diff options
Diffstat (limited to 'src/gallium/drivers/r300')
| -rw-r--r-- | src/gallium/drivers/r300/r300_context.c | 4 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_context.h | 14 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_cs.h | 18 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_debug.c | 90 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_debug.h | 146 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_emit.c | 83 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_emit.h | 2 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_flush.c | 2 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_render.c | 23 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_state.c | 9 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_state_tcl.c | 123 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_state_tcl.h | 16 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_surface.c | 7 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_texture.c | 2 | ||||
| -rw-r--r-- | src/gallium/drivers/r300/r300_winsys.h | 55 |
15 files changed, 416 insertions, 178 deletions
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c index 6bdf544a05..a4e89c37d1 100644 --- a/src/gallium/drivers/r300/r300_context.c +++ b/src/gallium/drivers/r300/r300_context.c @@ -34,10 +34,6 @@ static boolean r300_draw_range_elements(struct pipe_context* pipe, struct r300_context* r300 = r300_context(pipe); int i; - if (r300->dirty_state) { - r300_emit_dirty_state(r300); - } - for (i = 0; i < r300->vertex_buffer_count; i++) { void* buf = pipe_buffer_map(pipe->screen, r300->vertex_buffers[i].buffer, diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h index 4c695c1195..96f1f11246 100644 --- a/src/gallium/drivers/r300/r300_context.h +++ b/src/gallium/drivers/r300/r300_context.h @@ -264,6 +264,11 @@ struct r300_context { /* Draw module. Used mostly for SW TCL. */ struct draw_context* draw; + /* Vertex buffer for rendering. */ + struct pipe_buffer* vbo; + /* Offset into the VBO. */ + size_t vbo_offset; + /* Various CSO state objects. */ /* Blend state. */ struct r300_blend_state* blend_state; @@ -289,7 +294,7 @@ struct r300_context { /* Texture states. */ struct r300_texture* textures[8]; int texture_count; - /* Vertex buffers. */ + /* Vertex buffers for Gallium. */ struct pipe_vertex_buffer vertex_buffers[PIPE_MAX_ATTRIBS]; int vertex_buffer_count; /* Vertex information. */ @@ -314,11 +319,4 @@ struct draw_stage* r300_draw_stage(struct r300_context* r300); void r300_init_state_functions(struct r300_context* r300); void r300_init_surface_functions(struct r300_context* r300); -/* Fun with includes: r300_winsys also declares this prototype. - * We'll just step out in that case... */ -#ifndef R300_WINSYS_H -struct pipe_context* r300_create_context(struct pipe_screen* screen, - struct r300_winsys* r300_winsys); -#endif - #endif /* R300_CONTEXT_H */ diff --git a/src/gallium/drivers/r300/r300_cs.h b/src/gallium/drivers/r300/r300_cs.h index 5d9799dd72..82a3942248 100644 --- a/src/gallium/drivers/r300/r300_cs.h +++ b/src/gallium/drivers/r300/r300_cs.h @@ -49,27 +49,27 @@ #define CS_LOCALS(context) \ struct r300_winsys* cs_winsys = context->winsys; \ - struct radeon_cs* cs = cs_winsys->cs; \ int cs_count = 0; #define CHECK_CS(size) \ - cs_winsys->check_cs(cs, (size)) + cs_winsys->check_cs(cs_winsys, (size)) #define BEGIN_CS(size) do { \ CHECK_CS(size); \ debug_printf("r300: BEGIN_CS, count %d, in %s (%s:%d)\n", \ size, __FUNCTION__, __FILE__, __LINE__); \ - cs_winsys->begin_cs(cs, (size), __FILE__, __FUNCTION__, __LINE__); \ + cs_winsys->begin_cs(cs_winsys, (size), \ + __FILE__, __FUNCTION__, __LINE__); \ cs_count = size; \ } while (0) #define OUT_CS(value) do { \ - cs_winsys->write_cs_dword(cs, (value)); \ + cs_winsys->write_cs_dword(cs_winsys, (value)); \ cs_count--; \ } while (0) #define