freedreno: implement GL_ARB_indirect_parameters

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21593>

5 files changed, 45 insertions, 3 deletions

diff --git a/docs/features.txt b/docs/features.txt
index e241143c4a7..571e0c574d5 100644
--- a/docs/features.txt
+++ b/docs/features.txt
@@ -228,11 +228,11 @@ GL 4.5, GLSL 4.50 -- all DONE: freedreno/a6xx, nvc0, r600, radeonsi, llvmpipe, z
 GL 4.6, GLSL 4.60 -- all DONE: radeonsi, zink
 
   GL_ARB_gl_spirv                                       DONE (freedreno, i965/gen7+, llvmpipe)
-  GL_ARB_indirect_parameters                            DONE (i965/gen7+, nvc0, llvmpipe, virgl, d3d12)
+  GL_ARB_indirect_parameters                            DONE (freedreno/a6xx+, i965/gen7+, nvc0, llvmpipe, virgl, d3d12)
   GL_ARB_pipeline_statistics_query                      DONE (i965, nvc0, r600, llvmpipe, softpipe, )
   GL_ARB_polygon_offset_clamp                           DONE (freedreno, i965, nv50, nvc0, r600, llvmpipe, v3d, virgl)
   GL_ARB_shader_atomic_counter_ops                      DONE (freedreno/a5xx+, i965/gen7+, nvc0, r600, llvmpipe, softpipe, virgl, v3d)
-  GL_ARB_shader_draw_parameters                         DONE (i965, llvmpipe, nvc0, d3d12)
+  GL_ARB_shader_draw_parameters                         DONE (freedreno/a6xx+, i965, llvmpipe, nvc0, d3d12)
   GL_ARB_shader_group_vote                              DONE (freedreno/a6xx, i965, nvc0, llvmpipe)
   GL_ARB_spirv_extensions                               DONE (freedreno, i965/gen7+, llvmpipe)
   GL_ARB_texture_filter_anisotropic                     DONE (etnaviv/HALTI0, freedreno, i965, nv50, nvc0, r600, softpipe, llvmpipe, d3d12, virgl, v3d, panfrost/g72+, asahi)
diff --git a/src/freedreno/ci/freedreno-a618-fails.txt b/src/freedreno/ci/freedreno-a618-fails.txt
index 46b18581699..791c8cb0730 100644
--- a/src/freedreno/ci/freedreno-a618-fails.txt
+++ b/src/freedreno/ci/freedreno-a618-fails.txt
@@ -125,6 +125,11 @@ spec@arb_vertex_type_2_10_10_10_rev@attrib-p-type-size-match,Fail
 spec@arb_enhanced_layouts@execution@component-layout@vs-fs-array-dvec3,Crash
 spec@arb_enhanced_layouts@gs-stream-location-aliasing,Fail
 
+# fails on gen1 (a618/a630) with both fd and zink, but passes on gen4..
+# maybe gen1 sqe doesn't handle the count==0 case?
+spec@arb_indirect_parameters@tf-count-arrays,Fail
+spec@arb_indirect_parameters@tf-count-elements,Fail
+
 spec@egl 1.4@eglterminate then unbind context,Fail
 spec@egl_chromium_sync_control@conformance@eglGetSyncValuesCHROMIUM_msc_and_sbc_test,Fail
 spec@egl_chromium_sync_control@conformance,Fail
diff --git a/src/freedreno/ci/freedreno-a630-fails.txt b/src/freedreno/ci/freedreno-a630-fails.txt
index f16788ab4e9..f312e17d04b 100644
--- a/src/freedreno/ci/freedreno-a630-fails.txt
+++ b/src/freedreno/ci/freedreno-a630-fails.txt
@@ -131,6 +131,11 @@ spec@arb_vertex_type_2_10_10_10_rev@attrib-p-type-size-match,Fail
 spec@arb_enhanced_layouts@execution@component-layout@vs-fs-array-dvec3,Crash
 spec@arb_enhanced_layouts@gs-stream-location-aliasing,Fail
 
