intel-hard-graphics/src/gallium/drivers/r300/r300_fs.c

/*
 * Copyright 2008 Corbin Simpson <[email protected]>
 *                Joakim Sindholt <[email protected]>
 * Copyright 2009 Marek Olšák <[email protected]>
 * SPDX-License-Identifier: MIT
 */

#include "util/format/u_format.h"
#include "util/u_math.h"
#include "util/u_memory.h"

#include "r300_cb.h"
#include "r300_context.h"
#include "r300_emit.h"
#include "r300_screen.h"
#include "r300_fs.h"
#include "r300_reg.h"
#include "r300_texture.h"

#include "compiler/radeon_compiler.h"
#include "compiler/nir_to_rc.h"
#include "nir.h"
#include "compiler/nir/nir_builder.h"

static void allocate_hardware_inputs(
    struct r300_fragment_program_compiler * c,
    void (*allocate)(void * data, unsigned input, unsigned hwreg),
    void * mydata)
{
    struct r300_shader_semantics* inputs =
        (struct r300_shader_semantics*)c->UserData;
    int i, reg = 0;

    /* Allocate input registers. */
    for (i = 0; i < ATTR_COLOR_COUNT; i++) {
        if (inputs->color[i] != ATTR_UNUSED) {
            allocate(mydata, inputs->color[i], reg++);
        }
    }
    if (inputs->face != ATTR_UNUSED) {
        allocate(mydata, inputs->face, reg++);
    }
    for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
        if (inputs->generic[i] != ATTR_UNUSED) {
            allocate(mydata, inputs->generic[i], reg++);
        }
    }
    if (inputs->fog != ATTR_UNUSED) {
        allocate(mydata, inputs->fog, reg++);
    }
    if (inputs->wpos != ATTR_UNUSED) {
        allocate(mydata, inputs->wpos, reg++);
    }
}

void r300_fragment_program_get_external_state(
    struct r300_context* r300,
    struct r300_fragment_program_external_state* state)
{
    struct r300_textures_state *texstate = r300->textures_state.state;
    unsigned i;

    state->alpha_to_one = r300->alpha_to_one && r300->msaa_enable;
    state->sampler_state_count = texstate->sampler_state_count;

    for (i = 0; i < texstate->sampler_state_count; i++) {
        struct r300_sampler_state *s = texstate->sampler_states[i];
        struct r300_sampler_view *v = texstate->sampler_views[i];
        struct r300_resource *t;

        if (!s || !v) {
            continue;
        }

        t = r300_resource(v->base.texture);

        if (s->state.compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
            state->unit[i].compare_mode_enabled = 1;

            /* Fortunately, no need to translate this. */
            state->unit[i].texture_compare_func = s->state.compare_func;
        }

        /* Pass texture swizzling to the compiler, some lowering passes need it. */
        if (state->unit[i].compare_mode_enabled) {
            state->unit[i].texture_swizzle =
                RC_MAKE_SWIZZLE(v->swizzle[0], v->swizzle[1],
                                v->swizzle[2], v->swizzle[3]);
        }

        /* XXX this should probably take into account STR, not just S. */
        if (t->tex.is_npot) {
            switch (s->state.wrap_s) {
            case PIPE_TEX_WRAP_REPEAT:
                state->unit[i].wrap_mode = RC_WRAP_REPEAT;
                break;

            case PIPE_TEX_WRAP_MIRROR_REPEAT:
                state->unit[i].wrap_mode = RC_WRAP_MIRRORED_REPEAT;
                break;

            case PIPE_TEX_WRAP_MIRROR_CLAMP:
            case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
            case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
                state->unit[i].wrap_mode = RC_WRAP_MIRRORED_CLAMP;
                break;

            default:
                state->unit[i].wrap_mode = RC_WRAP_NONE;
            }

            if (t->b.target == PIPE_TEXTURE_3D)
                state->unit[i].clamp_and_scale_before_fetch = true;
        }
    }
}

static void r300_translate_fragment_shader(
    struct r300_context* r300,
    struct r300_fragment_shader_code* shader,
    struct pipe_shader_state state);

static void r300_dummy_fragment_shader(
    struct r300_context* r300,
    struct r300_fragment_shader_code* shader)
{
    struct pipe_shader_state state = {};
    const nir_shader_compiler_options *options =
        r300->screen->screen.nir_options[MESA_SHADER_FRAGMENT];

