From 855ebb26d1868d4c1e61a90c1533154b97bd41ff Mon Sep 17 00:00:00 2001 From: Brian Date: Tue, 9 Jan 2007 09:14:32 -0700 Subject: Implement shadow samplers and dFdx(), dFdy() code generation. --- .../shader/slang/library/slang_fragment_builtin.gc | 102 +++++++++++++-------- .../slang/library/slang_fragment_builtin_gc.h | 78 +++++++++------- src/mesa/shader/slang/slang_assemble.c | 2 + src/mesa/shader/slang/slang_codegen.c | 2 + src/mesa/shader/slang/slang_emit.c | 4 + src/mesa/shader/slang/slang_ir.h | 2 + 6 files changed, 120 insertions(+), 70 deletions(-) (limited to 'src/mesa') diff --git a/src/mesa/shader/slang/library/slang_fragment_builtin.gc b/src/mesa/shader/slang/library/slang_fragment_builtin.gc index 1c099a673b3..034672370ae 100644 --- a/src/mesa/shader/slang/library/slang_fragment_builtin.gc +++ b/src/mesa/shader/slang/library/slang_fragment_builtin.gc @@ -120,28 +120,48 @@ vec4 texture3DProj(const sampler3D sampler, const vec4 coord, const float bias) vec4 textureCube(const samplerCube sampler, const vec3 coord, const float bias) { - __asm vec4_texcube __retVal, sampler, coord, bias; + vec4 coord4; + coord4.xyz = coord; + coord4.w = bias; + __asm vec4_texcube __retVal, sampler, coord4; } +// For shadow textures, we use the regular tex instructions since they should +// do the depth comparison step. + vec4 shadow1D(const sampler1DShadow sampler, const vec3 coord, const float bias) { - __asm vec4_shad1d __retVal, sampler, coord, bias; + vec4 coord4; + coord4.xyz = coord; + coord4.w = bias; + __asm vec4_texb1d __retVal, sampler, coord4; } vec4 shadow1DProj(const sampler1DShadow sampler, const vec4 coord, const float bias) { - return shadow1D (sampler, vec3 (coord.s / coord.q, 0.0, coord.p / coord.q), bias); + vec4 pcoord; + pcoord.x = coord.x / coord.w; + pcoord.z = coord.z; + pcoord.w = bias; + __asm vec4_texb1d __retVal, sampler, pcoord; } vec4 shadow2D(const sampler2DShadow sampler, const vec3 coord, const float bias) { - __asm vec4_shad2d __retVal, sampler, coord, bias; + vec4 coord4; + coord4.xyz = coord; + coord4.w = bias; + __asm vec4_texb2d __retVal, sampler, coord4; } vec4 shadow2DProj(const sampler2DShadow sampler, const vec4 coord, const float bias) { - return shadow2D (sampler, vec3 (coord.s / coord.q, coord.t / coord.q, coord.p / coord.q), bias); + vec4 pcoord; + pcoord.xy = coord.xy / coord.w; + pcoord.z = coord.z; + pcoord.w = bias; + __asm vec4_texb2d __retVal, sampler, pcoord; } @@ -152,59 +172,67 @@ vec4 shadow2DProj(const sampler2DShadow sampler, const vec4 coord, const float b // 8.8 Fragment Processing Functions // -float dFdx (float p) { - // XXX: - return 0.001; +float dFdx(const float p) +{ + __asm vec4_ddx __retVal.x, p.xxxx; } -vec2 dFdx (vec2 p) { - // XXX: - return vec2 (0.001); +vec2 dFdx(const vec2 p) +{ + __asm vec4_ddx __retVal.xy, p.xyyy; } -vec3 dFdx (vec3 p) { - // XXX: - return vec3 (0.001); +vec3 dFdx(const vec3 p) +{ + __asm vec4_ddx __retVal.xyz, p.xyzz; } -vec4 dFdx (vec4 p) { - // XXX: - return vec4 (0.001); +vec4 dFdx(const vec4 p) +{ + __asm vec4_ddx __retVal, p; } -float dFdy (float p) { - // XXX: - return 0.001; +float dFdy(const float p) +{ + __asm vec4_ddy __retVal.x, p.xxxx; } -vec2 dFdy (vec2 p) { - // XXX: - return vec2 (0.001); +vec2 dFdy(const vec2 p) +{ + __asm vec4_ddy __retVal.xy, p.xyyy; } -vec3 dFdy (vec3 p) { - // XXX: - return vec3 (0.001); +vec3 dFdy(const vec3 p) +{ + __asm vec4_ddy __retVal.xyz, p.xyzz; } -vec4 dFdy (vec4 p) { - // XXX: - return vec4 (0.001); +vec4 dFdy(const vec4 p) +{ + __asm vec4_ddy __retVal, p; } -float fwidth (float p) { - return abs (dFdx (p)) + abs (dFdy (p)); +float fwidth (const float p) +{ + // XXX hand-write with __asm + return abs(dFdx(p)) + abs(dFdy(p)); } -vec2 fwidth (vec2 p) { - return abs (dFdx (p)) + abs (dFdy (p)); +vec2 fwidth(const vec2 p) +{ + // XXX hand-write with __asm + return abs(dFdx(p)) + abs(dFdy(p)); } -vec3 fwidth (vec3 p) { - return abs (dFdx (p)) + abs (dFdy (p)); +vec3 fwidth(const vec3 p) +{ + // XXX hand-write with __asm + return abs(dFdx(p)) + abs(dFdy(p)); } -vec4 fwidth (vec4 p) { - return abs (dFdx (p)) + abs (dFdy (p)); +vec4 fwidth(const vec4 p) +{ + // XXX hand-write with __asm + return abs(dFdx(p)) + abs(dFdy(p)); } diff --git a/src/mesa/shader/slang/library/slang_fragment_builtin_gc.h b/src/mesa/shader/slang/library/slang_fragment_builtin_gc.h index a7249ea561d..8b8202ea2f9 100644 --- a/src/mesa/shader/slang/library/slang_fragment_builtin_gc.h +++ b/src/mesa/shader/slang/library/slang_fragment_builtin_gc.h @@ -52,36 +52,48 @@ 119,0,18,98,105,97,115,0,20,0,4,118,101,99,52,95,116,101,120,98,51,100,0,18,95,95,114,101,116,86, 97,108,0,0,18,115,97,109,112,108,101,114,0,0,18,112,99,111,111,114,100,0,0,0,0,1,0,12,0,116,101, 120,116,117,114,101,67,117,98,101,0,1,1,0,19,115,97,109,112,108,101,114,0,0,1,1,0,11,99,111,111, -114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,4,118,101,99,52,95,116,101,120,99,117,98,101,0,18,95,95, -114,101,116,86,97,108,0,0,18,115,97,109,112,108,101,114,0,0,18,99,111,111,114,100,0,0,18,98,105,97, -115,0,0,0,0,1,0,12,0,115,104,97,100,111,119,49,68,0,1,1,0,20,115,97,109,112,108,101,114,0,0,1,1,0, -11,99,111,111,114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,4,118,101,99,52,95,115,104,97,100,49,100,0, -18,95,95,114,101,116,86,97,108,0,0,18,115,97,109,112,108,101,114,0,0,18,99,111,111,114,100,0,0,18, -98,105,97,115,0,0,0,0,1,0,12,0,115,104,97,100,111,119,49,68,80,114,111,106,0,1,1,0,20,115,97,109, -112,108,101,114,0,0,1,1,0,12,99,111,111,114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,8,58,115,104,97, -100,111,119,49,68,0,18,115,97,109,112,108,101,114,0,0,58,118,101,99,51,0,18,99,111,111,114,100,0, -59,115,0,18,99,111,111,114,100,0,59,113,0,49,0,17,48,0,48,0,0,0,18,99,111,111,114,100,0,59,112,0, -18,99,111,111,114,100,0,59,113,0,49,0,0,0,18,98,105,97,115,0,0,0,0,0,1,0,12,0,115,104,97,100,111, -119,50,68,0,1,1,0,21,115,97,109,112,108,101,114,0,0,1,1,0,11,99,111,111,114,100,0,0,1,1,0,9,98,105, -