diff options
Diffstat (limited to 'drivers')
-rw-r--r-- | drivers/gpu/nvgpu/gv11b/gr_gv11b.c | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/drivers/gpu/nvgpu/gv11b/gr_gv11b.c b/drivers/gpu/nvgpu/gv11b/gr_gv11b.c index 8d8f56f6..2b242978 100644 --- a/drivers/gpu/nvgpu/gv11b/gr_gv11b.c +++ b/drivers/gpu/nvgpu/gv11b/gr_gv11b.c | |||
@@ -2123,6 +2123,26 @@ void gr_gv11b_detect_sm_arch(struct gk20a *g) | |||
2123 | gr_gpc0_tpc0_sm_arch_warp_count_v(v); | 2123 | gr_gpc0_tpc0_sm_arch_warp_count_v(v); |
2124 | } | 2124 | } |
2125 | 2125 | ||
2126 | static u32 gr_gv11b_get_nonpes_aware_tpc(struct gk20a *g, u32 gpc, u32 tpc) | ||
2127 | { | ||
2128 | u32 tpc_new = 0; | ||
2129 | u32 temp; | ||
2130 | u32 pes; | ||
2131 | struct gr_gk20a *gr = &g->gr; | ||
2132 | |||
2133 | for (pes = 0; pes < gr->gpc_ppc_count[gpc]; pes++) { | ||
2134 | if (gr->pes_tpc_mask[pes][gpc] & BIT(tpc)) | ||
2135 | break; | ||
2136 | tpc_new += gr->pes_tpc_count[pes][gpc]; | ||
2137 | } | ||
2138 | temp = (BIT(tpc) - 1) & gr->pes_tpc_mask[pes][gpc]; | ||
2139 | temp = hweight32(temp); | ||
2140 | tpc_new += temp; | ||
2141 | |||
2142 | nvgpu_log_info(g, "tpc: %d -> new tpc: %d", tpc, tpc_new); | ||
2143 | return tpc_new; | ||
2144 | } | ||
2145 | |||
2126 | void gr_gv11b_program_sm_id_numbering(struct gk20a *g, | 2146 | void gr_gv11b_program_sm_id_numbering(struct gk20a *g, |
2127 | u32 gpc, u32 tpc, u32 smid) | 2147 | u32 gpc, u32 tpc, u32 smid) |
2128 | { | 2148 | { |
@@ -2133,6 +2153,8 @@ void gr_gv11b_program_sm_id_numbering(struct gk20a *g, | |||
2133 | u32 tpc_offset = tpc_in_gpc_stride * tpc; | 2153 | u32 tpc_offset = tpc_in_gpc_stride * tpc; |
2134 | u32 global_tpc_index = g->gr.sm_to_cluster[smid].global_tpc_index; | 2154 | u32 global_tpc_index = g->gr.sm_to_cluster[smid].global_tpc_index; |
2135 | 2155 | ||
2156 | tpc = gr_gv11b_get_nonpes_aware_tpc(g, gpc, tpc); | ||
2157 | |||
2136 | gk20a_writel(g, gr_gpc0_tpc0_sm_cfg_r() + gpc_offset + tpc_offset, | 2158 | gk20a_writel(g, gr_gpc0_tpc0_sm_cfg_r() + gpc_offset + tpc_offset, |
2137 | gr_gpc0_tpc0_sm_cfg_tpc_id_f(global_tpc_index)); | 2159 | gr_gpc0_tpc0_sm_cfg_tpc_id_f(global_tpc_index)); |
2138 | gk20a_writel(g, gr_gpc0_gpm_pd_sm_id_r(tpc) + gpc_offset, | 2160 | gk20a_writel(g, gr_gpc0_gpm_pd_sm_id_r(tpc) + gpc_offset, |