OUT_CS_32F(value) do { \ - cs_winsys->write_cs_dword(cs, fui(value)); \ + cs_winsys->write_cs_dword(cs_winsys, fui(value)); \ cs_count--; \ } while (0) @@ -97,7 +97,7 @@ bo, offset); \ assert(bo); \ OUT_CS(offset); \ - cs_winsys->write_cs_reloc(cs, bo, rd, wd, flags); \ + cs_winsys->write_cs_reloc(cs_winsys, bo, rd, wd, flags); \ cs_count -= 2; \ } while (0) @@ -106,13 +106,13 @@ __LINE__); \ if (cs_count != 0) \ debug_printf("r300: Warning: cs_count off by %d\n", cs_count); \ - cs_winsys->end_cs(cs, __FILE__, __FUNCTION__, __LINE__); \ + cs_winsys->end_cs(cs_winsys, __FILE__, __FUNCTION__, __LINE__); \ } while (0) #define FLUSH_CS do { \ debug_printf("r300: FLUSH_CS in %s (%s:%d)\n\n", __FUNCTION__, __FILE__, \ __LINE__); \ - cs_winsys->flush_cs(cs); \ + cs_winsys->flush_cs(cs_winsys); \ } while (0) #define RADEON_ONE_REG_WR (1 << 15) @@ -138,7 +138,7 @@ assert(bo); \ OUT_CS(offset); \ OUT_CS(count); \ - cs_winsys->write_cs_reloc(cs, bo, rd, wd, flags); \ + cs_winsys->write_cs_reloc(cs_winsys, bo, rd, wd, flags); \ cs_count -= 2; \ } while (0) diff --git a/src/gallium/drivers/r300/r300_debug.c b/src/gallium/drivers/r300/r300_debug.c index dd63136c9d..1ff72172eb 100644 --- a/src/gallium/drivers/r300/r300_debug.c +++ b/src/gallium/drivers/r300/r300_debug.c @@ -30,81 +30,6 @@ static void r300_dump_fs(struct r300_fragment_shader* fs) } } -static char* r500_fs_swiz[] = { - " R", - " G", - " B", - " A", - " 0", - ".5", - " 1", - " U", -}; - -static char* r500_fs_op_rgb[] = { - "MAD", - "DP3", - "DP4", - "D2A", - "MIN", - "MAX", - "---", - "CND", - "CMP", - "FRC", - "SOP", - "MDH", - "MDV", -}; - -static char* r500_fs_op_alpha[] = { - "MAD", - " DP", - "MIN", - "MAX", - "---", - "CND", - "CMP", - "FRC", - "EX2", - "LN2", - "RCP", - "RSQ", - "SIN", - "COS", - "MDH", - "MDV", -}; - -static char* r500_fs_mask[] = { - "NONE", - "R ", - " G ", - "RG ", - " B ", - "R B ", - " GB ", - "RGB ", - " A", - "R A", - " G A", - "RG A", - " BA", - "R BA", - " GBA", - "RGBA", -}; - -static char* r500_fs_tex[] = { - " NOP", - " LD", - "TEXKILL", - " PROJ", - "LODBIAS", - " LOD", - " DXDY", -}; - void r500_fs_dump(struct r500_fragment_shader* fs) { int i; @@ -225,12 +150,25 @@ void r500_fs_dump(struct r500_fragment_shader* fs) } } +static void r300_vs_op_dump(uint32_t op) +{ + if (op & 0x81) { + debug_printf("PVS_MACRO_OP_2CLK_M2X_ADD\n"); + } else if (op & 0x80) { + debug_printf(" PVS_MACRO_OP_2CLK_MADD\n"); + } else if (op & 0x40) { + debug_printf("%s\n", r300_vs_me_ops[op & 0x1f]); + } else { + debug_printf("%s\n", r300_vs_ve_ops[op & 0x1f]); + } +} + void r300_vs_dump(struct r300_vertex_shader* vs) { int i; for (i = 0; i < vs->instruction_count; i++) { - debug_printf("inst0: 0x%x\n", vs->instructions[i].inst0); + r300_vs_op_dump(vs->instructions[i].inst0); debug_printf("inst1: 0x%x\n", vs->instructions[i].inst1); debug_printf("inst2: 0x%x\n", vs->instructions[i].inst2); debug_printf("inst3: 0x%x\n", vs->instructions[i].inst3); diff --git a/src/gallium/drivers/r300/r300_debug.h b/src/gallium/drivers/r300/r300_debug.h index a1f873656d..6306594099 100644 --- a/src/gallium/drivers/r300/r300_debug.h +++ b/src/gallium/drivers/r300/r300_debug.h @@ -27,6 +27,152 @@ #include "r300_state_shader.h" #include "r300_state_tcl.h" +static char* r500_fs_swiz[] = { + " R", + " G", + " B", + " A", + " 0", + ".5", + " 1", + " U", +}; + +static char* r500_fs_op_rgb[] = { + "MAD", + "DP3", + "DP4", + "D2A", + "MIN", + "MAX", + "---", + "CND", + "CMP", + "FRC", + "SOP", + "MDH", + "MDV", +}; + +static char* r500_fs_op_alpha[] = { + "MAD", + " DP", + "MIN", + "MAX", + "---", + "CND", + "CMP", + "FRC", + "EX2", + "LN2", + "RCP", + "RSQ", + "SIN", + "COS", + "MDH", + "MDV", +}; + +static char* r500_fs_mask[] = { + "NONE", + "R ", + " G ", + "RG ", + " B ", + "R B ", + " GB ", + "RGB ", + " A", + "R A", + " G A", + "RG A", + " BA", + "R