    /* Make a simple fragment shader which outputs (0, 0, 0, 1). */
    nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_FRAGMENT,
                                                   options, "r300 dummy FS");
    nir_variable *out = nir_variable_create(b.shader, nir_var_shader_out,
                                            glsl_vec4_type(), "out_color");
    out->data.location = FRAG_RESULT_COLOR;
    nir_store_var(&b, out, nir_imm_vec4(&b, 0, 0, 0, 1), 0xf);

    shader->dummy = true;
    state.type = PIPE_SHADER_IR_NIR;
    state.ir.nir = b.shader;
    r300_translate_fragment_shader(r300, shader, state);
    ralloc_free(state.ir.nir);
}

static void r300_emit_fs_code_to_buffer(
    struct r300_context *r300,
    struct r300_fragment_shader_code *shader)
{
    struct rX00_fragment_program_code *generic_code = &shader->code;
    unsigned imm_count = shader->immediates_count;
    unsigned imm_first = shader->externals_count;
    unsigned imm_end = generic_code->constants.Count;
    struct rc_constant *constants = generic_code->constants.Constants;
    unsigned i;
    CB_LOCALS;

    if (r300->screen->caps.is_r500) {
        struct r500_fragment_program_code *code = &generic_code->code.r500;

        shader->cb_code_size = 19 +
                               ((code->inst_end + 1) * 6) +
                               imm_count * 7 +
                               code->int_constant_count * 2;

        NEW_CB(shader->cb_code, shader->cb_code_size);
        if (r300->screen->options.ieeemath)
            OUT_CB_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO_DEFAULT);
        else
            OUT_CB_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO_LEGACY);
        OUT_CB_REG(R500_US_PIXSIZE, code->max_temp_idx);
        OUT_CB_REG(R500_US_FC_CTRL, code->us_fc_ctrl);
        for(i = 0; i < code->int_constant_count; i++){
                OUT_CB_REG(R500_US_FC_INT_CONST_0 + (i * 4),
                                                code->int_constants[i]);
        }
        OUT_CB_REG(R500_US_CODE_RANGE,
                   R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(code->inst_end));
        OUT_CB_REG(R500_US_CODE_OFFSET, 0);
        OUT_CB_REG(R500_US_CODE_ADDR,
                   R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(code->inst_end));

        OUT_CB_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_INSTR);
        OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, (code->inst_end + 1) * 6);
        for (i = 0; i <= code->inst_end; i++) {
            OUT_CB(code->inst[i].inst0);
            OUT_CB(code->inst[i].inst1);
            OUT_CB(code->inst[i].inst2);
            OUT_CB(code->inst[i].inst3);
            OUT_CB(code->inst[i].inst4);
            OUT_CB(code->inst[i].inst5);
        }

        /* Emit immediates. */
        if (imm_count) {
            for(i = imm_first; i < imm_end; ++i) {
                if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
                    const float *data = constants[i].u.Immediate;

                    OUT_CB_REG(R500_GA_US_VECTOR_INDEX,
                               R500_GA_US_VECTOR_INDEX_TYPE_CONST |
                               (i & R500_GA_US_VECTOR_INDEX_MASK));
                    OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
                    OUT_CB_TABLE(data, 4);
                }
            }
        }
    } else { /* r300 */
        struct r300_fragment_program_code *code = &generic_code->code.r300;
        unsigned int alu_length = code->alu.length;
        unsigned int alu_iterations = ((alu_length - 1) / 64) + 1;
        unsigned int tex_length = code->tex.length;
        unsigned int tex_iterations =
            tex_length > 0 ? ((tex_length - 1) / 32) + 1 : 0;
        unsigned int iterations =
            alu_iterations > tex_iterations ? alu_iterations : tex_iterations;
        unsigned int bank = 0;

