherwise it hangs. */
-   if (sctx->screen->use_ngg_streamout && shader_with_so->info.enabled_streamout_buffer_mask)
+   if (sctx->gfx_level >= GFX11 && shader_with_so->info.enabled_streamout_buffer_mask)
       si_allocate_gds(sctx);
 }
 
@@ -3384,7 +3384,7 @@ bool si_update_ngg(struct si_context *sctx)
 
    if (sctx->shader.gs.cso && sctx->shader.tes.cso && sctx->shader.gs.cso->tess_turns_off_ngg) {
       new_ngg = false;
-   } else if (!sctx->screen->use_ngg_streamout) {
+   } else if (sctx->gfx_level < GFX11) {
       struct si_shader_selector *last = si_get_vs(sctx)->cso;
 
       if ((last && last->info.enabled_streamout_buffer_mask) ||

diff --git a/src/gallium/drivers/radeonsi/si_state_streamout.c b/src/gallium/drivers/radeonsi/si_state_streamout.c
index d6fe2cc..fefabf8 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_state_streamout.c
+++ b/src/gallium/drivers/radeonsi/si_state_streamout.c
@@ -92,13 +92,8 @@ static void si_set_streamout_targets(struct pipe_context *ctx, unsigned num_targ
        */
       sctx->flags |= SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE;
 
-      if (sctx->screen->use_ngg_streamout) {
-         if (sctx->gfx_level >= GFX11) {
-            sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH | SI_CONTEXT_PFP_SYNC_ME;
-         } else {
-            /* The BUFFER_FILLED_SIZE is written using a PS_DONE event. */
-            sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_PFP_SYNC_ME;
-         }
+      if (sctx->gfx_level >= GFX11) {
+         sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH | SI_CONTEXT_PFP_SYNC_ME;
 
          /* Wait now. This is needed to make sure that GDS is not
           * busy at the end of IBs.
@@ -149,7 +144,7 @@ static void si_set_streamout_targets(struct pipe_context *ctx, unsigned num_targ
       /* Allocate space for the filled buffer size. */
       struct si_streamout_target *t = sctx->streamout.targets[i];
       if (!t->buf_filled_size) {
-         unsigned buf_filled_size_size = sctx->screen->use_ngg_streamout ? 8 : 4;
+         unsigned buf_filled_size_size = sctx->gfx_level >= GFX11 ? 8 : 4;
          u_suballocator_alloc(&sctx->allocator_zeroed_memory, buf_filled_size_size, 4,
                               &t->buf_filled_size_offset,
                               (struct pipe_resource **)&t->buf_filled_size);
@@ -181,7 +176,7 @@ static void si_set_streamout_targets(struct pipe_context *ctx, unsigned num_targ
          struct pipe_shader_buffer sbuf;
          sbuf.buffer = targets[i]->buffer;
 
-         if (sctx->screen->use_ngg_streamout) {
+         if (sctx->gfx_level >= GFX11) {
             sbuf.buffer_offset = targets[i]->buffer_offset;
             sbuf.buffer_size = targets[i]->buffer_size;
          } else {
@@ -243,7 +238,7 @@ static void si_emit_streamout_begin(struct si_context *sctx)
    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
    struct si_streamout_target **t = sctx->streamout.targets;
 
-   if (!sctx->screen->use_ngg_streamout)
+   if (sctx->gfx_level < GFX11)
       si_flush_vgt_streamout(sctx);
 
