radeonsi: enable TGSI support cap for compute shaders

v2: Use chip_class instead of family.

v3: Check kernel version for SI.

v4: Preemptively allow amdgpu winsys for SI.

Signed-off-by: Bas Nieuwenhuizen <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Reviewed-by: Marek Olšák <[email protected]>

4 files changed, 33 insertions, 9 deletions

diff --git a/docs/GL3.txt b/docs/GL3.txt
index 3febd6e45f0..6214f8d3073 100644
--- a/docs/GL3.txt
+++ b/docs/GL3.txt
@@ -167,7 +167,7 @@ GL 4.3, GLSL 4.30:
   GL_ARB_arrays_of_arrays                               DONE (all drivers that support GLSL 1.30)
   GL_ARB_ES3_compatibility                              DONE (all drivers that support GLSL 3.30)
   GL_ARB_clear_buffer_object                            DONE (all drivers)
-  GL_ARB_compute_shader                                 DONE (i965)
+  GL_ARB_compute_shader                                 DONE (i965, radeonsi)
   GL_ARB_copy_image                                     DONE (i965, nv50, nvc0, r600, radeonsi)
   GL_KHR_debug                                          DONE (all drivers)
   GL_ARB_explicit_uniform_location                      DONE (all drivers that support GLSL)
@@ -225,7 +225,7 @@ GL 4.5, GLSL 4.50:
 These are the extensions cherry-picked to make GLES 3.1
 GLES3.1, GLSL ES 3.1
   GL_ARB_arrays_of_arrays                               DONE (all drivers that support GLSL 1.30)
-  GL_ARB_compute_shader                                 DONE (i965)
+  GL_ARB_compute_shader                                 DONE (i965, radeonsi)
   GL_ARB_draw_indirect                                  DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
   GL_ARB_explicit_uniform_location                      DONE (all drivers that support GLSL)
   GL_ARB_framebuffer_no_attachments                     DONE (i965, nvc0, r600, radeonsi, softpipe)
diff --git a/docs/relnotes/11.3.0.html b/docs/relnotes/11.3.0.html
index 0f9aed8fd1e..5a7083c5ba4 100644
--- a/docs/relnotes/11.3.0.html
+++ b/docs/relnotes/11.3.0.html
@@ -45,6 +45,7 @@ Note: some of the new features are only available with certain drivers.
 
 <ul>
 <li>OpenGL 4.2 on radeonsi</li>
+<li>GL_ARB_compute_shader on radeonsi</li>
 <li>GL_ARB_framebuffer_no_attachments on nvc0, r600, radeonsi, softpipe</li>
 <li>GL_ARB_internalformat_query2 on all drivers</li>
 <li>GL_ARB_robust_buffer_access_behavior on radeonsi</li>
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index a8660f20c86..9ed6da6a82b 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -646,23 +646,34 @@ static int r600_get_compute_param(struct pipe_screen *screen,
 			uint64_t *grid_size = ret;
 			grid_size[0] = 65535;
 			grid_size[1] = 65535;
-			grid_size[2] = 1;
+			grid_size[2] = 65535;
 		}
 		return 3 * sizeof(uint64_t) ;
 
 	case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 		if (ret) {
 			uint64_t *block_size = ret;
-			block_size[0] = 256;
-			block_size[1] = 256;
-			block_size[2] = 256;
+			if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+			    ir_type == PIPE_SHADER_IR_TGSI) {
+				block_size[0] = 2048;
+				block_size[1] = 2048;
+				block_size[2] = 2048;
+			} else {
+				block_size[0] = 256;
+				block_size[1] = 256;
+				block_size[2] = 256;
+			}
 		}
 		return 3 * sizeof(uint64_t);
 
 	case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 		if (ret) {
 			uint64_t *max_threads_per_block = ret;
-			*max_threads_per_block = 256;
+			if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+			    ir_type == PIPE_SHADER_IR_TGSI)
+				*max_threads_per_block = 2048;
+			else
+				*max_threads_per_block = 256;
 		}
 		return sizeof(uint64_t);
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index dabd28a4bc2..17d59b60d06 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -473,6 +473,8 @@ static int si_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 
 static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
 {
+	struct si_screen *sscreen = (struct si_screen *)pscreen;
+
 	switch(shader)
 	{
 	case PIPE_SHADER_FRAGMENT:
@@ -490,9 +492,19 @@ static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enu
 		case PIPE_SHADER_CAP_PREFERRED_IR:
 			return PIPE_SHADER_IR_NATIVE;
 
-		case PIPE_SHADER_CAP_SUPPORTED_IRS:
-			return 0;
+		case PIPE_SHADER_CAP_SUPPORTED_IRS: {
+			int ir = 1 << PIPE_SHADER_IR_NATIVE;
 
+			/* Old kernels disallowed some register writes for SI
+			 * that are used for indirect dispatches. */
+			if (HAVE_LLVM >= 0x309 && (sscreen->b.chip_class >= CIK ||
+			                           sscreen->b.info.drm_major == 3 ||
+			                           (sscreen->b.info.drm_major == 2 &&
+			                            sscreen->b.info.drm_minor >= 45)))
+				ir |= 1 << PIPE_SHADER_IR_TGSI;
+
+			return ir;
+		}
 		case PIPE_SHADER_CAP_DOUBLES:
 			return HAVE_LLVM >= 0x0307;