        shader->cb_code_size = 15 +
            /* R400_US_CODE_BANK */
            (r300->screen->caps.is_r400 ? 2 * (iterations + 1): 0) +
            /* R400_US_CODE_EXT */
            (r300->screen->caps.is_r400 ? 2 : 0) +
            /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0, R400_US_ALU_EXT_ADDR_0 */
            (code->r390_mode ? (5 * alu_iterations) : 4) +
            /* R400_US_ALU_EXT_ADDR_[0-63] */
            (code->r390_mode ? (code->alu.length) : 0) +
            /* R300_US_ALU_{RGB,ALPHA}_{INST,ADDR}_0 */
            code->alu.length * 4 +
            /* R300_US_TEX_INST_0, R300_US_TEX_INST_[0-31] */
            (code->tex.length > 0 ? code->tex.length + tex_iterations : 0) +
            imm_count * 5;

        NEW_CB(shader->cb_code, shader->cb_code_size);

        OUT_CB_REG(R300_US_CONFIG, code->config);
        OUT_CB_REG(R300_US_PIXSIZE, code->pixsize);
        OUT_CB_REG(R300_US_CODE_OFFSET, code->code_offset);

        if (code->r390_mode) {
            OUT_CB_REG(R400_US_CODE_EXT, code->r400_code_offset_ext);
        } else if (r300->screen->caps.is_r400) {
            /* This register appears to affect shaders even if r390_mode is
             * disabled, so it needs to be set to 0 for shaders that
             * don't use r390_mode. */
            OUT_CB_REG(R400_US_CODE_EXT, 0);
        }

        OUT_CB_REG_SEQ(R300_US_CODE_ADDR_0, 4);
        OUT_CB_TABLE(code->code_addr, 4);

        do {
            unsigned int bank_alu_length = (alu_length < 64 ? alu_length : 64);
            unsigned int bank_alu_offset = bank * 64;
            unsigned int bank_tex_length = (tex_length < 32 ? tex_length : 32);
            unsigned int bank_tex_offset = bank * 32;

            if (r300->screen->caps.is_r400) {
                OUT_CB_REG(R400_US_CODE_BANK, code->r390_mode ?
                                (bank << R400_BANK_SHIFT) | R400_R390_MODE_ENABLE : 0);//2
            }

            if (bank_alu_length > 0) {
                OUT_CB_REG_SEQ(R300_US_ALU_RGB_INST_0, bank_alu_length);
                for (i = 0; i < bank_alu_length; i++)
                    OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_inst);

                OUT_CB_REG_SEQ(R300_US_ALU_RGB_ADDR_0, bank_alu_length);
                for (i = 0; i < bank_alu_length; i++)
                    OUT_CB(code->alu.inst[i + bank_alu_offset].rgb_addr);

                OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_INST_0, bank_alu_length);
                for (i = 0; i < bank_alu_length; i++)
                    OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_inst);

                OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_ADDR_0, bank_alu_length);
                for (i = 0; i < bank_alu_length; i++)
                    OUT_CB(code->alu.inst[i + bank_alu_offset].alpha_addr);

                if (code->r390_mode) {
                    OUT_CB_REG_SEQ(R400_US_ALU_EXT_ADDR_0, bank_alu_length);
                    for (i = 0; i < bank_alu_length; i++)
                        OUT_CB(code->alu.inst[i + bank_alu_offset].r400_ext_addr);
                }
            }

            if (bank_tex_length > 0) {
                OUT_CB_REG_SEQ(R300_US_TEX_INST_0, bank_tex_length);
                OUT_CB_TABLE(code->tex.inst + bank_tex_offset, bank_tex_length);
            }

            alu_length -= bank_alu_length;
            tex_length -= bank_tex_length;
            bank++;
        } while(code->r390_mode && (alu_length > 0 || tex_length > 0));