    for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
@@ -265,27 +260,6 @@ static void si_emit_streamout_begin(struct si_context *sctx)
             radeon_set_uconfig_reg(R_031088_GDS_STRMOUT_DWORDS_WRITTEN_0 + i * 4, 0);
             radeon_end();
          }
-      } else if (sctx->screen->use_ngg_streamout) {
-         bool append = sctx->streamout.append_bitmask & (1 << i);
-         uint64_t va = 0;
-
-         if (append) {
-            radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size,
-                                      RADEON_USAGE_READ | RADEON_PRIO_SO_FILLED_SIZE);
-
-            va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
-         }
-
-         radeon_begin(cs);
-         radeon_emit(PKT3(PKT3_DMA_DATA, 5, 0));
-         radeon_emit(S_411_SRC_SEL(append ? V_411_SRC_ADDR_TC_L2 : V_411_DATA) |
-                     S_411_DST_SEL(V_411_GDS) | S_411_CP_SYNC(1));
-         radeon_emit(va);
-         radeon_emit(va >> 32);
-         radeon_emit(4 * i); /* destination in GDS */
-         radeon_emit(0);
-         radeon_emit(S_415_BYTE_COUNT_GFX9(4));
-         radeon_end();
       } else {
          /* Legacy streamout.
           *
@@ -337,7 +311,7 @@ void si_emit_streamout_end(struct si_context *sctx)
       /* Wait for streamout to finish before reading GDS_STRMOUT registers. */
       sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH;
       sctx->emit_cache_flush(sctx, &sctx->gfx_cs);
-   } else if (!sctx->screen->use_ngg_streamout) {
+   } else {
       si_flush_vgt_streamout(sctx);
    }
 
@@ -353,11 +327,6 @@ void si_emit_streamout_end(struct si_context *sctx)
                          COPY_DATA_REG, NULL,
                          (R_031088_GDS_STRMOUT_DWORDS_WRITTEN_0 >> 2) + i);
          sctx->flags |= SI_CONTEXT_PFP_SYNC_ME;
-      } else if (sctx->screen->use_ngg_streamout) {
-         /* TODO: PS_DONE doesn't ensure completion of VS if there are no PS waves. */
-         si_cp_release_mem(sctx, &sctx->gfx_cs, V_028A90_PS_DONE, 0, EOP_DST_SEL_TC_L2,
-                           EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM, EOP_DATA_SEL_GDS,
-                           t[i]->buf_filled_size, va, EOP_DATA_GDS(i, 1), 0);
       } else {
          radeon_begin(cs);
          radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
@@ -394,7 +363,7 @@ void si_emit_streamout_end(struct si_context *sctx)
 
 static void si_emit_streamout_enable(struct si_context *sctx)
 {
-   assert(!sctx->screen->use_ngg_streamout);
+   assert(sctx->gfx_level < GFX11);
 
    radeon_begin(&sctx->gfx_cs);
    radeon_set_context_reg_seq(R_028B94_VGT_STRMOUT_CONFIG, 2);
@@ -418,7 +387,7 @@ static void si_set_streamout_enable(struct si_context *sctx, bool enable)
       sctx->streamout.enabled_mask | (sctx->streamout.enabled_mask << 4) |
       (sctx->streamout.enabled_mask << 8) | (sctx->streamout.enabled_mask << 12);
 
-   if (!sctx->screen->use_ngg_streamout &&
+   if (sctx->gfx_level < GFX11 &&
        ((old_strmout_en != si_get_strmout_en(sctx)) ||
         (old_hw_enabled_mask != sctx->streamout.hw_enabled_mask)))
       si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
@@ -426,7 +395,7 @@ static void si_set_streamout_enable(struct si_context *sctx, bool enable)
 
 void si_update_prims_generated_query_state(struct si_context *sctx, unsigned type, int diff)
 {
-   if (!sctx->screen->use_ngg_streamout && type == PIPE_QUERY_PRIMITIVES_GENERATED) {
+   if (sctx->gfx_level < GFX11 && type == PIPE_QUERY_PRIMITIVES_GENERATED) {
       bool old_strmout_en = si_get_strmout_en(sctx);
 
       sctx->streamout.num_prims_gen_queries += diff;
@@ -451,6 +420,6 @@ void si_init_streamout_functions(struct si_context *sctx)
    sctx->b.set_stream_output_targets = si_set_streamout_targets;
    sctx->atoms.s.streamout_begin.emit = si_emit_streamout_begin;
 
-   if (!sctx->screen->use_ngg_streamout)
+   if (sctx->gfx_level < GFX11)
       sctx->atoms.s.streamout_enable.emit = si_emit_streamout_enable;
 }