From: Rob Clark Date: Wed, 13 Jun 2018 13:50:34 +0000 (-0400) Subject: freedreno/a5xx: better FOUR_QUAD/TWO_QUAD decision for compute X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=1a6150207c3d28adf8ebf51cc12d44bd3e830beb;p=mesa.git freedreno/a5xx: better FOUR_QUAD/TWO_QUAD decision for compute If we aren't going to get full occupancy, then use TWO_QUAD. Signed-off-by: Rob Clark --- diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_compute.c b/src/gallium/drivers/freedreno/a5xx/fd5_compute.c index c7635a98c9f..8f3847ea843 100644 --- a/src/gallium/drivers/freedreno/a5xx/fd5_compute.c +++ b/src/gallium/drivers/freedreno/a5xx/fd5_compute.c @@ -70,13 +70,21 @@ fd5_delete_compute_state(struct pipe_context *pctx, void *hwcso) /* maybe move to fd5_program? */ static void -cs_program_emit(struct fd_ringbuffer *ring, struct ir3_shader_variant *v) +cs_program_emit(struct fd_ringbuffer *ring, struct ir3_shader_variant *v, + const struct pipe_grid_info *info) { + const unsigned *local_size = info->block; const struct ir3_info *i = &v->info; enum a3xx_threadsize thrsz; - /* note: blob uses local_size_x/y/z threshold to choose threadsize: */ - thrsz = FOUR_QUADS; + /* maybe the limit should be 1024.. basically if we can't have full + * occupancy, use TWO_QUAD mode to reduce divergence penalty. + */ + if ((local_size[0] * local_size[1] * local_size[2]) < 512) { + thrsz = TWO_QUADS; + } else { + thrsz = FOUR_QUADS; + } OUT_PKT4(ring, REG_A5XX_SP_SP_CNTL, 1); OUT_RING(ring, 0x00000000); /* SP_SP_CNTL */ @@ -175,7 +183,7 @@ fd5_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info) v = ir3_shader_variant(so->shader, key, &ctx->debug); if (ctx->dirty_shader[PIPE_SHADER_COMPUTE] & FD_DIRTY_SHADER_PROG) - cs_program_emit(ring, v); + cs_program_emit(ring, v, info); fd5_emit_cs_state(ctx, ring, v); ir3_emit_cs_consts(v, ring, ctx, info);