<div dir="ltr"><div dir="ltr"><div class="gmail_quote"><div dir="ltr">On Tue, Nov 13, 2018 at 9:48 AM Karol Herbst <<a href="mailto:kherbst@redhat.com">kherbst@redhat.com</a>> wrote: </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Not complete, mostly just adding things as I encounter them in CTS. But not getting far enough yet to hit most of the OpenCL.std instructions. Anyway, this is better than nothing and covers the most common builtins. Signed-off-by: Karol Herbst <<a href="mailto:kherbst@redhat.com" target="_blank">kherbst@redhat.com</a>> --- src/compiler/nir/meson.build | 1 + src/compiler/nir/nir_builtin_builder.c | 249 +++++++++++++++++++++- src/compiler/nir/nir_builtin_builder.h | 150 ++++++++++++- src/compiler/spirv/spirv_to_nir.c | 2 + src/compiler/spirv/vtn_alu.c | 15 ++ src/compiler/spirv/vtn_glsl450.c | 2 +- src/compiler/spirv/vtn_opencl.c | 284 +++++++++++++++++++++++++ src/compiler/spirv/vtn_private.h | 3 + 8 files changed, 701 insertions(+), 5 deletions(-) create mode 100644 src/compiler/spirv/vtn_opencl.c diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build index b0c3a7feb31..00d7f56e6eb 100644 --- a/src/compiler/nir/meson.build +++ b/src/compiler/nir/meson.build @@ -206,6 +206,7 @@ files_libnir = files( '../spirv/vtn_amd.c', '../spirv/vtn_cfg.c', '../spirv/vtn_glsl450.c', + '../spirv/vtn_opencl.c', '../spirv/vtn_private.h', '../spirv/vtn_subgroup.c', '../spirv/vtn_variables.c', diff --git a/src/compiler/nir/nir_builtin_builder.c b/src/compiler/nir/nir_builtin_builder.c index 252a7691f36..e37915e92ca 100644 --- a/src/compiler/nir/nir_builtin_builder.c +++ b/src/compiler/nir/nir_builtin_builder.c @@ -21,11 +21,43 @@ * IN THE SOFTWARE. */ +#include <math.h> + #include "nir.h" #include "nir_builtin_builder.h" nir_ssa_def* -nir_cross(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +nir_iadd_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) </blockquote><div> </div><div>For some of these, I'd rather we add NIR opcodes and some lowering code in nir_opt_algebraic. Intel hardware can do them directly so I'd like to be able to take advantage of that. In particular, [iu]add_sat, [iu]sub_sat, and [i]hadd. For more information, see</div><div> </div><div><a href="https://cgit.freedesktop.org/~idr/mesa/tree/docs/specs/INTEL_shader_integer_functions2.txt?h=INTEL_shader_integer_functions2">https://cgit.freedesktop.org/~idr/mesa/tree/docs/specs/INTEL_shader_integer_functions2.txt?h=INTEL_shader_integer_functions2</a></div><div> </div><div>We do have to do a small bit of lowering for some of them but it's way better than the pile of instructions you're emitting below. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +{ + int64_t max; + switch (x->bit_size) { + case 64: + max = INT64_MAX; + break; + case 32: + max = INT32_MAX; + break; + case 16: + max = INT16_MAX; + break; + case 8: + max = INT8_MAX; + break; + } + + nir_ssa_def *sum = nir_iadd(b, x, y); + + nir_ssa_def *hi = nir_bcsel(b, nir_ilt(b, sum, x), + nir_imm_intN_t(b, max, x->bit_size), sum); </blockquote><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + + nir_ssa_def *lo = nir_bcsel(b, nir_ilt(b, x, sum), + nir_imm_intN_t(b, max + 1, x->bit_size), sum); + + return nir_bcsel(b, nir_ige(b, y, nir_imm_intN_t(b, 1, y->bit_size)), hi, lo); +} + +nir_ssa_def* +nir_cross3(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) { unsigned