97,115,0,0,0,1,4,118,101,99,52,95,115,104,97,100,50,100,0,18,95,95,114,101,116,86,97,108,0,0,18, -115,97,109,112,108,101,114,0,0,18,99,111,111,114,100,0,0,18,98,105,97,115,0,0,0,0,1,0,12,0,115,104, -97,100,111,119,50,68,80,114,111,106,0,1,1,0,21,115,97,109,112,108,101,114,0,0,1,1,0,12,99,111,111, -114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,8,58,115,104,97,100,111,119,50,68,0,18,115,97,109,112, -108,101,114,0,0,58,118,101,99,51,0,18,99,111,111,114,100,0,59,115,0,18,99,111,111,114,100,0,59,113, -0,49,0,18,99,111,111,114,100,0,59,116,0,18,99,111,111,114,100,0,59,113,0,49,0,18,99,111,111,114, -100,0,59,112,0,18,99,111,111,114,100,0,59,113,0,49,0,0,0,18,98,105,97,115,0,0,0,0,0,1,0,9,0,100,70, -100,120,0,1,0,0,9,112,0,0,0,1,8,17,48,0,48,48,49,0,0,0,0,1,0,10,0,100,70,100,120,0,1,0,0,10,112,0, -0,0,1,8,58,118,101,99,50,0,17,48,0,48,48,49,0,0,0,0,0,0,1,0,11,0,100,70,100,120,0,1,0,0,11,112,0,0, -0,1,8,58,118,101,99,51,0,17,48,0,48,48,49,0,0,0,0,0,0,1,0,12,0,100,70,100,120,0,1,0,0,12,112,0,0,0, -1,8,58,118,101,99,52,0,17,48,0,48,48,49,0,0,0,0,0,0,1,0,9,0,100,70,100,121,0,1,0,0,9,112,0,0,0,1,8, -17,48,0,48,48,49,0,0,0,0,1,0,10,0,100,70,100,121,0,1,0,0,10,112,0,0,0,1,8,58,118,101,99,50,0,17,48, -0,48,48,49,0,0,0,0,0,0,1,0,11,0,100,70,100,121,0,1,0,0,11,112,0,0,0,1,8,58,118,101,99,51,0,17,48,0, -48,48,49,0,0,0,0,0,0,1,0,12,0,100,70,100,121,0,1,0,0,12,112,0,0,0,1,8,58,118,101,99,52,0,17,48,0, -48,48,49,0,0,0,0,0,0,1,0,9,0,102,119,105,100,116,104,0,1,0,0,9,112,0,0,0,1,8,58,97,98,115,0,58,100, -70,100,120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0,46,0,0,1,0,10,0, -102,119,105,100,116,104,0,1,0,0,10,112,0,0,0,1,8,58,97,98,115,0,58,100,70,100,120,0,18,112,0,0,0,0, -0,58,97,98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0,46,0,0,1,0,11,0,102,119,105,100,116,104,0,1, -0,0,11,112,0,0,0,1,8,58,97,98,115,0,58,100,70,100,120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70, -100,121,0,18,112,0,0,0,0,0,46,0,0,1,0,12,0,102,119,105,100,116,104,0,1,0,0,12,112,0,0,0,1,8,58,97, -98,115,0,58,100,70,100,120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0, -46,0,0,0 +114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,3,2,0,12,1,99,111,111,114,100,52,0,0,0,9,18,99,111,111, +114,100,52,0,59,120,121,122,0,18,99,111,111,114,100,0,20,0,9,18,99,111,111,114,100,52,0,59,119,0, +18,98,105,97,115,0,20,0,4,118,101,99,52,95,116,101,120,99,117,98,101,0,18,95,95,114,101,116,86,97, +108,0,0,18,115,97,109,112,108,101,114,0,0,18,99,111,111,114,100,52,0,0,0,0,1,0,12,0,115,104,97,100, +111,119,49,68,0,1,1,0,20,115,97,109,112,108,101,114,0,0,1,1,0,11,99,111,111,114,100,0,0,1,1,0,9,98, +105,97,115,0,0,0,1,3,2,0,12,1,99,111,111,114,100,52,0,0,0,9,18,99,111,111,114,100,52,0,59,120,121, +122,0,18,99,111,111,114,100,0,20,0,9,18,99,111,111,114,100,52,0,59,119,0,18,98,105,97,115,0,20,0,4, +118,101,99,52,95,116,101,120,98,49,100,0,18,95,95,114,101,116,86,97,108,0,0,18,115,97,109,112,108, +101,114,0,0,18,99,111,111,114,100,52,0,0,0,0,1,0,12,0,115,104