X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fsvga%2Fsvga_state_framebuffer.c;h=b1362a981689ffe0ca183db9ca5a1ea159f0fc5f;hb=c72dcd9a718628638957bfd99549bf788c2e6b36;hp=cf3310ac14a70a9f807533ea48096823863ccdd8;hpb=2786343896030ed4abcf6674a3e66f3a0d246e0d;p=mesa.git diff --git a/src/gallium/drivers/svga/svga_state_framebuffer.c b/src/gallium/drivers/svga/svga_state_framebuffer.c index cf3310ac14a..b1362a98168 100644 --- a/src/gallium/drivers/svga/svga_state_framebuffer.c +++ b/src/gallium/drivers/svga/svga_state_framebuffer.c @@ -26,57 +26,69 @@ #include "util/u_inlines.h" #include "pipe/p_defines.h" #include "util/u_math.h" +#include "util/u_format.h" #include "svga_context.h" #include "svga_state.h" #include "svga_cmd.h" #include "svga_debug.h" +#include "svga_screen.h" +#include "svga_surface.h" -/*********************************************************************** - * Hardware state update +/* + * flush our command buffer after the 8th distinct render target + * + * This helps improve the surface cache behaviour in the face of the + * large number of single-use render targets generated by EXA and the xorg + * state tracker. Without this we can reference hundreds of individual + * render targets from a command buffer, which leaves little scope for + * sharing or reuse of those targets. */ +#define MAX_RT_PER_BATCH 8 + static enum pipe_error -emit_framebuffer( struct svga_context *svga, - unsigned dirty ) +emit_fb_vgpu9(struct svga_context *svga) { + struct svga_screen *svgascreen = svga_screen(svga->pipe.screen); const struct pipe_framebuffer_state *curr = &svga->curr.framebuffer; struct pipe_framebuffer_state *hw = &svga->state.hw_clear.framebuffer; - boolean reemit = svga->rebind.rendertargets; + boolean reemit = svga->rebind.flags.rendertargets; unsigned i; enum pipe_error ret; + assert(!svga_have_vgpu10(svga)); + /* * We need to reemit non-null surface bindings, even when they are not * dirty, to ensure that the resources are paged in. */ - - for(i = 0; i < PIPE_MAX_COLOR_BUFS; ++i) { - if (curr->cbufs[i] != hw->cbufs[i] || - (reemit && hw->cbufs[i])) { - if (svga->curr.nr_fbs++ > 8) + + for (i = 0; i < svgascreen->max_color_buffers; i++) { + if ((curr->cbufs[i] != hw->cbufs[i]) || (reemit && hw->cbufs[i])) { + if (svga->curr.nr_fbs++ > MAX_RT_PER_BATCH) return PIPE_ERROR_OUT_OF_MEMORY; - ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_COLOR0 + i, curr->cbufs[i]); + ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_COLOR0 + i, + curr->cbufs[i]); if (ret != PIPE_OK) return ret; - + pipe_surface_reference(&hw->cbufs[i], curr->cbufs[i]); } } - - if (curr->zsbuf != hw->zsbuf || - (reemit && hw->zsbuf)) { + if ((curr->zsbuf != hw->zsbuf) || (reemit && hw->zsbuf)) { ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_DEPTH, curr->zsbuf); if (ret != PIPE_OK) return ret; if (curr->zsbuf && - curr->zsbuf->format == PIPE_FORMAT_S8_UINT_Z24_UNORM) { - ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_STENCIL, curr->zsbuf); + util_format_is_depth_and_stencil(curr->zsbuf->format)) { + ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_STENCIL, + curr->zsbuf); if (ret != PIPE_OK) return ret; } @@ -85,12 +97,10 @@ emit_framebuffer( struct svga_context *svga, if (ret != PIPE_OK) return ret; } - + pipe_surface_reference(&hw->zsbuf, curr->zsbuf); } - svga->rebind.rendertargets = FALSE; - return PIPE_OK; } @@ -103,18 +113,20 @@ emit_framebuffer( struct svga_context *svga, * Called at the beginning of every new command buffer to ensure that * non-dirty rendertargets are properly paged-in. */ -enum pipe_error -svga_reemit_framebuffer_bindings(struct svga_context *svga) +static enum pipe_error +svga_reemit_framebuffer_bindings_vgpu9(struct svga_context *svga) { + struct svga_screen *svgascreen = svga_screen(svga->pipe.screen); struct pipe_framebuffer_state *hw = &svga->state.hw_clear.framebuffer; unsigned i; enum pipe_error ret; - assert(svga->rebind.rendertargets); + assert(!svga_have_vgpu10(svga)); - for (i = 0; i < MIN2(PIPE_MAX_COLOR_BUFS, 8); ++i) { + for (i = 0; i < svgascreen->max_color_buffers; i++) { if (hw->cbufs[i]) { - ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_COLOR0 + i, hw->cbufs[i]); + ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_COLOR0 + i, + hw->cbufs[i]); if (ret != PIPE_OK) { return ret; } @@ -128,7 +140,7 @@ svga_reemit_framebuffer_bindings(struct svga_context *svga) } if (hw->zsbuf && - hw->zsbuf->format == PIPE_FORMAT_S8_UINT_Z24_UNORM) { + util_format_is_depth_and_stencil(hw->zsbuf->format)) { ret = SVGA3D_SetRenderTarget(svga->swc, SVGA3D_RT_STENCIL, hw->zsbuf); if (ret != PIPE_OK) { return ret; @@ -142,13 +154,181 @@ svga_reemit_framebuffer_bindings(struct svga_context *svga) } } - svga->rebind.rendertargets = FALSE; + return PIPE_OK; +} + + + +static enum pipe_error +emit_fb_vgpu10(struct svga_context *svga) +{ + const struct svga_screen *ss = svga_screen(svga->pipe.screen); + struct pipe_surface *rtv[SVGA3D_MAX_RENDER_TARGETS]; + struct pipe_surface *dsv; + struct pipe_framebuffer_state *curr = &svga->curr.framebuffer; + struct pipe_framebuffer_state *hw = &svga->state.hw_clear.framebuffer; + const unsigned num_color = MAX2(curr->nr_cbufs, hw->nr_cbufs); + int last_rtv = -1; + unsigned i; + enum pipe_error ret = PIPE_OK; + + assert(svga_have_vgpu10(svga)); + + /* Setup render targets array. Note that we loop over the max of the + * number of previously bound buffers and the new buffers to unbind + * any previously bound buffers when the new number of buffers is less + * than the old number of buffers. + */ + for (i = 0; i < num_color; i++) { + if (curr->cbufs[i]) { + rtv[i] = svga_validate_surface_view(svga, + svga_surface(curr->cbufs[i])); + if (rtv[i] == NULL) { + return PIPE_ERROR_OUT_OF_MEMORY; + } + + assert(svga_surface(rtv[i])->view_id != SVGA3D_INVALID_ID); + last_rtv = i; + } + else { + rtv[i] = NULL; + } + } + + /* Setup depth stencil view */ + if (curr->zsbuf) { + dsv = svga_validate_surface_view(svga, svga_surface(curr->zsbuf)); + if (!dsv) { + return PIPE_ERROR_OUT_OF_MEMORY; + } + } + else { + dsv = NULL; + } + + /* avoid emitting redundant SetRenderTargets command */ + if ((num_color != svga->state.hw_draw.num_rendertargets) || + (dsv != svga->state.hw_draw.dsv) || + memcmp(rtv, svga->state.hw_draw.rtv, num_color * sizeof(rtv[0]))) { + + ret = SVGA3D_vgpu10_SetRenderTargets(svga->swc, num_color, rtv, dsv); + if (ret != PIPE_OK) + return ret; + + /* number of render targets sent to the device, not including trailing + * unbound render targets. + */ + svga->state.hw_draw.num_rendertargets = last_rtv + 1; + svga->state.hw_draw.dsv = dsv; + memcpy(svga->state.hw_draw.rtv, rtv, num_color * sizeof(rtv[0])); + + for (i = 0; i < ss->max_color_buffers; i++) { + if (hw->cbufs[i] != curr->cbufs[i]) { + /* propagate the backed view surface before unbinding it */ + if (hw->cbufs[i] && svga_surface(hw->cbufs[i])->backed) { + svga_propagate_surface(svga, + &svga_surface(hw->cbufs[i])->backed->base); + } + pipe_surface_reference(&hw->cbufs[i], curr->cbufs[i]); + } + } + hw->nr_cbufs = curr->nr_cbufs; + + if (hw->zsbuf != curr->zsbuf) { + /* propagate the backed view surface before unbinding it */ + if (hw->zsbuf && svga_surface(hw->zsbuf)->backed) { + svga_propagate_surface(svga, &svga_surface(hw->zsbuf)->backed->base); + } + pipe_surface_reference(&hw->zsbuf, curr->zsbuf); + } + } + + return ret; +} + + +static enum pipe_error +emit_framebuffer(struct svga_context *svga, unsigned dirty) +{ + if (svga_have_vgpu10(svga)) { + return emit_fb_vgpu10(svga); + } + else { + return emit_fb_vgpu9(svga); + } +} + + +/* + * Rebind rendertargets. + * + * Similar to emit_framebuffer, but without any state checking/update. + * + * Called at the beginning of every new command buffer to ensure that + * non-dirty rendertargets are properly paged-in. + */ +enum pipe_error +svga_reemit_framebuffer_bindings(struct svga_context *svga) +{ + enum pipe_error ret; + + assert(svga->rebind.flags.rendertargets); + + if (svga_have_vgpu10(svga)) { + ret = emit_fb_vgpu10(svga); + } + else { + ret = svga_reemit_framebuffer_bindings_vgpu9(svga); + } + + svga->rebind.flags.rendertargets = FALSE; + + return ret; +} + + +/* + * Send a private allocation command to page in rendertargets resource. + */ +enum pipe_error +svga_rebind_framebuffer_bindings(struct svga_context *svga) +{ + struct svga_hw_draw_state *hw = &svga->state.hw_draw; + unsigned i; + enum pipe_error ret; + + assert(svga_have_vgpu10(svga)); + + if (!svga->rebind.flags.rendertargets) + return PIPE_OK; + + for (i = 0; i < hw->num_rendertargets; i++) { + if (hw->rtv[i]) { + ret = svga->swc->resource_rebind(svga->swc, + svga_surface(hw->rtv[i])->handle, + NULL, + SVGA_RELOC_WRITE); + if (ret != PIPE_OK) + return ret; + } + } + + if (hw->dsv) { + ret = svga->swc->resource_rebind(svga->swc, + svga_surface(hw->dsv)->handle, + NULL, + SVGA_RELOC_WRITE); + if (ret != PIPE_OK) + return ret; + } + + svga->rebind.flags.rendertargets = 0; return PIPE_OK; } -struct svga_tracked_state svga_hw_framebuffer = +struct svga_tracked_state svga_hw_framebuffer = { "hw framebuffer state", SVGA_NEW_FRAME_BUFFER, @@ -158,7 +338,7 @@ struct svga_tracked_state svga_hw_framebuffer = -/*********************************************************************** +/*********************************************************************** */ static enum pipe_error @@ -178,13 +358,14 @@ emit_viewport( struct svga_context *svga, boolean invertY = FALSE; enum pipe_error ret; - float fb_width = svga->curr.framebuffer.width; - float fb_height = svga->curr.framebuffer.height; + float