BA", + " GBA", + "RGBA", +}; + +static char* r500_fs_tex[] = { + " NOP", + " LD", + "TEXKILL", + " PROJ", + "LODBIAS", + " LOD", + " DXDY", +}; + +static char* r300_vs_ve_ops[] = { + /* R300 vector ops */ + " VE_NO_OP", + " VE_DOT_PRODUCT", + " VE_MULTIPLY", + " VE_ADD", + " VE_MULTIPLY_ADD", + " VE_DISTANCE_FACTOR", + " VE_FRACTION", + " VE_MAXIMUM", + " VE_MINIMUM", + "VE_SET_GREATER_THAN_EQUAL", + " VE_SET_LESS_THAN", + " VE_MULTIPLYX2_ADD", + " VE_MULTIPLY_CLAMP", + " VE_FLT2FIX_DX", + " VE_FLT2FIX_DX_RND", + /* R500 vector ops */ + " VE_PRED_SET_EQ_PUSH", + " VE_PRED_SET_GT_PUSH", + " VE_PRED_SET_GTE_PUSH", + " VE_PRED_SET_NEQ_PUSH", + " VE_COND_WRITE_EQ", + " VE_COND_WRITE_GT", + " VE_COND_WRITE_GTE", + " VE_COND_WRITE_NEQ", + " VE_SET_GREATER_THAN", + " VE_SET_EQUAL", + " VE_SET_NOT_EQUAL", + " (reserved)", + " (reserved)", + " (reserved)", +}; + +static char* r300_vs_me_ops[] = { + /* R300 math ops */ + " ME_NO_OP", + " ME_EXP_BASE2_DX", + " ME_LOG_BASE2_DX", + " ME_EXP_BASEE_FF", + " ME_LIGHT_COEFF_DX", + " ME_POWER_FUNC_FF", + " ME_RECIP_DX", + " ME_RECIP_FF", + " ME_RECIP_SQRT_DX", + " ME_RECIP_SQRT_FF", + " ME_MULTIPLY", + " ME_EXP_BASE2_FULL_DX", + " ME_LOG_BASE2_FULL_DX", + " ME_POWER_FUNC_FF_CLAMP_B", + "ME_POWER_FUNC_FF_CLAMP_B1", + "ME_POWER_FUNC_FF_CLAMP_01", + " ME_SIN", + " ME_COS", + /* R500 math ops */ + " ME_LOG_BASE2_IEEE", + " ME_RECIP_IEEE", + " ME_RECIP_SQRT_IEEE", + " ME_PRED_SET_EQ", + " ME_PRED_SET_GT", + " ME_PRED_SET_GTE", + " ME_PRED_SET_NEQ", + " ME_PRED_SET_CLR", + " ME_PRED_SET_INV", + " ME_PRED_SET_POP", + " ME_PRED_SET_RESTORE", + " (reserved)", + " (reserved)", + " (reserved)", +}; + void r500_fs_dump(struct r500_fragment_shader* fs); void r300_vs_dump(struct r300_vertex_shader* vs); diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c index 417d5f6307..c73d5a0b44 100644 --- a/src/gallium/drivers/r300/r300_emit.c +++ b/src/gallium/drivers/r300/r300_emit.c @@ -152,34 +152,39 @@ void r500_emit_fragment_shader(struct r300_context* r300, END_CS; } -/* XXX add pitch, stride, clean up */ void r300_emit_fb_state(struct r300_context* r300, struct pipe_framebuffer_state* fb) { - int i; struct r300_texture* tex; + unsigned pixpitch; + int i; CS_LOCALS(r300); - BEGIN_CS((6 * fb->nr_cbufs) + (fb->zsbuf ? 6 : 0) + 4); + BEGIN_CS((8 * fb->nr_cbufs) + (fb->zsbuf ? 8 : 0) + 4); for (i = 0; i < fb->nr_cbufs; i++) { tex = (struct r300_texture*)fb->cbufs[i]->texture; + pixpitch = tex->stride / tex->tex.block.size; + OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1); OUT_CS_RELOC(tex->buffer, 0, 0, RADEON_GEM_DOMAIN_VRAM, 0); + OUT_CS_REG(R300_RB3D_COLORPITCH0 + (4 * i), pixpitch | + r300_translate_colorformat(tex->tex.format)); + OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), r300_translate_out_fmt(fb->cbufs[i]->format)); } if (fb->zsbuf) { tex = (struct r300_texture*)fb->zsbuf->texture; + pixpitch = (tex->stride / tex->tex.block.size); + OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1); OUT_CS_RELOC(tex->buffer, 0, 0, RADEON_GEM_DOMAIN_VRAM, 0); - if (fb->zsbuf->format == PIPE_FORMAT_Z24S8_UNORM) { - OUT_CS_REG(R300_ZB_FORMAT, - R300_DEPTHFORMAT_24BIT_INT_Z_8BIT_STENCIL); - } else { - OUT_CS_REG(R300_ZB_FORMAT, 0x0); - } + + OUT_CS_REG(R300_ZB_FORMAT, r300_translate_zsformat(tex->tex.format)); + + OUT_CS_REG(R300_ZB_DEPTHPITCH, pixpitch); } OUT_CS_REG(R300_RB3D_DSTCACHE_CTLSTAT, @@ -291,6 +296,30 @@ void r300_emit_texture(struct r300_context* r300, END_CS; } +void r300_emit_vertex_buffer(struct r300_context* r300) +{ + CS_LOCALS(r300); + + debug_printf("r300: Preparing vertex buffer %p for render, " + "vertex size %d\n", r300->vbo, + r300->vertex_info.vinfo.size); + /* Set the pointer to our vertex buffer. The emitted values are this: + * PACKET3 [3D_LOAD_VBPNTR] + * COUNT [1] + * FORMAT [size | stride << 8] + * OFFSET [offset into