X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_fs_visitor.cpp;h=cdaba7f47d9d267bab5bd5f5030747acf3c93521;hb=e4f26acc08a3d852e60a27d0f0da7001944cb607;hp=13a3bf29da81b8d2f90d4e2c1f8036295dbd8658;hpb=8eb6c109994de2827b0a1340a2dc8d933edaf5e0;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp index 13a3bf29da8..cdaba7f47d9 100644 --- a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp +++ b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp @@ -27,8 +27,6 @@ * makes it easier to do backend-specific optimizations than doing so * in the GLSL IR or in the native code. */ -extern "C" { - #include #include "main/macros.h" @@ -41,7 +39,6 @@ extern "C" { #include "brw_context.h" #include "brw_eu.h" #include "brw_wm.h" -} #include "brw_vec4.h" #include "brw_fs.h" #include "main/uniforms.h" @@ -430,21 +427,16 @@ fs_visitor::try_emit_mad(ir_expression *ir) if (ir->type != glsl_type::float_type) return false; - ir_rvalue *nonmul = ir->operands[1]; - ir_expression *mul = ir->operands[0]->as_expression(); + ir_rvalue *nonmul; + ir_expression *mul; + bool mul_negate, mul_abs; - bool mul_negate = false, mul_abs = false; - if (mul && mul->operation == ir_unop_abs) { - mul = mul->operands[0]->as_expression(); - mul_abs = true; - } else if (mul && mul->operation == ir_unop_neg) { - mul = mul->operands[0]->as_expression(); - mul_negate = true; - } + for (int i = 0; i < 2; i++) { + mul_negate = false; + mul_abs = false; - if (!mul || mul->operation != ir_binop_mul) { - nonmul = ir->operands[0]; - mul = ir->operands[1]->as_expression(); + mul = ir->operands[i]->as_expression(); + nonmul = ir->operands[1 - i]; if (mul && mul->operation == ir_unop_abs) { mul = mul->operands[0]->as_expression(); @@ -454,10 +446,13 @@ fs_visitor::try_emit_mad(ir_expression *ir) mul_negate = true; } - if (!mul || mul->operation != ir_binop_mul) - return false; + if (mul && mul->operation == ir_binop_mul) + break; } + if (!mul || mul->operation != ir_binop_mul) + return false; + nonmul->accept(this); fs_reg src0 = this->result; @@ -1252,6 +1247,7 @@ fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r, case GLSL_TYPE_ATOMIC_UINT: break; + case GLSL_TYPE_DOUBLE: case GLSL_TYPE_VOID: case GLSL_TYPE_ERROR: case GLSL_TYPE_INTERFACE: @@ -2029,7 +2025,7 @@ fs_visitor::emit_texture(ir_texture_opcode op, fs_reg shadow_c, fs_reg lod, fs_reg lod2, int grad_components, fs_reg sample_index, - fs_reg offset_value, unsigned offset_components, + fs_reg offset_value, fs_reg mcs, int gather_component, bool is_cube_array, @@ -2195,7 +2191,6 @@ fs_visitor::visit(ir_texture *ir) } fs_reg offset_value; - int offset_components = 0; if (ir->offset) { ir_constant *const_offset = ir->offset->as_constant(); if (const_offset) { @@ -2210,7 +2205,6 @@ fs_visitor::visit(ir_texture *ir) ir->offset->accept(this); offset_value = this->result; } - offset_components = ir->offset->type->vector_elements; } fs_reg lod, lod2, sample_index, mcs; @@ -2267,7 +2261,7 @@ fs_visitor::visit(ir_texture *ir) emit_texture(ir->op, ir->type, coordinate, coord_components, shadow_comparitor, lod, lod2, grad_components, - sample_index, offset_value, offset_components, mcs, + sample_index, offset_value, mcs, gather_component, is_cube_array, is_rect, sampler, sampler_reg, texunit); } @@ -2422,8 +2416,9 @@ fs_visitor::visit(ir_discard *ir) */ fs_inst *cmp; if (ir->condition) { - ir->condition->accept(this); - cmp = emit(CMP(reg_null_f, this->result, fs_reg(0), BRW_CONDITIONAL_Z)); + emit_bool_to_cond_code(ir->condition); + cmp = (fs_inst *) this->instructions.get_tail(); + cmp->conditional_mod = brw_negate_cmod(cmp->conditional_mod); } else { fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW)); @@ -3366,7 +3361,8 @@ fs_visitor::emit_interpolation_setup_gen6() } int -fs_visitor::setup_color_payload(fs_reg *dst, fs_reg color, unsigned components) +fs_visitor::setup_color_payload(fs_reg *dst, fs_reg color, unsigned components, + bool use_2nd_half) { brw_wm_prog_key *key = (brw_wm_prog_key*) this->key; fs_inst *inst; @@ -3384,7 +3380,7 @@ fs_visitor::setup_color_payload(fs_reg *dst, fs_reg color, unsigned components) colors_enabled = (1 << components) - 1; } - if (dispatch_width == 8 || brw->gen >= 6) { + if (dispatch_width == 8 || (brw->gen >= 6 && !do_dual_src)) { /* SIMD8 write looks like: * m + 0: r0 * m + 1: r1 @@ -3415,6 +3411,33 @@ fs_visitor::setup_color_payload(fs_reg *dst, fs_reg color, unsigned components) len++; } return len; + } else if (brw->gen >= 6 && do_dual_src) { + /* SIMD16 dual source blending for gen6+. + * + * From the SNB PRM, volume 4, part 1, page 193: + * + * "The dual source render target messages only have SIMD8 forms due to + * maximum message length limitations. SIMD16 pixel shaders must send two + * of these messages to cover all of the pixels. Each message contains + * two colors (4 channels each) for each pixel in the message payload." + * + * So in SIMD16 dual source blending we will send 2 SIMD8 messages, + * each one will call this function twice (one for each color involved), + * so in each pass we only write 4 registers. Notice that the second + * SIMD8 message needs to read color data from the 2nd half of the color + * registers, so it needs to call this with use_2nd_half = true. + */ + for (unsigned i = 0; i < 4; ++i) { + if (colors_enabled & (1 << i)) { + dst[i] = fs_reg(GRF, alloc.allocate(1), color.type); + inst = emit(MOV(dst[i], half(offset(color, i), + use_2nd_half ? 1 : 0))); + inst->saturate = key->clamp_fragment_color; + if (use_2nd_half) + inst->force_sechalf = true; + } + } + return 4; } else { /* pre-gen6 SIMD16 single source DP write looks like: * m + 0: r0 @@ -3498,7 +3521,8 @@ fs_visitor::emit_alpha_test() fs_inst * fs_visitor::emit_single_fb_write(fs_reg color0, fs_reg color1, - fs_reg src0_alpha, unsigned components) + fs_reg src0_alpha, unsigned components, + bool use_2nd_half) { assert(stage == MESA_SHADER_FRAGMENT); brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data; @@ -3558,7 +3582,8 @@ fs_visitor::emit_single_fb_write(fs_reg color0, fs_reg color1, * alpha out the pipeline to our null renderbuffer to support * alpha-testing, alpha-to-coverage, and so on. */ - length += setup_color_payload(sources + length, this->outputs[0], 0); + length += setup_color_payload(sources + length, this->outputs[0], 0, + false); } else if (color1.file == BAD_FILE) { if (src0_alpha.file != BAD_FILE) { sources[length] = fs_reg(GRF, alloc.allocate(reg_size), @@ -3568,10 +3593,13 @@ fs_visitor::emit_single_fb_write(fs_reg color0, fs_reg color1, length++; } - length += setup_color_payload(sources + length, color0, components); + length += setup_color_payload(sources + length, color0, components, + false); } else { - length += setup_color_payload(sources + length, color0, components); - length += setup_color_payload(sources + length, color1, components); + length += setup_color_payload(sources + length, color0, components, + use_2nd_half); + length += setup_color_payload(sources + length, color1, components, + use_2nd_half); } if (source_depth_to_render_target) { @@ -3640,12 +3668,6 @@ fs_visitor::emit_fb_writes() brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data; brw_wm_prog_key *key = (brw_wm_prog_key*) this->key; - if (do_dual_src) { - no16("GL_ARB_blend_func_extended not yet supported in SIMD16."); - if (dispatch_width == 16) - do_dual_src = false; - } - fs_inst *inst; if (do_dual_src) { if (INTEL_DEBUG & DEBUG_SHADER_TIME) @@ -3656,6 +3678,30 @@ fs_visitor::emit_fb_writes() inst = emit_single_fb_write(this->outputs[0], this->dual_src_output, reg_undef, 4); inst->target = 0; + + /* SIMD16 dual source blending requires to send two SIMD8 dual source + * messages, where each message contains color data for 8 pixels. Color + * data for the first group of pixels is stored in the "lower" half of + * the color registers, so in SIMD16, the previous message did: + * m + 0: r0 + * m + 1: g0 + * m + 2: b0 + * m + 3: a0 + * + * Here goes the second message, which packs color data for the + * remaining 8 pixels. Color data for these pixels is stored in the + * "upper" half of the color registers, so we need to do: + * m + 0: r1 + * m + 1: g1 + * m + 2: b1 + * m + 3: a1 + */ + if (dispatch_width == 16) { + inst = emit_single_fb_write(this->outputs[0], this->dual_src_output, + reg_undef, 4, true); + inst->target = 0; + } + prog_data->dual_src_blend = true; } else if (key->nr_color_regions > 0) { for (int target = 0; target < key->nr_color_regions; target++) {