fb_width = (float) svga->curr.framebuffer.width; + float fb_height = (float) svga->curr.framebuffer.height; - float fx = viewport->scale[0] * -1.0 + viewport->translate[0]; - float fy = flip * viewport->scale[1] * -1.0 + viewport->translate[1]; - float fw = viewport->scale[0] * 2; - float fh = flip * viewport->scale[1] * 2; + float fx = viewport->scale[0] * -1.0f + viewport->translate[0]; + float fy = flip * viewport->scale[1] * -1.0f + viewport->translate[1]; + float fw = viewport->scale[0] * 2.0f; + float fh = flip * viewport->scale[1] * 2.0f; + boolean emit_vgpu10_viewport = FALSE; memset( &prescale, 0, sizeof(prescale) ); @@ -208,30 +389,44 @@ emit_viewport( struct svga_context *svga, prescale.translate[1] = 0; prescale.translate[2] = 0; prescale.translate[3] = 0; - prescale.enabled = TRUE; - + /* Enable prescale to adjust vertex positions to match + VGPU10 convention only if rasterization is enabled. + */ + if (svga->curr.rast->templ.rasterizer_discard) { + degenerate = TRUE; + goto out; + } else { + prescale.enabled = TRUE; + } if (fw < 0) { - prescale.scale[0] *= -1.0; + prescale.scale[0] *= -1.0f; prescale.translate[0] += -fw; fw = -fw; - fx = viewport->scale[0] * 1.0 + viewport->translate[0]; + fx = viewport->scale[0] * 1.0f + viewport->translate[0]; } if (fh < 0.0) { - prescale.translate[1] = fh - 1 + fy * 2; + if (svga_have_vgpu10(svga)) { + /* floating point viewport params below */ + prescale.translate[1] = fh + fy * 2.0f; + } + else { + /* integer viewport params below */ + prescale.translate[1] = fh - 1.0f + fy * 2.0f; + } fh = -fh; fy -= fh; - prescale.scale[1] = -1.0; + prescale.scale[1] = -1.0f; invertY = TRUE; } if (fx < 0) { prescale.translate[0] += fx; - prescale.scale[0] *= fw / (fw + fx); + prescale.scale[0] *= fw / (fw + fx); fw += fx; - fx = 0; + fx = 0.0f; } if (fy < 0) { @@ -241,17 +436,16 @@ emit_viewport( struct svga_context *svga, else { prescale.translate[1] += fy; } - prescale.scale[1] *= fh / (fh + fy); + prescale.scale[1] *= fh / (fh + fy); fh += fy; - fy = 0; + fy = 0.0f; } if (fx + fw > fb_width) { - prescale.scale[0] *= fw / (fb_width - fx); + prescale.scale[0] *= fw / (fb_width - fx); prescale.translate[0] -= fx * (fw / (fb_width - fx)); prescale.translate[0] += fx; fw = fb_width - fx; - } if (fy + fh > fb_height) { @@ -274,17 +468,16 @@ emit_viewport( struct svga_context *svga, goto out; } - /* D3D viewport is integer space. Convert fx,fy,etc. to * integers. * * TODO: adjust pretranslate correct for any subpixel error * introduced converting to integers. */ - rect.x = fx; - rect.y = fy; - rect.w = fw; - rect.h = fh; + rect.x = (uint32) fx; + rect.y = (uint32) fy; + rect.w = (uint32) fw; + rect.h = (uint32) fh; SVGA_DBG(DEBUG_VIEWPORT, "viewport error %f,%f %fx%f\n", @@ -300,25 +493,39 @@ emit_viewport( struct svga_context *svga, rect.w, rect.h); - /* Finally, to get GL rasterization rules, need to tweak the * screen-space coordinates slightly relative to D3D which is * what hardware implements natively. */ - if (svga->curr.rast->templ.gl_rasterization_rules) { + if (svga->curr.rast->templ.half_pixel_center) { float adjust_x = 0.0; float adjust_y = 0.0; - switch (svga->curr.reduced_prim) { - case