BO] + * VBPNTR [relocated BO] + */ + BEGIN_CS(7); + OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, 3); + OUT_CS(1); + OUT_CS(r300->vertex_info.vinfo.size | + (r300->vertex_info.vinfo.size << 8)); + OUT_CS(r300->vbo_offset); + OUT_CS_RELOC(r300->vbo, 0, RADEON_GEM_DOMAIN_GTT, 0, 0); + END_CS; +} + void r300_emit_vertex_format_state(struct r300_context* r300) { int i; @@ -416,16 +445,45 @@ void r300_flush_textures(struct r300_context* r300) void r300_emit_dirty_state(struct r300_context* r300) { struct r300_screen* r300screen = r300_screen(r300->context.screen); + struct r300_texture* tex; int i; int dirty_tex = 0; - if (!(r300->dirty_state) && !(r300->dirty_hw)) { + if (!(r300->dirty_state)) { return; } r300_update_derived_state(r300); /* XXX check size */ + /* Color buffers... */ + for (i = 0; i < r300->framebuffer_state.nr_cbufs; i++) { + tex = (struct r300_texture*)r300->framebuffer_state.cbufs[i]->texture; + assert(tex && tex->buffer && "cbuf is marked, but NULL!"); + if (!tex->buffer) return; + r300->winsys->add_buffer(r300->winsys, tex->buffer, + 0, RADEON_GEM_DOMAIN_VRAM); + } + /* ...depth buffer... */ + if (r300->framebuffer_state.zsbuf) { + tex = (struct r300_texture*)r300->framebuffer_state.zsbuf->texture; + assert(tex && tex->buffer && "zsbuf is marked, but NULL!"); + if (!tex->buffer) return; + r300->winsys->add_buffer(r300->winsys, tex->buffer, + 0, RADEON_GEM_DOMAIN_VRAM); + } + /* ...and vertex buffer. */ + if (r300->vbo) { + r300->winsys->add_buffer(r300->winsys, r300->vbo, + RADEON_GEM_DOMAIN_GTT, 0); + } else { + debug_printf("No VBO while emitting dirty state!\n"); + } + + if (r300->winsys->validate(r300->winsys)) { + /* XXX */ + r300->context.flush(&r300->context, 0, NULL); + } if (r300->dirty_state & R300_NEW_BLEND) { r300_emit_blend_state(r300, r300->blend_state); @@ -506,4 +564,9 @@ void r300_emit_dirty_state(struct r300_context* r300) r300_emit_vertex_format_state(r300); r300->dirty_state &= ~R300_NEW_VERTEX_FORMAT; } + + /* Finally, emit the VBO. */ + r300_emit_vertex_buffer(r300); + + r300->dirty_hw++; } diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h index 31dbc7ab85..36e14f69f7 100644 --- a/src/gallium/drivers/r300/r300_emit.h +++ b/src/gallium/drivers/r300/r300_emit.h @@ -62,6 +62,8 @@ void r300_emit_scissor_state(struct r300_context* r300, void r300_emit_texture(struct r300_context* r300, struct r300_texture* tex, unsigned offset); +void r300_emit_vertex_buffer(struct r300_context* r300); + void r300_emit_vertex_format_state(struct r300_context* r300); void r300_emit_vertex_shader(struct r300_context* r300, diff --git a/src/gallium/drivers/r300/r300_flush.c b/src/gallium/drivers/r300/r300_flush.c index 20ca6905ad..89a5f2b20c 100644 --- a/src/gallium/drivers/r300/r300_flush.c +++ b/src/gallium/drivers/r300/r300_flush.c @@ -29,6 +29,8 @@ static void r300_flush(struct pipe_context* pipe, struct r300_context* r300 = r300_context(pipe); CS_LOCALS(r300); + draw_flush(r300->draw); + if (r300->dirty_hw) { FLUSH_CS; r300_emit_invariant_state(r300); diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c index cbd84d7c56..29b66cee7e 100644 --- a/src/gallium/drivers/r300/r300_render.c +++ b/src/gallium/drivers/r300/r300_render.c @@ -180,27 +180,10 @@ static void prepare_render(struct r300_render* render, unsigned count) CS_LOCALS(r300); - /* Make sure that all possible state is emitted. */ - r300_emit_dirty_state(r300); + r300->vbo = render->vbo; + r300->vbo_offset = render->vbo_offset; - debug_printf("r300: Preparing vertex buffer %p for render, " - "vertex size %d, vertex count %d\n", render->vbo, - r300->vertex_info.vinfo.size, count); - /* Set the pointer to our vertex buffer. The emitted values are this: - * PACKET3 [3D_LOAD_VBPNTR] - * COUNT [1] - * FORMAT [size | stride << 8] - * OFFSET [0] - * VBPNTR [relocated BO] - */ - BEGIN_CS(7); - OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, 3); - OUT_CS(1); - OUT_CS(r300->vertex_info.vinfo.size | - (r300->vertex_info.vinfo.size << 8)); - OUT_CS(render->vbo_offset); - OUT_CS_RELOC(render->vbo, 0, RADEON_GEM_DOMAIN_GTT, 0, 0); - END_CS; + r300_emit_dirty_state(r300); } static void r300_render_draw_arrays(struct vbuf_render* render, diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c index c9507ae193..0143e228c4 100644 --- a/src/gallium/drivers/r300/r300_state.c +++ b/src/gallium/drivers/r300/r300_state.c @@ -421,6 +421,7 @@ static void r300_bind_rs_state(struct pipe_context* pipe, void* state) struct r300_context* r300 = r300_context(pipe); struct r300_rs_state* rs = (struct r300_rs_state*)state; + draw_flush(r300->draw); draw_set_rasterizer_state(r300->draw, &rs->rs); r300->rs_state = rs; @@ -528,7 +529,6 @@ static void r300_set_scissor_state(struct pipe_context* pipe, const struct pipe_scissor_state* state) { struct r300_context* r300 = r300_context(pipe); - draw_flush(r300->draw); if (r300_screen(r300->context.screen)->caps->is_r500) { r300->scissor_state->scissor_top_left = @@ -555,19 +555,24 @@ static void r300_set_viewport_state(struct pipe_context* pipe, { struct r300_context* r300 = r300_context(pipe); + draw_flush(r300->draw); + if (r300_screen(r300->context.screen)->caps->has_tcl) { /* Do the transform in HW. */ r300->viewport_state->vte_control = R300_VTX_W0_FMT; if (state->scale[0] != 1.0f) { + assert(state->scale[0] != 0.0f); r300->viewport_state->xscale = state->scale[0]; r300->viewport_state->vte_control |= R300_VPORT_X_SCALE_ENA; } if (state->scale[1] != 1.0f) { + assert(state->scale[1] != 0.0f); r300->viewport_state->yscale = state->scale[1]; r300->viewport_state->vte_control |= R300_VPORT_Y_SCALE_ENA; } if (state->scale[2] != 1.0f) { + assert(state->scale[2] != 0.0f); r300->viewport_state->zscale = state->scale[2]; r300->viewport_state->vte_control |= R300_VPORT_Z_SCALE_ENA; } @@ -642,6 +647,8 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader) { struct r300_context* r300 = r300_context(pipe); + draw_flush(r300->draw); + if (r300_screen(pipe->screen)->caps->has_tcl) { struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader; diff --git a/src/gallium/drivers/r300/r300_state_tcl.c b/src/gallium/drivers/r300/r300_state_tcl.c index bb96e2ad67..ed9164db49 100644 --- a/src/gallium/drivers/r300/r300_state_tcl.c +++ b/src/gallium/drivers/r300/r300_state_tcl.c @@ -34,14 +34,20 @@ static void r300_vs_declare(struct r300_vs_asm* assembler, assembler->tab[decl->DeclarationRange.First] = 0; break; case TGSI_SEMANTIC_COLOR: - assembler->tab[decl->DeclarationRange.First] = 2; + assembler->tab[decl->DeclarationRange.First] = + (assembler->point_size ? 1 : 0) + + assembler->out_colors; break; + case TGSI_SEMANTIC_FOG: case TGSI_SEMANTIC_GENERIC: /* XXX multiple? */ - assembler->tab[decl->DeclarationRange.First] = 6; + assembler->tab[decl->DeclarationRange.First] = + (assembler->point_size ? 1 : 0) + + assembler->out_colors + + assembler->out_texcoords; break; case TGSI_SEMANTIC_PSIZE: - assembler->tab[decl->DeclarationRange.First] = 15; + assembler->tab[decl->DeclarationRange.First] = 1; break; default: debug_printf("r300: vs: Bad semantic declaration %d\n", @@ -65,16 +71,13 @@ static INLINE unsigned r300_vs_src_type(struct r300_vs_asm* assembler, { switch (src->File) { case TGSI_FILE_NULL: - /* Probably a zero or one swizzle */ - return R300_PVS_SRC_REG_INPUT; - break; case TGSI_FILE_INPUT: + /* Probably a zero or one swizzle */ return R300_PVS_SRC_REG_INPUT; - break; case TGSI_FILE_TEMPORARY: return R300_PVS_SRC_REG_TEMPORARY; - break; case TGSI_FILE_CONSTANT: + case TGSI_FILE_IMMEDIATE: return R300_PVS_SRC_REG_CONSTANT; default: debug_printf("r300: vs: Unimplemented src type %d\n", src->File); @@ -83,16 +86,32 @@ static INLINE unsigned r300_vs_src_type(struct