+# fails on gen1 (a618/a630) with both fd and zink, but passes on gen4..
+# maybe gen1 sqe doesn't handle the count==0 case?
+spec@arb_indirect_parameters@tf-count-arrays,Fail
+spec@arb_indirect_parameters@tf-count-elements,Fail
+
 spec@egl 1.4@eglterminate then unbind context,Fail
 spec@egl_chromium_sync_control@conformance@eglGetSyncValuesCHROMIUM_msc_and_sbc_test,Fail
 spec@egl_chromium_sync_control@conformance,Fail
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_draw.cc b/src/gallium/drivers/freedreno/a6xx/fd6_draw.cc
index 015cde6f3f2..aefd347d184 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_draw.cc
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_draw.cc
@@ -83,7 +83,26 @@ draw_emit_indirect(struct fd_context *ctx,
 {
    struct fd_resource *ind = fd_resource(indirect->buffer);
 
-   if (info->index_size) {
+   if (indirect->indirect_draw_count && info->index_size) {
+      //On some firmwares CP_DRAW_INDIRECT_MULTI waits for WFIs before
+      //reading the draw parameters but after reading the count, so commands
+      //that use indirect draw count need a WFM anyway.
+      OUT_PKT7(ring, CP_WAIT_FOR_ME, 0);
+      OUT_PKT7(ring, CP_DRAW_INDIRECT_MULTI, 11);
+      OUT_RING(ring, pack_CP_DRAW_INDX_OFFSET_0(*draw0).value);
+      OUT_RING(ring,
+         (A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT_INDEXED)
+         | A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(driver_param)));
+      struct fd_resource *count_buf = fd_resource(indirect->indirect_draw_count);
+      struct pipe_resource *idx = info->index.resource;
+      unsigned max_indices = (idx->width0 - index_offset) / info->index_size;
+      OUT_RING(ring, indirect->draw_count);
+      OUT_RELOC(ring, fd_resource(idx)->bo, index_offset, 0, 0);
+      OUT_RING(ring, max_indices);
+      OUT_RELOC(ring, ind->bo, indirect->offset, 0, 0);
+      OUT_RELOC(ring, count_buf->bo, indirect->indirect_draw_count_offset, 0, 0);
+      OUT_RING(ring, indirect->stride);
+   } else if (info->index_size) {
       OUT_PKT7(ring, CP_DRAW_INDIRECT_MULTI, 9);
       OUT_RING(ring, pack_CP_DRAW_INDX_OFFSET_0(*draw0).value);
       OUT_RING(ring,
@@ -98,6 +117,18 @@ draw_emit_indirect(struct fd_context *ctx,
       OUT_RING(ring, max_indices);
       OUT_RELOC(ring, ind->bo, indirect->offset, 0, 0);
       OUT_RING(ring, indirect->stride);
+   }  else if(indirect->indirect_draw_count) {
+      OUT_PKT7(ring, CP_WAIT_FOR_ME, 0);
+      OUT_PKT7(ring, CP_DRAW_INDIRECT_MULTI, 8);
+      OUT_RING(ring, pack_CP_DRAW_INDX_OFFSET_0(*draw0).value);
+      OUT_RING(ring,
+         (A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT)
+         | A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(driver_param)));
+      struct fd_resource *count_buf = fd_resource(indirect->indirect_draw_count);
+      OUT_RING(ring, indirect->draw_count);
+      OUT_RELOC(ring, ind->bo, indirect->offset, 0, 0);
+      OUT_RELOC(ring, count_buf->bo, indirect->indirect_draw_count_offset, 0, 0);
+      OUT_RING(ring, indirect->stride);
    } else {
       OUT_PKT7(ring, CP_DRAW_INDIRECT_MULTI, 6);
       OUT_RING(ring, pack_CP_DRAW_INDX_OFFSET_0(*draw0).value);
diff --git a/src/gallium/drivers/freedreno/freedreno_screen.c b/src/gallium/drivers/freedreno/freedreno_screen.c
index f5081a3c10a..bf2b37b5391 100644
--- a/src/gallium/drivers/freedreno/freedreno_screen.c
+++ b/src/gallium/drivers/freedreno/freedreno_screen.c
@@ -215,6 +215,7 @@ fd_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_CLEAR_TEXTURE:
    case PIPE_CAP_MULTI_DRAW_INDIRECT:
    case PIPE_CAP_DRAW_PARAMETERS:
+   case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
       return is_a6xx(screen);
 
    case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY: