X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fgen8_depth_state.c;h=0eb993fcd7bfc41bfd59b29cea19a01266043f50;hb=1be1114e6b977ef70cc8bba9485adbaace4ecde4;hp=31d6d565f08a515ad61596a737964d271c1fc784;hpb=23e9f06569f58062820ab058d8057e9325042105;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/gen8_depth_state.c b/src/mesa/drivers/dri/i965/gen8_depth_state.c index 31d6d565f08..0eb993fcd7b 100644 --- a/src/mesa/drivers/dri/i965/gen8_depth_state.c +++ b/src/mesa/drivers/dri/i965/gen8_depth_state.c @@ -28,6 +28,8 @@ #include "brw_context.h" #include "brw_state.h" #include "brw_defines.h" +#include "brw_wm.h" +#include "main/framebuffer.h" /** * Helper function to emit depth related command packets. @@ -40,7 +42,6 @@ emit_depth_packets(struct brw_context *brw, bool depth_writable, struct intel_mipmap_tree *stencil_mt, bool stencil_writable, - uint32_t stencil_offset, bool hiz, uint32_t width, uint32_t height, @@ -48,13 +49,15 @@ emit_depth_packets(struct brw_context *brw, uint32_t lod, uint32_t min_array_element) { + uint32_t mocs_wb = brw->gen >= 9 ? SKL_MOCS_WB : BDW_MOCS_WB; + /* Skip repeated NULL depth/stencil emits (think 2D rendering). */ if (!depth_mt && !stencil_mt && brw->no_depth_or_stencil) { assert(brw->hw_ctx); return; } - intel_emit_depth_stall_flushes(brw); + brw_emit_depth_stall_flushes(brw); /* _NEW_BUFFERS, _NEW_DEPTH, _NEW_STENCIL */ BEGIN_BATCH(8); @@ -73,9 +76,9 @@ emit_depth_packets(struct brw_context *brw, OUT_BATCH(0); } OUT_BATCH(((width - 1) << 4) | ((height - 1) << 18) | lod); - OUT_BATCH(((depth - 1) << 21) | (min_array_element << 10) | BDW_MOCS_WB); + OUT_BATCH(((depth - 1) << 21) | (min_array_element << 10) | mocs_wb); OUT_BATCH(0); - OUT_BATCH(depth_mt ? depth_mt->qpitch >> 2 : 0); + OUT_BATCH(((depth - 1) << 21) | (depth_mt ? depth_mt->qpitch >> 2 : 0)); ADVANCE_BATCH(); if (!hiz) { @@ -89,10 +92,10 @@ emit_depth_packets(struct brw_context *brw, } else { BEGIN_BATCH(5); OUT_BATCH(GEN7_3DSTATE_HIER_DEPTH_BUFFER << 16 | (5 - 2)); - OUT_BATCH((depth_mt->hiz_mt->pitch - 1) | BDW_MOCS_WB << 25); - OUT_RELOC64(depth_mt->hiz_mt->bo, + OUT_BATCH((depth_mt->hiz_buf->pitch - 1) | mocs_wb << 25); + OUT_RELOC64(depth_mt->hiz_buf->bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, 0); - OUT_BATCH(depth_mt->hiz_mt->qpitch >> 2); + OUT_BATCH(depth_mt->hiz_buf->qpitch >> 2); ADVANCE_BATCH(); } @@ -121,11 +124,10 @@ emit_depth_packets(struct brw_context *brw, * page (which would imply that it does). Experiments with the hardware * indicate that it does. */ - OUT_BATCH(HSW_STENCIL_ENABLED | BDW_MOCS_WB << 22 | + OUT_BATCH(HSW_STENCIL_ENABLED | mocs_wb << 22 | (2 * stencil_mt->pitch - 1)); OUT_RELOC64(stencil_mt->bo, - I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, - stencil_offset); + I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, 0); OUT_BATCH(stencil_mt ? stencil_mt->qpitch >> 2 : 0); ADVANCE_BATCH(); } @@ -185,8 +187,20 @@ gen8_emit_depth_stencil_hiz(struct brw_context *brw, depth *= 6; break; case GL_TEXTURE_3D: - assert(rb); - depth = MAX2(rb->Depth, 1); + assert(mt); + depth = MAX2(mt->logical_depth0, 1); + surftype = translate_tex_target(gl_target); + break; + case GL_TEXTURE_1D_ARRAY: + case GL_TEXTURE_1D: + if (brw->gen >= 9) { + /* WaDisable1DDepthStencil. Skylake+ doesn't support 1D depth + * textures but it does allow pretending it's a 2D texture + * instead. + */ + surftype = BRW_SURFACE_2D; + break; + } /* fallthrough */ default: surftype = translate_tex_target(gl_target); @@ -205,10 +219,176 @@ gen8_emit_depth_stencil_hiz(struct brw_context *brw, emit_depth_packets(brw, depth_mt, brw_depthbuffer_format(brw), surftype, ctx->Depth.Mask != 0, stencil_mt, ctx->Stencil._WriteEnabled, - brw->depthstencil.stencil_offset, hiz, width, height, depth, lod, min_array_element); } +/** + * Should we set the PMA FIX ENABLE bit? + * + * To avoid unnecessary depth related stalls, we need to set this bit. + * However, there is a very complicated formula which governs when it + * is legal to do so. This function computes that. + * + * See the documenation for the CACHE_MODE_1 register, bit 11. + */ +static bool +pma_fix_enable(const struct brw_context *brw) +{ + const struct gl_context *ctx = &brw->ctx; + /* _NEW_BUFFERS */ + struct intel_renderbuffer *depth_irb = + intel_get_renderbuffer(ctx->DrawBuffer, BUFFER_DEPTH); + + /* 3DSTATE_WM::ForceThreadDispatch is never used. */ + const bool wm_force_thread_dispatch = false; + + /* 3DSTATE_RASTER::ForceSampleCount is never used. */ + const bool raster_force_sample_count_nonzero = false; + + /* _NEW_BUFFERS: + * 3DSTATE_DEPTH_BUFFER::SURFACE_TYPE != NULL && + * 3DSTATE_DEPTH_BUFFER::HIZ Enable + */ + const bool hiz_enabled = depth_irb && intel_renderbuffer_has_hiz(depth_irb); + + /* BRW_NEW_FS_PROG_DATA: + * 3DSTATE_WM::Early Depth/Stencil Control != EDSC_PREPS (2). + */ + const bool edsc_not_preps = !brw->wm.prog_data->early_fragment_tests; + + /* 3DSTATE_PS_EXTRA::PixelShaderValid is always true. */ + const bool pixel_shader_valid = true; + + /* !(3DSTATE_WM_HZ_OP::DepthBufferClear || + * 3DSTATE_WM_HZ_OP::DepthBufferResolve || + * 3DSTATE_WM_HZ_OP::Hierarchical Depth Buffer Resolve Enable || + * 3DSTATE_WM_HZ_OP::StencilBufferClear) + * + * HiZ operations are done outside of the normal state upload, so they're + * definitely not happening now. + */ + const bool in_hiz_op = false; + + /* _NEW_DEPTH: + * DEPTH_STENCIL_STATE::DepthTestEnable + */ + const bool depth_test_enabled = depth_irb && ctx->Depth.Test; + + /* _NEW_DEPTH: + * 3DSTATE_WM_DEPTH_STENCIL::DepthWriteEnable && + * 3DSTATE_DEPTH_BUFFER::DEPTH_WRITE_ENABLE. + */ + const bool depth_writes_enabled = ctx->Depth.Mask; + + /* _NEW_STENCIL: + * !DEPTH_STENCIL_STATE::Stencil Buffer Write Enable || + * !3DSTATE_DEPTH_BUFFER::Stencil Buffer Enable || + * !3DSTATE_STENCIL_BUFFER::Stencil Buffer Enable + */ + const bool stencil_writes_enabled = ctx->Stencil._WriteEnabled; + + /* BRW_NEW_FS_PROG_DATA: + * 3DSTATE_PS_EXTRA::Pixel Shader Computed Depth Mode != PSCDEPTH_OFF + */ + const bool ps_computes_depth = + brw->wm.prog_data->computed_depth_mode != BRW_PSCDEPTH_OFF; + + /* BRW_NEW_FS_PROG_DATA: 3DSTATE_PS_EXTRA::PixelShaderKillsPixels + * BRW_NEW_FS_PROG_DATA: 3DSTATE_PS_EXTRA::oMask Present to RenderTarget + * _NEW_MULTISAMPLE: 3DSTATE_PS_BLEND::AlphaToCoverageEnable + * _NEW_COLOR: 3DSTATE_PS_BLEND::AlphaTestEnable + * + * 3DSTATE_WM_CHROMAKEY::ChromaKeyKillEnable is always false. + * 3DSTATE_WM::ForceKillPix != ForceOff is always true. + */ + const bool kill_pixel = + brw->wm.prog_data->uses_kill || + brw->wm.prog_data->uses_omask || + (_mesa_is_multisample_enabled(ctx) && ctx->Multisample.SampleAlphaToCoverage) || + ctx->Color.AlphaEnabled; + + /* The big formula in CACHE_MODE_1::NP PMA FIX ENABLE. */ + return !wm_force_thread_dispatch && + !raster_force_sample_count_nonzero && + hiz_enabled && + edsc_not_preps && + pixel_shader_valid && + !in_hiz_op && + depth_test_enabled && + (ps_computes_depth || + (kill_pixel && (depth_writes_enabled || stencil_writes_enabled))); +} + +void +gen8_write_pma_stall_bits(struct brw_context *brw, uint32_t pma_stall_bits) +{ + struct gl_context *ctx = &brw->ctx; + + /* If we haven't actually changed the value, bail now to avoid unnecessary + * pipeline stalls and register writes. + */ + if (brw->pma_stall_bits == pma_stall_bits) + return; + + brw->pma_stall_bits = pma_stall_bits; + + /* According to the PIPE_CONTROL documentation, software should emit a + * PIPE_CONTROL with the CS Stall and Depth Cache Flush bits set prior + * to the LRI. If stencil buffer writes are enabled, then a Render Cache + * Flush is also necessary. + */ + const uint32_t render_cache_flush = + ctx->Stencil._WriteEnabled ? PIPE_CONTROL_RENDER_TARGET_FLUSH : 0; + brw_emit_pipe_control_flush(brw, + PIPE_CONTROL_CS_STALL | + PIPE_CONTROL_DEPTH_CACHE_FLUSH | + render_cache_flush); + + /* CACHE_MODE_1 is a non-privileged register. */ + BEGIN_BATCH(3); + OUT_BATCH(MI_LOAD_REGISTER_IMM | (3 - 2)); + OUT_BATCH(GEN7_CACHE_MODE_1); + OUT_BATCH(GEN8_HIZ_PMA_MASK_BITS | pma_stall_bits); + ADVANCE_BATCH(); + + /* After the LRI, a PIPE_CONTROL with both the Depth Stall and Depth Cache + * Flush bits is often necessary. We do it regardless because it's easier. + * The render cache flush is also necessary if stencil writes are enabled. + */ + brw_emit_pipe_control_flush(brw, + PIPE_CONTROL_DEPTH_STALL | + PIPE_CONTROL_DEPTH_CACHE_FLUSH | + render_cache_flush); + +} + +static void +gen8_emit_pma_stall_workaround(struct brw_context *brw) +{ + uint32_t bits = 0; + + if (brw->gen >= 9) + return; + + if (pma_fix_enable(brw)) + bits |= GEN8_HIZ_NP_PMA_FIX_ENABLE | GEN8_HIZ_NP_EARLY_Z_FAILS_DISABLE; + + gen8_write_pma_stall_bits(brw, bits); +} + +const struct brw_tracked_state gen8_pma_fix = { + .dirty = { + .mesa = _NEW_BUFFERS | + _NEW_COLOR | + _NEW_DEPTH | + _NEW_MULTISAMPLE | + _NEW_STENCIL, + .brw = BRW_NEW_BLORP | + BRW_NEW_FS_PROG_DATA, + }, + .emit = gen8_emit_pma_stall_workaround +}; + /** * Emit packets to perform a depth/HiZ resolve or fast