r300_vs_asm* assembler, return 0; } +static INLINE unsigned r300_vs_src(struct r300_vs_asm* assembler, + struct tgsi_src_register* src) +{ + switch (src->File) { + case TGSI_FILE_NULL: + case TGSI_FILE_INPUT: + case TGSI_FILE_TEMPORARY: + case TGSI_FILE_CONSTANT: + return src->Index; + case TGSI_FILE_IMMEDIATE: + return src->Index + assembler->imm_offset; + default: + debug_printf("r300: vs: Unimplemented src type %d\n", src->File); + break; + } + return 0; +} + static INLINE unsigned r300_vs_dst_type(struct r300_vs_asm* assembler, struct tgsi_dst_register* dst) { switch (dst->File) { case TGSI_FILE_TEMPORARY: return R300_PVS_DST_REG_TEMPORARY; - break; case TGSI_FILE_OUTPUT: return R300_PVS_DST_REG_OUT; - break; default: debug_printf("r300: vs: Unimplemented dst type %d\n", dst->File); break; @@ -106,10 +125,8 @@ static INLINE unsigned r300_vs_dst(struct r300_vs_asm* assembler, switch (dst->File) { case TGSI_FILE_TEMPORARY: return dst->Index; - break; case TGSI_FILE_OUTPUT: return assembler->tab[dst->Index]; - break; default: debug_printf("r300: vs: Unimplemented dst %d\n", dst->File); break; @@ -129,6 +146,12 @@ static uint32_t r300_vs_op(unsigned op) case TGSI_OPCODE_MOV: case TGSI_OPCODE_SWZ: return R300_VE_ADD; + case TGSI_OPCODE_MAX: + return R300_VE_MAXIMUM; + case TGSI_OPCODE_SLT: + return R300_VE_SET_LESS_THAN; + case TGSI_OPCODE_RSQ: + return R300_PVS_DST_MATH_INST | R300_ME_RECIP_DX; case TGSI_OPCODE_MAD: return R300_PVS_DST_MACRO_INST | R300_PVS_MACRO_OP_2CLK_MADD; default: @@ -152,39 +175,62 @@ static uint32_t r300_vs_swiz(struct tgsi_full_src_register* reg) } } +/* XXX icky icky icky icky */ +static uint32_t r300_vs_scalar_swiz(struct tgsi_full_src_register* reg) +{ + if (reg->SrcRegister.Extended) { + return reg->SrcRegisterExtSwz.ExtSwizzleX | + (reg->SrcRegisterExtSwz.ExtSwizzleX << 3) | + (reg->SrcRegisterExtSwz.ExtSwizzleX << 6) | + (reg->SrcRegisterExtSwz.ExtSwizzleX << 9); + } else { + return reg->SrcRegister.SwizzleX | + (reg->SrcRegister.SwizzleX << 3) | + (reg->SrcRegister.SwizzleX << 6) | + (reg->SrcRegister.SwizzleX << 9); + } +} + +/* XXX scalar stupidity */ static void r300_vs_emit_inst(struct r300_vertex_shader* vs, struct r300_vs_asm* assembler, struct tgsi_full_src_register* src, struct tgsi_full_dst_register* dst, unsigned op, - unsigned count) + unsigned count, + boolean is_scalar) { int i = vs->instruction_count; vs->instructions[i].inst0 = R300_PVS_DST_OPCODE(r300_vs_op(op)) | R300_PVS_DST_REG_TYPE(r300_vs_dst_type(assembler, &dst->DstRegister)) | R300_PVS_DST_OFFSET(r300_vs_dst(assembler, &dst->DstRegister)) | - R300_PVS_DST_WE_XYZW; + R300_PVS_DST_WE(dst->DstRegister.WriteMask); switch (count) { case 3: vs->instructions[i].inst3 = R300_PVS_SRC_REG_TYPE(r300_vs_src_type(assembler, &src[2].SrcRegister)) | - R300_PVS_SRC_OFFSET(src[2].SrcRegister.Index) | + R300_PVS_SRC_OFFSET(r300_vs_src(assembler, + &src[2].SrcRegister)) | R300_PVS_SRC_SWIZZLE(r300_vs_swiz(&src[2])); /* Fall through */ case 2: vs->instructions[i].inst2 = R300_PVS_SRC_REG_TYPE(r300_vs_src_type(assembler, &src[1].SrcRegister)) | - R300_PVS_SRC_OFFSET(src[1].SrcRegister.Index) | + R300_PVS_SRC_OFFSET(r300_vs_src(assembler, + &src[1].SrcRegister)) | R300_PVS_SRC_SWIZZLE(r300_vs_swiz(&src[1])); /* Fall through */ case 1: vs->instructions[i].inst1 = R300_PVS_SRC_REG_TYPE(r300_vs_src_type(assembler, &src[0].SrcRegister)) | - R300_PVS_SRC_OFFSET(src[0].SrcRegister.Index) | - R300_PVS_SRC_SWIZZLE(r300_vs_swiz(&src[0])); + R300_PVS_SRC_OFFSET(r300_vs_src(assembler, + &src[0].SrcRegister)) | + /* XXX the icky, it burns */ + R300_PVS_SRC_SWIZZLE(is_scalar ? r300_vs_scalar_swiz(&src[0]) + : r300_vs_swiz(&src[0])); break; } vs->instruction_count++; @@ -195,11 +241,18 @@ static void r300_vs_instruction(struct r300_vertex_shader* vs, struct tgsi_full_instruction* inst) { switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_RSQ: + r300_vs_emit_inst(vs, assembler, inst->FullSrcRegisters, + &inst->FullDstRegisters[0], inst->Instruction.Opcode, + 1, TRUE); + break; case TGSI_OPCODE_ADD: case TGSI_OPCODE_MUL: + case TGSI_OPCODE_MAX: + case TGSI_OPCODE_SLT: r300_vs_emit_inst(vs, assembler, inst->FullSrcRegisters, &inst->FullDstRegisters[0], inst->Instruction.Opcode, - 2); + 2, FALSE); break; case TGSI_OPCODE_DP3: /* Set alpha swizzle to zero for src0 and src1 */ @@ -229,19 +282,19 @@ static void r300_vs_instruction(struct r300_vertex_shader* vs, case TGSI_OPCODE_DP4: r300_vs_emit_inst(vs, assembler, inst->FullSrcRegisters, &inst->FullDstRegisters[0], inst->Instruction.Opcode, - 2); + 2, FALSE); break; case TGSI_OPCODE_MOV: case TGSI_OPCODE_SWZ: inst->FullSrcRegisters[1] = r300_constant_zero; r300_vs_emit_inst(vs, assembler, inst->FullSrcRegisters, &inst->FullDstRegisters[0], inst->Instruction.Opcode, - 2); + 2, FALSE); break; case TGSI_OPCODE_MAD: r300_vs_emit_inst(vs, assembler, inst->FullSrcRegisters, &inst->FullDstRegisters[0], inst->Instruction.Opcode, - 3); + 3, FALSE); break; case TGSI_OPCODE_END: break; @@ -252,6 +305,28 @@ static void r300_vs_instruction(struct r300_vertex_shader* vs, } } +static void r300_vs_init(struct r300_vertex_shader* vs, + struct r300_vs_asm* assembler) +{ + struct tgsi_shader_info* info = &vs->info; + int i; + + for (i = 0; i < info->num_outputs; i++) { + switch (info->output_semantic_name[i]) { + case TGSI_SEMANTIC_PSIZE: + assembler->point_size = TRUE; + break; + case TGSI_SEMANTIC_COLOR: + assembler->out_colors++; + break; + case TGSI_SEMANTIC_FOG: + case TGSI_SEMANTIC_GENERIC: + assembler->out_texcoords++; + break; + } + } +} + void r300_translate_vertex_shader(struct r300_context* r300, struct r300_vertex_shader* vs) { @@ -264,6 +339,10 @@ void r300_translate_vertex_shader(struct r300_context* r300, if (assembler == NULL) { return; } + + /* Init assembler. */ + r300_vs_init(vs, assembler); + /* Setup starting offset for immediates. */ assembler->imm_offset = consts->user_count; diff --git a/src/gallium/drivers/r300/r300_state_tcl.h b/src/gallium/drivers/r300/r300_state_tcl.h index de944028ba..d5d425e9d6 100644 --- a/src/gallium/drivers/r300/r300_state_tcl.h +++ b/src/gallium/drivers/r300/r300_state_tcl.h @@ -35,6 +35,10 @@ # define R300_VE_DOT_PRODUCT 1 # define R300_VE_MULTIPLY 2 # define R300_VE_ADD 3 +# define R300_VE_MAXIMUM 7 +# define R300_VE_SET_LESS_THAN 10 +#define R300_PVS_DST_MATH_INST (1 << 6) +# define R300_ME_RECIP_DX 6 #define R300_PVS_DST_MACRO_INST (1 << 7) # define R300_PVS_MACRO_OP_2CLK_MADD 0 #define R300_PVS_DST_REG_TYPE(x) ((x) << 8) @@ -99,7 +103,13 @@ struct r300_vs_asm { unsigned imm_offset; /* Number of immediate constants. */ unsigned imm_count; - /* Offsets into vertex output memory. */ + /* Number of colors to write. */ + unsigned out_colors; + /* Number of texcoords to write. */ + unsigned out_texcoords; + /* Whether to emit point size. */ + boolean point_size; + /* Tab of declared outputs to OVM outputs. */ unsigned tab[16]; }; @@ -115,7 +125,7 @@ static struct r300_vertex_shader r300_passthrough_vertex_shader = { .instructions[0].inst3 = 0x0, .instructions[1].inst0 = R300_PVS_DST_OPCODE(R300_VE_ADD) | R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) | - R300_PVS_DST_OFFSET(2) | R300_PVS_DST_WE_XYZW, + R300_PVS_DST_OFFSET(1) | R300_PVS_DST_WE_XYZW, .instructions[1].inst1 = R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) | R300_PVS_SRC_OFFSET(1) | R300_PVS_SRC_SWIZZLE_XYZW, .instructions[1].inst2 = R300_PVS_SRC_SWIZZLE_ZERO, @@ -134,7 +144,7 @@ static struct r300_vertex_shader r300_texture_vertex_shader = { .instructions[0].inst3 = 0x0, .instructions[1].inst0 = R300_PVS_DST_OPCODE(R300_VE_ADD) | R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) | - R300_PVS_DST_OFFSET(6) | R300_PVS_DST_WE_XYZW, + R300_PVS_DST_OFFSET(1) | R300_PVS_DST_WE_XYZW, .instructions[1].inst1 = R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) | R300_PVS_SRC_OFFSET(1) | R300_PVS_SRC_SWIZZLE_XYZW, .instructions[1].inst2 = R300_PVS_SRC_SWIZZLE_ZERO, diff --git a/src/gallium/drivers/r300/r300_surface.c b/src/gallium/drivers/r300/r300_surface.c index 79bed03253..4dd5b8af99 100644 --- a/src/gallium/drivers/r300/r300_surface.c +++ b/src/gallium/drivers/r300/r300_surface.c @@ -34,6 +34,13 @@ static void r300_surface_setup(struct pipe_context* pipe, unsigned pixpitch = tex->stride / tex->tex.block.size; CS_LOCALS(r300); + /* Make sure our target BO is okay. */ + r300->winsys->add_buffer(r300->winsys, tex->buffer, + 0, RADEON_GEM_DOMAIN_VRAM); + if (r300->winsys->validate(r300->winsys)) { + r300->context.flush(&r300->context, 0, NULL); + } + r300_emit_blend_state(r300, &blend_clear_state); r300_emit_blend_color_state(r300, &blend_color_clear_state); r300_emit_dsa_state(r300, &dsa_clear_state); diff --git a/src/gallium/drivers/r300/r300_texture.c b/src/gallium/drivers/r300/r300_texture.c index fe91f4e184..5ea9f56247 100644 --- a/src/gallium/drivers/r300/r300_texture.c +++ b/src/gallium/drivers/r300/r300_texture.c @@ -86,8 +86,6 @@ static struct pipe_texture* r300_texture_create(struct pipe_screen* screen, const struct pipe_texture* template) { - /* XXX struct r300_screen* r300screen = r300_screen(screen); */ - struct r300_texture* tex = CALLOC_STRUCT(r300_texture); if (!tex) { diff --git a/src/gallium/drivers/r300/r300_winsys.h b/src/gallium/drivers/r300/r300_winsys.h index baa95282c3..a833bb0399 100644 --- a/src/gallium/drivers/r300/r300_winsys.h +++ b/src/gallium/drivers/r300/r300_winsys.h @@ -35,8 +35,6 @@ extern "C" { #include "pipe/p_state.h" #include "pipe/internal/p_winsys_screen.h" -struct radeon_cs; - struct r300_winsys { /* Parent class */ struct pipe_winsys base; @@ -50,39 +48,50 @@ struct r300_winsys { /* GB pipe count */ uint32_t gb_pipes; - /* CS object. This is very much like Intel's batchbuffer. - * Fill it full of dwords and relocs and then submit. - * Repeat as needed. */ - struct radeon_cs* cs; + /* GART size. */ + uint32_t gart_size; + + /* VRAM size. */ + uint32_t vram_size; + + /* Add a pipe_buffer to the list of buffer objects to validate. */ + void (*add_buffer)(struct r300_winsys* winsys, + struct pipe_buffer* pbuffer, + uint32_t rd, + uint32_t wd); + + /* Revalidate all currently setup pipe_buffers. + * Returns TRUE if a flush is required. */ + boolean (*validate)(struct r300_winsys* winsys); /* Check to see if there's room for commands. */ - boolean (*check_cs)(struct radeon_cs* cs, int size); + boolean (*check_cs)(struct r300_winsys* winsys, int size); /* Start a command emit. */ - void (*begin_cs)(struct radeon_cs* cs, - int size, - const char* file, - const char* function, - int line); + void (*begin_cs)(struct r300_winsys* winsys, + int size, + const char* file, + const char* function, + int line); /* Write a dword to the command buffer. */ - void (*write_cs_dword)(struct radeon_cs* cs, uint32_t dword); + void (*write_cs_dword)(struct r300_winsys* winsys, uint32_t dword); /* Write a relocated dword to the command buffer. */ - void (*write_cs_reloc)(struct radeon_cs* cs, - struct pipe_buffer* bo, - uint32_t rd, - uint32_t wd, - uint32_t flags); + void (*write_cs_reloc)(struct r300_winsys* winsys, + struct pipe_buffer* bo, + uint32_t rd, + uint32_t wd, + uint32_t flags); /* Finish a command emit. */ - void (*end_cs)(struct radeon_cs* cs, - const char* file, - const char* function, - int line); + void (*end_cs)(struct r300_winsys* winsys, + const char* file, + const char* function, + int line); /* Flush the CS. */ - void (*flush_cs)(struct radeon_cs* cs); + void (*flush_cs)(struct r300_winsys* winsys); }; struct pipe_context* r300_create_context(struct pipe_screen* screen, |