yzx[3] = { 1, 2, 0 }; unsigned zxy[3] = { 2, 0, 1 }; @@ -36,6 +68,63 @@ nir_cross(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) nir_swizzle(b, y, yzx, 3, true))); } +nir_ssa_def* +nir_cross4(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *cross = nir_cross3(b, x, y); + + return nir_vec4(b, + nir_channel(b, cross, 0), + nir_channel(b, cross, 1), + nir_channel(b, cross, 2), + nir_imm_intN_t(b, 0, cross->bit_size)); +} + +static nir_ssa_def* +nir_hadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, bool sign) </blockquote><div> </div><div>We have this in hardware too. :-) </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +{ + nir_ssa_def *imm1 = nir_imm_int(b, 1); + + nir_ssa_def *t0 = nir_ixor(b, x, y); + nir_ssa_def *t1 = nir_iand(b, x, y); </blockquote><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + + nir_ssa_def *t2; + if (sign) + t2 = nir_ishr(b, t0, imm1); + else + t2 = nir_ushr(b, t0, imm1); + return nir_iadd(b, t1, t2); </blockquote><div> </div><div>This is seriously clever and needs a comment. :-) Perhaps something like this: </div></div><div class="gmail_quote"> </div><div class="gmail_quote">x + y = x - (x & ~y) + (x & ~y) + y - (~x & y) + (~x & y)</div><div class="gmail_quote"> = (x & y) + (x & ~y) + (x & y) + (~x & y)</div><div class="gmail_quote"> = 2 * (x & y) + (x & ~y) + (~x & y)</div><div class="gmail_quote"> = ((x & y) << 1) + (x ^ y)</div><div class="gmail_quote"> </div><div class="gmail_quote">and, this works for any pair of N-bit numbers. Since we know that the bottom bit of (x & y) << 1 is zero, </div><div class="gmail_quote"> </div><div class="gmail_quote">(x + y) >> 1 = (((x & y) << 1) + (x ^ y)) >> 1</div><div class="gmail_quote"> = (x & y) + ((x ^ y) >> 1)</div><div class="gmail_quote"> <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +} + +nir_ssa_def* +nir_ihadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + return nir_hadd(b, x, y, true); +} + +nir_ssa_def* +nir_uhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + return nir_hadd(b, x, y, false); +} + +nir_ssa_def* +nir_length(nir_builder *b, nir_ssa_def *vec) +{ + nir_ssa_def *finf = nir_imm_floatN_t(b, INFINITY, vec->bit_size); + + nir_ssa_def *abs = nir_fabs(b, vec); + if (vec->num_components == 1) + return abs; + + nir_ssa_def *maxc = nir_fmax(b, nir_channel(b, abs, 0), nir_channel(b, abs, 1)); + for (int i = 2; i < vec->num_components; ++i) + maxc = nir_fmax(b, maxc, nir_channel(b, abs, i)); + abs = nir_fdiv(b, abs, maxc); + nir_ssa_def *res = nir_fmul(b, nir_fsqrt(b, nir_fdot(b, abs, abs)), maxc); </blockquote><div> </div><div>We could make this even more precise if we made maxc a power of two so we don't loose anything multiplying and dividing by it. Unfortunately, I don't have a good way off-hand to do that quickly that handles denormals properly. This is probably good enough. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + return nir_bcsel(b, nir_feq(b, maxc, finf), maxc, res); +} + nir_ssa_def* nir_fast_length(nir_builder *b, nir_ssa_def *vec) { @@ -49,6 +138,107 @@ nir_fast_length(nir_builder *b, nir_ssa_def *vec) } } +nir_ssa_def* +nir_nextafter(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *zero = nir_imm_intN_t(b, 0, x->bit_size); + nir_ssa_def *one = nir_imm_intN_t(b, 1, x->bit_size); + nir_ssa_def *nzero = nir_imm_intN_t(b, 1ull << (x->bit_size - 1), x->bit_size); + + nir_ssa_def *condeq = nir_feq(b, x, y); + nir_ssa_def *conddir = nir_flt(b, x, y); + nir_ssa_def *condnzero = nir_feq(b, x, nzero); + + // beware of -0.0 - 1 == NaN + nir_ssa_def *xn = + nir_bcsel(b, + condnzero, + nir_imm_intN_t(b, (1 << (x->bit_size - 1)) + 1, x->bit_size), + nir_isub(b, x, one)); </blockquote><div> </div><div>What about 0.0 - 1? That's ~0 which isn't what we want either. Also, what is nextafter(-0.0, 1) suppsed to be? 0.0? Or is it the same as nextafter(0.0, 1)? </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + + // beware of -0.0 + 1 == -0x1p-149 + nir_ssa_def *xp = nir_bcsel(b, condnzero, one, nir_iadd(b, x, one)); + + // nextafter can be implemented by just +/- 1 on the int value + nir_ssa_def *resp = nir_bcsel(b, conddir, xp, xn); + nir_ssa_def *resn = nir_bcsel(b, conddir, xn, xp); + + nir_ssa_def *res = nir_bcsel(b, nir_flt(b, x, zero), resn, resp); </blockquote><div> </div><div>This would be more efficient as nir_bcsel(b, nir_ixor(b, conddir, nir_flt(b, x, zero)), xp, xn); </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + + return nir_nan_check2(b, x, y, nir_bcsel(b, condeq, x, res)); +} + +nir_ssa_def* +nir_normalize(nir_builder *b, nir_ssa_def *vec) +{ + nir_ssa_def *f0 = nir_imm_floatN_t(b, 0.0, vec->bit_size); + + nir_ssa_def *maxc; + nir_ssa_def *res; + if (vec->num_components == 1) { + nir_ssa_def *f1p = nir_imm_floatN_t(b, 1.0, vec->bit_size); + nir_ssa_def *f1n = nir_imm_floatN_t(b, -1.0, vec->bit_size); + + nir_ssa_def *cond = nir_flt(b, vec, f0); + res = nir_bcsel(b, cond, f1n, f1p); </blockquote><div> </div><div>Isn't this just fsign(x)? </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + maxc = vec; + } else { + maxc = nir_fmax(b, nir_fabs(b, nir_channel(b, vec, 0)), + nir_fabs(b, nir_channel(b, vec, 1))); + for (int i = 2; i < vec->num_components; ++i) + maxc = nir_fmax(b, maxc, nir_fabs(b, nir_channel(b, vec, i))); </blockquote><div> </div><div>Might be worth adding a helper for this now that I've seen it twice. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + nir_ssa_def *temp = nir_fdiv(b, vec, maxc); </blockquote><div> </div><div>What if maxc is inf or zero? </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + res = nir_fmul(b, temp, nir_frsq(b, nir_fdot(b, temp, temp))); + } + return nir_bcsel(b, nir_feq(b, maxc, f0), vec, res); +} + +static nir_ssa_def* +nir_rhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, bool sign) </blockquote><div> </div><div>We have this in hardware too. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +{ + nir_ssa_def *imm1 = nir_imm_int(b, 1); + + nir_ssa_def *t0 = nir_ixor(b, x, y); + nir_ssa_def *t1 = nir_ior(b, x, y); + + nir_ssa_def *t2; + if (sign) + t2 = nir_ishr(b, t0, imm1); + else + t2 = nir_ushr(b, t0, imm1); + + return nir_isub(b, t1, t2); </blockquote><div> </div><div>Comment:</div><div> </div><div>x + y + 1 = x + (~x & y) - (~x & y) + y + (x & ~y) - (x & ~y) + 1 </div><div> = (x | y) - (~x & y) + (x | y) - (x & ~y) + 1 </div><div> = 2 * (x | y) - ((~x & y) + (x & ~y)) + 1</div><div> = ((x | y) << 1) - (x ^ y) + 1</div></div><div class="gmail_quote"> </div><div class="gmail_quote">and, this works for any pair of N-bit numbers. Since we know that the bottom bit of (x & y) << 1 is zero,</div><div