depth/stencil clear. * @@ -222,9 +402,30 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt, if (op == GEN6_HIZ_OP_NONE) return; + /* Disable the PMA stall fix since we're about to do a HiZ operation. */ + if (brw->gen == 8) + gen8_write_pma_stall_bits(brw, 0); + assert(mt->first_level == 0); assert(mt->logical_depth0 >= 1); + /* If we're operating on LOD 0, align to 8x4 to meet the alignment + * requirements for most HiZ operations. Otherwise, use the actual size + * to allow the hardware to calculate the miplevel offsets correctly. + */ + uint32_t surface_width = ALIGN(mt->logical_width0, level == 0 ? 8 : 1); + uint32_t surface_height = ALIGN(mt->logical_height0, level == 0 ? 4 : 1); + + /* From the documentation for 3DSTATE_WM_HZ_OP: "3DSTATE_MULTISAMPLE packet + * must be used prior to this packet to change the Number of Multisamples. + * This packet must not be used to change Number of Multisamples in a + * rendering sequence." + */ + if (brw->num_samples != mt->num_samples) { + gen8_emit_3dstate_multisample(brw, mt->num_samples); + brw->NewGLState |= _NEW_MULTISAMPLE; + } + /* The basic algorithm is: * - If needed, emit 3DSTATE_{DEPTH,HIER_DEPTH,STENCIL}_BUFFER and * 3DSTATE_CLEAR_PARAMS packets to set up the relevant buffers. @@ -237,16 +438,21 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt, brw_depth_format(brw, mt->format), BRW_SURFACE_2D, true, /* depth writes */ - NULL, false, 0, /* no stencil for now */ + NULL, false, /* no stencil for now */ true, /* hiz */ - mt->logical_width0, - mt->logical_height0, + surface_width, + surface_height, mt->logical_depth0, level, layer); /* min_array_element */ - unsigned rect_width = minify(mt->logical_width0, level); - unsigned rect_height = minify(mt->logical_height0, level); + /* Depth buffer clears and HiZ resolves must use an 8x4 aligned rectangle. + * Note that intel_miptree_level_enable_hiz disables HiZ for miplevels > 0 + * which aren't 8x4 aligned, so expanding the size is safe - it'll just + * draw into empty padding space. + */ + unsigned rect_width = ALIGN(minify(mt->logical_width0, level), 8); + unsigned rect_height = ALIGN(minify(mt->logical_height0, level), 4); BEGIN_BATCH(4); OUT_BATCH(_3DSTATE_DRAWING_RECTANGLE << 16 | (4 - 2)); @@ -271,7 +477,7 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt, dw1 |= GEN8_WM_HZ_DEPTH_CLEAR; break; case GEN6_HIZ_OP_NONE: - assert(!"Should not get here."); + unreachable("Should not get here."); } if (mt->num_samples > 0) @@ -292,7 +498,7 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt, */ brw_emit_pipe_control_write(brw, PIPE_CONTROL_WRITE_IMMEDIATE, - brw->batch.workaround_bo, 0, 0, 0); + brw->workaround_bo, 0, 0, 0); /* Emit 3DSTATE_WM_HZ_OP again to disable the state overrides. */ BEGIN_BATCH(5); @@ -312,5 +518,5 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt, * * Setting _NEW_DEPTH and _NEW_BUFFERS covers it, but is rather overkill. */ - brw->state.dirty.mesa |= _NEW_DEPTH | _NEW_BUFFERS; + brw->NewGLState |= _NEW_DEPTH | _NEW_BUFFERS; }