PIPE_PRIM_LINES: - adjust_x = -0.5; - adjust_y = 0; - break; - case PIPE_PRIM_POINTS: - case PIPE_PRIM_TRIANGLES: - adjust_x = -0.5; - adjust_y = -0.5; - break; + if (svga_have_vgpu10(svga)) { + /* Normally, we don't have to do any sub-pixel coordinate + * adjustments for VGPU10. But when we draw wide points with + * a GS we need an X adjustment in order to be conformant. + */ + if (svga->curr.reduced_prim == PIPE_PRIM_POINTS && + svga->curr.rast->pointsize > 1.0f) { + adjust_x = 0.5; + } + } + else { + switch (svga->curr.reduced_prim) { + case PIPE_PRIM_POINTS: + adjust_x = -0.375; + adjust_y = -0.75; + break; + case PIPE_PRIM_LINES: + adjust_x = -0.5; + adjust_y = 0; + break; + case PIPE_PRIM_TRIANGLES: + adjust_x = -0.5; + adjust_y = -0.5; + break; + } } if (invertY) @@ -330,9 +537,8 @@ emit_viewport( struct svga_context *svga, prescale.scale[2] = 0.5; /* D3D clip space */ } - - range_min = viewport->scale[2] * -1.0 + viewport->translate[2]; - range_max = viewport->scale[2] * 1.0 + viewport->translate[2]; + range_min = viewport->scale[2] * -1.0f + viewport->translate[2]; + range_max = viewport->scale[2] * 1.0f + viewport->translate[2]; /* D3D (and by implication SVGA) doesn't like dealing with zmax * less than zmin. Detect that case, flip the depth range and @@ -340,10 +546,21 @@ emit_viewport( struct svga_context *svga, */ if (range_min > range_max) { float range_tmp; - range_tmp = range_min; - range_min = range_max; + range_tmp = range_min; + range_min = range_max; range_max = range_tmp; - prescale.scale[2] = -prescale.scale[2]; + prescale.scale[2] = -prescale.scale[2]; + } + + /* If zmin is less than 0, clamp zmin to 0 and adjust the prescale. + * zmin can be set to -1 when viewport->scale[2] is set to 1 and + * viewport->translate[2] is set to 0 in the blit code. + */ + if (range_min < 0.0f) { + range_min = -0.5f * viewport->scale[2] + 0.5f + viewport->translate[2]; + range_max = 0.5f * viewport->scale[2] + 0.5f + viewport->translate[2]; + prescale.scale[2] *= 2.0f; + prescale.translate[2] -= 0.5f; } if (prescale.enabled) { @@ -358,10 +575,10 @@ emit_viewport( struct svga_context *svga, prescale.scale[0], prescale.scale[1]); - H[0] = (float)rect.w / 2.0; - H[1] = -(float)rect.h / 2.0; - J[0] = (float)rect.x + (float)rect.w / 2.0; - J[1] = (float)rect.y + (float)rect.h / 2.0; + H[0] = (float)rect.w / 2.0f; + H[1] = -(float)rect.h / 2.0f; + J[0] = (float)rect.x + (float)rect.w / 2.0f; + J[1] = (float)rect.y + (float)rect.h / 2.0f; SVGA_DBG(DEBUG_VIEWPORT, "H %f,%f\n" @@ -374,7 +591,7 @@ emit_viewport( struct svga_context *svga, /* Adjust prescale to take into account the fact that it is * going to be applied prior to the perspective divide and * viewport transformation. - * + * * Vwin = H(Vc/Vc.w) + J * * We want to tweak Vwin with scale and translation from above, @@ -394,7 +611,7 @@ emit_viewport( struct svga_context *svga, */ for (i = 0; i < 2; i++) { prescale.translate[i] = ((prescale.translate[i] + - (prescale.scale[i] - 1.0) * J[i]) / H[i]); + (prescale.scale[i] - 1.0f) * J[i]) / H[i]); } SVGA_DBG(DEBUG_VIEWPORT, @@ -414,22 +631,49 @@ out: prescale.enabled = FALSE; } - if (memcmp(&rect, &svga->state.hw_clear.viewport, sizeof(rect)) != 