,97,100,111,119,49,68,80,114,111,106, +0,1,1,0,20,115,97,109,112,108,101,114,0,0,1,1,0,12,99,111,111,114,100,0,0,1,1,0,9,98,105,97,115,0, +0,0,1,3,2,0,12,1,112,99,111,111,114,100,0,0,0,9,18,112,99,111,111,114,100,0,59,120,0,18,99,111,111, +114,100,0,59,120,0,18,99,111,111,114,100,0,59,119,0,49,20,0,9,18,112,99,111,111,114,100,0,59,122,0, +18,99,111,111,114,100,0,59,122,0,20,0,9,18,112,99,111,111,114,100,0,59,119,0,18,98,105,97,115,0,20, +0,4,118,101,99,52,95,116,101,120,98,49,100,0,18,95,95,114,101,116,86,97,108,0,0,18,115,97,109,112, +108,101,114,0,0,18,112,99,111,111,114,100,0,0,0,0,1,0,12,0,115,104,97,100,111,119,50,68,0,1,1,0,21, +115,97,109,112,108,101,114,0,0,1,1,0,11,99,111,111,114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,3,2,0, +12,1,99,111,111,114,100,52,0,0,0,9,18,99,111,111,114,100,52,0,59,120,121,122,0,18,99,111,111,114, +100,0,20,0,9,18,99,111,111,114,100,52,0,59,119,0,18,98,105,97,115,0,20,0,4,118,101,99,52,95,116, +101,120,98,50,100,0,18,95,95,114,101,116,86,97,108,0,0,18,115,97,109,112,108,101,114,0,0,18,99,111, +111,114,100,52,0,0,0,0,1,0,12,0,115,104,97,100,111,119,50,68,80,114,111,106,0,1,1,0,21,115,97,109, +112,108,101,114,0,0,1,1,0,12,99,111,111,114,100,0,0,1,1,0,9,98,105,97,115,0,0,0,1,3,2,0,12,1,112, +99,111,111,114,100,0,0,0,9,18,112,99,111,111,114,100,0,59,120,121,0,18,99,111,111,114,100,0,59,120, +121,0,18,99,111,111,114,100,0,59,119,0,49,20,0,9,18,112,99,111,111,114,100,0,59,122,0,18,99,111, +111,114,100,0,59,122,0,20,0,9,18,112,99,111,111,114,100,0,59,119,0,18,98,105,97,115,0,20,0,4,118, +101,99,52,95,116,101,120,98,50,100,0,18,95,95,114,101,116,86,97,108,0,0,18,115,97,109,112,108,101, +114,0,0,18,112,99,111,111,114,100,0,0,0,0,1,0,9,0,100,70,100,120,0,1,1,0,9,112,0,0,0,1,4,118,101, +99,52,95,100,100,120,0,18,95,95,114,101,116,86,97,108,0,59,120,0,0,18,112,0,59,120,120,120,120,0,0, +0,0,1,0,10,0,100,70,100,120,0,1,1,0,10,112,0,0,0,1,4,118,101,99,52,95,100,100,120,0,18,95,95,114, +101,116,86,97,108,0,59,120,121,0,0,18,112,0,59,120,121,121,121,0,0,0,0,1,0,11,0,100,70,100,120,0,1, +1,0,11,112,0,0,0,1,4,118,101,99,52,95,100,100,120,0,18,95,95,114,101,116,86,97,108,0,59,120,121, +122,0,0,18,112,0,59,120,121,122,122,0,0,0,0,1,0,12,0,100,70,100,120,0,1,1,0,12,112,0,0,0,1,4,118, +101,99,52,95,100,100,120,0,18,95,95,114,101,116,86,97,108,0,0,18,112,0,0,0,0,1,0,9,0,100,70,100, +121,0,1,1,0,9,112,0,0,0,1,4,118,101,99,52,95,100,100,121,0,18,95,95,114,101,116,86,97,108,0,59,120, +0,0,18,112,0,59,120,120,120,120,0,0,0,0,1,0,10,0,100,70,100,121,0,1,1,0,10,112,0,0,0,1,4,118,101, +99,52,95,100,100,121,0,18,95,95,114,101,116,86,97,108,0,59,120,121,0,0,18,112,0,59,120,121,121,121, +0,0,0,0,1,0,11,0,100,70,100,121,0,1,1,0,11,112,0,0,0,1,4,118,101,99,52,95,100,100,121,0,18,95,95, +114,101,116,86,97,108,0,59,120,121,122,0,0,18,112,0,59,120,121,122,122,0,0,0,0,1,0,12,0,100,70,100, +121,0,1,1,0,12,112,0,0,0,1,4,118,101,99,52,95,100,100,121,0,18,95,95,114,101,116,86,97,108,0,0,18, +112,0,0,0,0,1,0,9,0,102,119,105,100,116