        /* R400_US_CODE_BANK needs to be reset to 0, otherwise some shaders
         * will be rendered incorrectly. */
        if (r300->screen->caps.is_r400) {
            OUT_CB_REG(R400_US_CODE_BANK,
                code->r390_mode ? R400_R390_MODE_ENABLE : 0);
        }

        /* Emit immediates. */
        if (imm_count) {
            for(i = imm_first; i < imm_end; ++i) {
                if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
                    const float *data = constants[i].u.Immediate;

                    OUT_CB_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
                    OUT_CB(pack_float24(data[0]));
                    OUT_CB(pack_float24(data[1]));
                    OUT_CB(pack_float24(data[2]));
                    OUT_CB(pack_float24(data[3]));
                }
            }
        }
    }

    OUT_CB_REG(R300_FG_DEPTH_SRC, shader->fg_depth_src);
    OUT_CB_REG(R300_US_W_FMT, shader->us_out_w);
    END_CB;
}

static void r300_translate_fragment_shader(
    struct r300_context* r300,
    struct r300_fragment_shader_code* shader,
    struct pipe_shader_state state)
{
    struct r300_fragment_program_compiler compiler;
    int wpos, face;
    unsigned i;
    union r300_shader_code code;
    code.f = shader;

    r300_shader_semantics_reset(&shader->inputs);

    /* gl_FragColor (vs. gl_FragData[0]) makes the FS write the same value
     * to all bound color buffers. */
    shader->write_all = false;
    nir_foreach_shader_out_variable(var, state.ir.nir) {
        if (var->data.location == FRAG_RESULT_COLOR) {
            shader->write_all = true;
            break;
        }
    }

    /* Setup the compiler. */
    memset(&compiler, 0, sizeof(compiler));
    rc_init(&compiler.Base, &r300->fs_regalloc_state);
    DBG_ON(r300, DBG_FP) ? compiler.Base.Debug |= RC_DBG_LOG : 0;

    compiler.code = &shader->code;
    compiler.state = shader->compare_state;
    if (!shader->dummy)
        compiler.Base.debug = &r300->context.debug;
    compiler.Base.is_r500 = r300->screen->caps.is_r500;
    compiler.Base.is_r400 = r300->screen->caps.is_r400;
    compiler.Base.disable_optimizations = DBG_ON(r300, DBG_NO_OPT);
    compiler.Base.has_half_swizzles = true;
    compiler.Base.has_presub = true;
    compiler.Base.has_omod = true;
    compiler.Base.max_temp_regs =
        compiler.Base.is_r500 ? 128 : (compiler.Base.is_r400 ? 64 : 32);
    compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;
    compiler.Base.max_alu_insts =
        (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 64;
    compiler.Base.max_tex_insts =
        (compiler.Base.is_r500 || compiler.Base.is_r400) ? 512 : 32;
    compiler.AllocateHwInputs = &allocate_hardware_inputs;
    compiler.UserData = &shader->inputs;

    nir_shader *clone = nir_shader_clone(NULL, state.ir.nir);
    nir_to_rc(clone, (struct pipe_screen *)r300->screen, shader->compare_state,
              code, &compiler.Base);

    if (compiler.Base.Error) {
        shader->error = strdup(compiler.Base.ErrorMsg ? compiler.Base.ErrorMsg
                                                      : "Cannot translate shader from NIR.");
        rc_destroy(&compiler.Base);
        r300_dummy_fragment_shader(r300, shader);
        return;
    }

    wpos = shader->inputs.wpos;
    face = shader->inputs.face;

    if (!r300->screen->caps.is_r500 ||
        compiler.Base.Program.Constants.Count > 200) {
        compiler.Base.remove_unused_constants = true;
    }

