From ca9e0871fb29412301a2e3058c7229c2a6223bbf Mon Sep 17 00:00:00 2001 From: Iago Toral Quiroga Date: Mon, 3 May 2021 11:12:43 +0200 Subject: [PATCH] v3d: enable NIR loop unrolling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The GL driver was getting loop unrolling from the GLSL compiler frontend, but NIR unrolling is more sophisticated, so prefer that. The only caveat is that loop unrolling is implemented in the Mesa state tracker, so our backend won't have a chance to undo the optimization if it causes us to lower thread count or spill, so we choose to be a bit more conservative with the configuration than what we were doing with GLSL. Shader-db results follow. Increase in instruction counts is expected due to additional unrolling. We lose threads in very few shaders, but we make up for this with the additional unrolling and reduced spilling. We also managed to get 3 more shaders to compile successfully. total instructions in shared programs: 13416427 -> 13461431 (0.34%) instructions in affected programs: 96936 -> 141940 (46.43%) helped: 58 HURT: 216 Instructions are HURT. total threads in shared programs: 410626 -> 410598 (<.01%) threads in affected programs: 56 -> 28 (-50.00%) helped: 0 HURT: 14 Threads are HURT. total loops in shared programs: 2121 -> 1708 (-19.47%) loops in affected programs: 468 -> 55 (-88.25%) helped: 446 HURT: 47 Loops are helped. total uniforms in shared programs: 3676567 -> 3691185 (0.40%) uniforms in affected programs: 25304 -> 39922 (57.77%) helped: 23 HURT: 199 Uniforms are HURT. total spills in shared programs: 5902 -> 5727 (-2.97%) spills in affected programs: 285 -> 110 (-61.40%) helped: 19 HURT: 0 total fills in shared programs: 13308 -> 13121 (-1.41%) fills in affected programs: 301 -> 114 (-62.13%) helped: 19 HURT: 0 total sfu-stalls in shared programs: 31860 -> 32856 (3.13%) sfu-stalls in affected programs: 1692 -> 2688 (58.87%) helped: 25 HURT: 196 Sfu-stalls are HURT. total inst-and-stalls in shared programs: 13448287 -> 13494287 (0.34%) inst-and-stalls in affected programs: 98404 -> 144404 (46.75%) helped: 57 HURT: 217 Inst-and-stalls are HURT. total nops in shared programs: 329276 -> 329551 (0.08%) nops in affected programs: 2189 -> 2464 (12.56%) helped: 58 HURT: 181 Nops are HURT. LOST: 0 GAINED: 3 Reviewed-by: Alejandro Piñeiro Part-of: --- src/gallium/drivers/v3d/v3d_screen.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/src/gallium/drivers/v3d/v3d_screen.c b/src/gallium/drivers/v3d/v3d_screen.c index 09e21a33782..a6d497f0ae9 100644 --- a/src/gallium/drivers/v3d/v3d_screen.c +++ b/src/gallium/drivers/v3d/v3d_screen.c @@ -423,7 +423,8 @@ v3d_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader, case PIPE_SHADER_CAP_SUPPORTED_IRS: return 1 << PIPE_SHADER_IR_NIR; case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT: - return 32; + /* Disable GLSL loop unrolling, we'll use NIR's */ + return 0; case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD: case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS: return 0; @@ -668,6 +669,12 @@ static const nir_shader_compiler_options v3d_nir_options = { .has_isub = true, .divergence_analysis_options = nir_divergence_multiple_workgroup_per_compute_subgroup, + /* This will enable loop unrolling in the state tracker so we won't + * be able to selectively disable it in backend if it leads to + * lower thread counts or TMU spills. Choose a conservative maximum to + * limit register pressure impact. + */ + .max_unroll_iterations = 16, }; static const void *