,104,0,1,1,0,9,112,0,0,0,1,8,58,97,98,115,0,58,100,70,100, +120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0,46,0,0,1,0,10,0,102,119, +105,100,116,104,0,1,1,0,10,112,0,0,0,1,8,58,97,98,115,0,58,100,70,100,120,0,18,112,0,0,0,0,0,58,97, +98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0,46,0,0,1,0,11,0,102,119,105,100,116,104,0,1,1,0,11, +112,0,0,0,1,8,58,97,98,115,0,58,100,70,100,120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70,100,121, +0,18,112,0,0,0,0,0,46,0,0,1,0,12,0,102,119,105,100,116,104,0,1,1,0,12,112,0,0,0,1,8,58,97,98,115,0, +58,100,70,100,120,0,18,112,0,0,0,0,0,58,97,98,115,0,58,100,70,100,121,0,18,112,0,0,0,0,0,46,0,0,0 diff --git a/src/mesa/shader/slang/slang_assemble.c b/src/mesa/shader/slang/slang_assemble.c index 3995f0f31fb..05574d0e561 100644 --- a/src/mesa/shader/slang/slang_assemble.c +++ b/src/mesa/shader/slang/slang_assemble.c @@ -756,6 +756,8 @@ static const struct {"vec4_texcube", slang_asm_vec4_texcube, slang_asm_none}, {"vec4_shad1d", slang_asm_vec4_shad1d, slang_asm_none}, {"vec4_shad2d", slang_asm_vec4_shad2d, slang_asm_none}, + {"vec4_ddx", 0, slang_asm_none}, + {"vec4_ddy", 0, slang_asm_none}, /* GL_MESA_shader_debug */ {"float_print", slang_asm_float_deref, slang_asm_float_print}, {"int_print", slang_asm_int_deref, slang_asm_int_print}, diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c index 968fdcdd16c..79261ee29a6 100644 --- a/src/mesa/shader/slang/slang_codegen.c +++ b/src/mesa/shader/slang/slang_codegen.c @@ -619,6 +619,8 @@ static slang_asm_info AsmInfo[] = { { "vec4_frac", IR_FRAC, 1, 1 }, { "vec4_abs", IR_ABS, 1, 1 }, { "vec4_negate", IR_NEG, 1, 1 }, + { "vec4_ddx", IR_DDX, 1, 1 }, + { "vec4_ddy", IR_DDY, 1, 1 }, /* float binary op */ { "float_add", IR_ADD, 1, 2 }, { "float_subtract", IR_SUB, 1, 2 }, diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c index c62a07d97a1..87a3f3f337c 100644 --- a/src/mesa/shader/slang/slang_emit.c +++ b/src/mesa/shader/slang/slang_emit.c @@ -77,6 +77,8 @@ static slang_ir_info IrInfo[] = { { IR_FRAC, "IR_FRAC", OPCODE_FRC, 4, 1 }, { IR_ABS, "IR_ABS", OPCODE_ABS, 4, 1 }, { IR_NEG, "IR_NEG", 0/*spec case*/, 4, 1 }, + { IR_DDX, "IR_DDX", OPCODE_DDX, 4, 1 }, + { IR_DDX, "IR_DDY", OPCODE_DDX, 4, 1 }, { IR_SIN, "IR_SIN", OPCODE_SIN, 1, 1 }, { IR_COS, "IR_COS", OPCODE_COS, 1, 1 }, /* other */ @@ -701,6 +703,8 @@ emit(slang_gen_context *gc, slang_ir_node *n, struct gl_program *prog) case IR_ABS: case IR_SIN: case IR_COS: + case IR_DDX: + case IR_DDY: return emit_unop(gc, n, prog); case IR_TEX: case IR_TEXB: diff --git a/src/mesa/shader/slang/slang_ir.h b/src/mesa/shader/slang/slang_ir.h index 273964a8f06..b106e85dbb5 100644 --- a/src/mesa/shader/slang/slang_ir.h +++ b/src/mesa/shader/slang/slang_ir.h @@ -74,6 +74,8 @@ typedef enum IR_FRAC, IR_ABS, /* absolute value */ IR_NEG, /* negate */ + IR_DDX, /* derivative w.r.t. X */ + IR_DDY, /* derivative w.r.t. Y */ IR_SIN, /* sine */ IR_COS, /* cosine */ IR_NOT, /* logical not */ -- cgit v1.2.3