    /**
     * Transform the program to support WPOS.
     *
     * Introduce a small fragment at the start of the program that will be
     * the only code that directly reads the WPOS input.
     * All other code pieces that reference that input will be rewritten
     * to read from a newly allocated temporary. */
    if (wpos != ATTR_UNUSED) {
        /* Moving the input to some other reg is not really necessary. */
        rc_transform_fragment_wpos(&compiler.Base, wpos, wpos, true);
    }

    if (face != ATTR_UNUSED) {
        rc_transform_fragment_face(&compiler.Base, face);
    }

    /* Invoke the compiler */
    r3xx_compile_fragment_program(&compiler);

    if (compiler.Base.Error) {
        shader->error = strdup(compiler.Base.ErrorMsg);

        if (shader->dummy) {
            fprintf(stderr, "r300 FP: Cannot compile the dummy shader! "
                    "Giving up...\n");
            abort();
        }

        free(compiler.code->constants.Constants);
        free(compiler.code->constants_remap_table);
        rc_destroy(&compiler.Base);
        r300_dummy_fragment_shader(r300, shader);
        return;
    }

    /* Shaders with zero instructions are invalid,
     * use the dummy shader instead. */
    if (shader->code.code.r500.inst_end == -1) {
        rc_destroy(&compiler.Base);
        r300_dummy_fragment_shader(r300, shader);
        return;
    }

    /* Initialize numbers of constants for each type. */
    shader->externals_count = 0;
    for (i = 0;
         i < shader->code.constants.Count &&
         shader->code.constants.Constants[i].Type == RC_CONSTANT_EXTERNAL; i++) {
        shader->externals_count = i+1;
    }
    shader->immediates_count = 0;
    shader->rc_state_count = 0;

    for (i = shader->externals_count; i < shader->code.constants.Count; i++) {
        switch (shader->code.constants.Constants[i].Type) {
            case RC_CONSTANT_IMMEDIATE:
                ++shader->immediates_count;
                break;
            case RC_CONSTANT_STATE:
                ++shader->rc_state_count;
                break;
            default:
                assert(0);
        }
    }

    /* Setup shader depth output. */
    if (shader->code.writes_depth) {
        shader->fg_depth_src = R300_FG_DEPTH_SRC_SHADER;
        shader->us_out_w = R300_W_FMT_W24 | R300_W_SRC_US;
    } else {
        shader->fg_depth_src = R300_FG_DEPTH_SRC_SCAN;
        shader->us_out_w = R300_W_FMT_W0 | R300_W_SRC_US;
    }

    /* And, finally... */
    rc_destroy(&compiler.Base);

    /* Build the command buffer. */
    r300_emit_fs_code_to_buffer(r300, shader);
}

bool r300_pick_fragment_shader(struct r300_context *r300,
                               struct r300_fragment_shader* fs,
                               struct r300_fragment_program_external_state *state)
{
    struct r300_fragment_shader_code* ptr;

    if (!fs->first) {
        /* Build the fragment shader for the first time. */
        fs->first = fs->shader = CALLOC_STRUCT(r300_fragment_shader_code);

        memcpy(&fs->shader->compare_state, state, sizeof(*state));
        r300_translate_fragment_shader(r300, fs->shader, fs->state);
        return true;

    } else {
        /* Check if the currently-bound shader has been compiled
         * with the texture-compare state we need. */
        if (memcmp(&fs->shader->compare_state, state, sizeof(*state)) != 0) {
            /* Search for the right shader. */
            ptr = fs->first;
            while (ptr) {
                if (memcmp(&ptr->compare_state, state, sizeof(*state)) == 0) {
                    if (fs->shader != ptr) {
                        fs->shader = ptr;
                        return true;
                    }
                    /* The currently-bound one is OK. */
                    return false;
                }
                ptr = ptr->next;
            }

            /* Not found, gotta compile a new one. */
            ptr = CALLOC_STRUCT(r300_fragment_shader_code);
            ptr->next = fs->first;
            fs->first = fs->shader = ptr;

            memcpy(&ptr->compare_state, state, sizeof(*state));
            r300_translate_fragment_shader(r300, ptr, fs->state);
            return true;
        }
    }

    return false;
}