radeonsi: move VS_STATE.LS_OUT_PATCH_SIZE a few bits higher to make space there

[mesa.git] / src / gallium / drivers / radeonsi / si_test_dma_perf.c

diff --git a/src/gallium/drivers/radeonsi/si_test_dma_perf.c b/src/gallium/drivers/radeonsi/si_test_dma_perf.c
index be2ad079e1a946d41862877e130a81cdca9bbbe9..4eec3d124594bfd9d9e1659c494dae55a7ddd90b 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_test_dma_perf.c
+++ b/src/gallium/drivers/radeonsi/si_test_dma_perf.c
@@ -112,11 +112,14 @@ void si_test_dma_perf(struct si_screen *sscreen)
                        unsigned cs_dwords_per_thread =
                                test_cs ? cs_dwords_per_thread_list[cs_method % NUM_SHADERS] : 0;
 
-                       if (sctx->chip_class == SI) {
-                               /* SI doesn't support CP DMA operations through L2. */
+                       if (test_sdma && !sctx->sdma_cs)
+                               continue;
+
+                       if (sctx->chip_class == GFX6) {
+                               /* GFX6 doesn't support CP DMA operations through L2. */
                                if (test_cp && cache_policy != L2_BYPASS)
                                        continue;
-                               /* WAVES_PER_SH is in multiples of 16 on SI. */
+                               /* WAVES_PER_SH is in multiples of 16 on GFX6. */
                                if (test_cs && cs_waves_per_sh % 16 != 0)
                                        continue;
                        }
@@ -151,7 +154,7 @@ void si_test_dma_perf(struct si_screen *sscreen)
                                unsigned query_type = PIPE_QUERY_TIME_ELAPSED;
 
                                if (test_sdma) {
-                                       if (sctx->chip_class == SI)
+                                       if (sctx->chip_class == GFX6)
                                                query_type = SI_QUERY_TIME_ELAPSED_SDMA_SI;
                                        else
                                                query_type = SI_QUERY_TIME_ELAPSED_SDMA;
@@ -178,20 +181,19 @@ void si_test_dma_perf(struct si_screen *sscreen)
                                        if (test_cp) {
                                                /* CP DMA */
                                                if (is_copy) {
-                                                       si_copy_buffer(sctx, dst, src, 0, 0, size, 0,
-                                                                      cache_policy);
+                                                       si_cp_dma_copy_buffer(sctx, dst, src, 0, 0, size, 0,
+                                                                             SI_COHERENCY_NONE, cache_policy);
                                                } else {
-                                                       si_cp_dma_clear_buffer(sctx, dst, 0, size, clear_value,
+                                                       si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, dst, 0, size,
+                                                                              clear_value, 0,
                                                                               SI_COHERENCY_NONE, cache_policy);
                                                }
                                        } else if (test_sdma) {
                                                /* SDMA */
                                                if (is_copy) {
-                                                       struct pipe_box box;
-                                                       u_box_1d(0, size, &box);
-                                                       sctx->dma_copy(ctx, dst, 0, 0, 0, 0, src, 0, &box);
+                                                       si_sdma_copy_buffer(sctx, dst, src, 0, 0, size);
                                                } else {
-                                                       sctx->dma_clear_buffer(sctx, dst, 0, size, clear_value);
+                                                       si_sdma_clear_buffer(sctx, dst, 0, size, clear_value);
                                                }
                                        } else {
                                                /* Compute */
@@ -229,10 +231,11 @@ void si_test_dma_perf(struct si_screen *sscreen)
                                                                sctx->cs_user_data[i] = clear_value;
                                                }
 
-                                               sctx->flags |= SI_CONTEXT_INV_VMEM_L1 |
-                                                              SI_CONTEXT_INV_SMEM_L1;
+                                               sctx->flags |= SI_CONTEXT_INV_VCACHE |
+                                                              SI_CONTEXT_INV_SCACHE;
 
-                                               ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0, is_copy ? 2 : 1, sb);
+                                               ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0,
+                                                                       is_copy ? 2 : 1, sb, 0x1);
                                                ctx->bind_compute_state(ctx, cs);
                                                sctx->cs_max_waves_per_sh = cs_waves_per_sh;
 
@@ -247,8 +250,8 @@ void si_test_dma_perf(struct si_screen *sscreen)
 
                                        /* Flush L2, so that we don't just test L2 cache performance. */
                                        if (!test_sdma) {
-                                               sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
-                                               si_emit_cache_flush(sctx);
+                                               sctx->flags |= SI_CONTEXT_WB_L2;
+                                               sctx->emit_cache_flush(sctx);
                                        }
 
                                        ctx->end_query(ctx, q[iter]);
@@ -344,10 +347,10 @@ void si_test_dma_perf(struct si_screen *sscreen)
                                        if (!r->is_valid)
                                                continue;
 
-                                       /* Ban CP DMA clears via MC on <= VI. They are super slow
+                                       /* Ban CP DMA clears via MC on <= GFX8. They are super slow
                                         * on GTT, which we can get due to BO evictions.
                                         */
-                                       if (sctx->chip_class <= VI && placement == 1 &&
+                                       if (sctx->chip_class <= GFX8 && placement == 1 &&
                                            r->is_cp && r->cache_policy == L2_BYPASS)
                                                continue;