0) { - ret = SVGA3D_SetViewport(svga->swc, &rect); - if(ret != PIPE_OK) - return ret; + if (!svga_rects_equal(&rect, &svga->state.hw_clear.viewport)) { + if (svga_have_vgpu10(svga)) { + emit_vgpu10_viewport = TRUE; + } + else { + ret = SVGA3D_SetViewport(svga->swc, &rect); + if (ret != PIPE_OK) + return ret; - memcpy(&svga->state.hw_clear.viewport, &rect, sizeof(rect)); - assert(sizeof(rect) == sizeof(svga->state.hw_clear.viewport)); + svga->state.hw_clear.viewport = rect; + } } if (svga->state.hw_clear.depthrange.zmin != range_min || - svga->state.hw_clear.depthrange.zmax != range_max) + svga->state.hw_clear.depthrange.zmax != range_max) { - ret = SVGA3D_SetZRange(svga->swc, range_min, range_max ); - if(ret != PIPE_OK) + if (svga_have_vgpu10(svga)) { + emit_vgpu10_viewport = TRUE; + } + else { + ret = SVGA3D_SetZRange(svga->swc, range_min, range_max ); + if (ret != PIPE_OK) + return ret; + + svga->state.hw_clear.depthrange.zmin = range_min; + svga->state.hw_clear.depthrange.zmax = range_max; + } + } + + if (emit_vgpu10_viewport) { + SVGA3dViewport vp; + vp.x = (float) rect.x; + vp.y = (float) rect.y; + vp.width = (float) rect.w; + vp.height = (float) rect.h; + vp.minDepth = range_min; + vp.maxDepth = range_max; + ret = SVGA3D_vgpu10_SetViewports(svga->swc, 1, &vp); + if (ret != PIPE_OK) return ret; + svga->state.hw_clear.viewport = rect; + svga->state.hw_clear.depthrange.zmin = range_min; svga->state.hw_clear.depthrange.zmax = range_max; } @@ -443,7 +687,7 @@ out: } -struct svga_tracked_state svga_hw_viewport = +struct svga_tracked_state svga_hw_viewport = { "hw viewport state", ( SVGA_NEW_FRAME_BUFFER | @@ -462,18 +706,31 @@ emit_scissor_rect( struct svga_context *svga, unsigned dirty ) { const struct pipe_scissor_state *scissor = &svga->curr.scissor; - SVGA3dRect rect; - rect.x = scissor->minx; - rect.y = scissor->miny; - rect.w = scissor->maxx - scissor->minx; /* + 1 ?? */ - rect.h = scissor->maxy - scissor->miny; /* + 1 ?? */ - - return SVGA3D_SetScissorRect(svga->swc, &rect); + if (svga_have_vgpu10(svga)) { + SVGASignedRect rect; + + rect.left = scissor->minx; + rect.top = scissor->miny; + rect.right = scissor->maxx; + rect.bottom = scissor->maxy; + + return SVGA3D_vgpu10_SetScissorRects(svga->swc, 1, &rect); + } + else { + SVGA3dRect rect; + + rect.x = scissor->minx; + rect.y = scissor->miny; + rect.w = scissor->maxx - scissor->minx; /* + 1 ?? */ + rect.h = scissor->maxy - scissor->miny; /* + 1 ?? */ + + return SVGA3D_SetScissorRect(svga->swc, &rect); + } } -struct svga_tracked_state svga_hw_scissor = +struct svga_tracked_state svga_hw_scissor = { "hw scissor state", SVGA_NEW_SCISSOR, @@ -514,16 +771,22 @@ emit_clip_planes( struct svga_context *svga, plane[2] = 2.0f * c; plane[3] = d - c; - ret = SVGA3D_SetClipPlane(svga->swc, i, plane); - if(ret != PIPE_OK) - return ret; + if (svga_have_vgpu10(svga)) { + //debug_printf("XXX emit DX10 clip plane\n"); + ret = PIPE_OK; + } + else { + ret = SVGA3D_SetClipPlane(svga->swc, i, plane); + if (ret != PIPE_OK) + return ret; + } } return PIPE_OK; } -struct svga_tracked_state svga_hw_clip_planes = +struct svga_tracked_state svga_hw_clip_planes = { "hw viewport state", SVGA_NEW_CLIP,