class="gmail_quote"> </div><div class="gmail_quote">(x + y + 1) >> 1 = (x | y) + (-(x ^ y) + 1) >> 1)</div><div class="gmail_quote"> = (x | y) - ((x ^ y) >> 1)</div><div class="gmail_quote"> </div><div class="gmail_quote">I'm really not sure why that last shift step works. I'll have to think on it some more. I wanted to send this out now so that you can look at it and respond. I'll comment more tomorrow. </div><div class="gmail_quote"><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +} + +nir_ssa_def* +nir_irhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + return nir_rhadd(b, x, y, true); +} + +nir_ssa_def* +nir_urhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + return nir_rhadd(b, x, y, false); +} + +nir_ssa_def* +nir_rotate(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *shift_mask = nir_imm_int(b, x->bit_size - 1); + + if (y->bit_size != 32) + y = nir_u2u32(b, y); + + nir_ssa_def *lshift = nir_iand(b, y, shift_mask); + nir_ssa_def *rshift = nir_isub(b, nir_imm_int(b, x->bit_size), lshift); + + nir_ssa_def *hi = nir_ishl(b, x, lshift); + nir_ssa_def *lo = nir_ushr(b, x, rshift); + + return nir_ior(b, hi, lo); +} + nir_ssa_def* nir_smoothstep(nir_builder *b, nir_ssa_def *edge0, nir_ssa_def *edge1, nir_ssa_def *x) { @@ -63,3 +253,60 @@ nir_smoothstep(nir_builder *b, nir_ssa_def *edge0, nir_ssa_def *edge1, nir_ssa_d /* result = t * t * (3 - 2 * t) */ return nir_fmul(b, t, nir_fmul(b, t, nir_fsub(b, f3, nir_fmul(b, f2, t)))); } + +nir_ssa_def* +nir_isub_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) </blockquote><div> </div><div>We have this in hardware too. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> +{ + uint64_t max = (1ull << (x->bit_size - 1)) - 1; + + nir_ssa_def *diff = nir_isub(b, x, y); + nir_ssa_def *hi = nir_bcsel(b, nir_ilt(b, diff, x), + nir_imm_intN_t(b, max, x->bit_size), diff); + nir_ssa_def *lo = nir_bcsel(b, nir_ilt(b, x, diff), + nir_imm_intN_t(b, max + 1, x->bit_size), diff); + return nir_bcsel(b, nir_ilt(b, y, nir_imm_intN_t(b, 0, y->bit_size)), hi, lo); +} + +nir_ssa_def* +nir_iupsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo) +{ + nir_ssa_def *hiup; + nir_ssa_def *loup; + switch (hi->bit_size) { + case 32: + hiup = nir_i2i64(b, hi); + loup = nir_i2i64(b, lo); + break; + case 16: + hiup = nir_i2i32(b, hi); + loup = nir_i2i32(b, lo); + break; + case 8: + hiup = nir_i2i16(b, hi); + loup = nir_i2i16(b, lo); </blockquote><div> </div><div>We really need a nir_i2i helper that takes a bit size. Mind making one and making this use it? </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + break; + } + return nir_ior(b, nir_ishl(b, hiup, nir_imm_int(b, hi->bit_size)), loup); +} + +nir_ssa_def* +nir_uupsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo) +{ + nir_ssa_def *hiup; + nir_ssa_def *loup; + switch (hi->bit_size) { + case 32: + hiup = nir_u2u64(b, hi); + loup = nir_u2u64(b, lo); + break; + case 16: + hiup = nir_u2u32(b, hi); + loup = nir_u2u32(b, lo); + break; + case 8: + hiup = nir_u2u16(b, hi); + loup = nir_u2u16(b, lo); </blockquote><div> </div><div>Same here. </div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"> + break; + } + return nir_ior(b, nir_ishl(b, hiup, nir_imm_int(b, hi->bit_size)), loup); +} diff --git a/src/compiler/nir/nir_builtin_builder.h b/src/compiler/nir/nir_builtin_builder.h index 0e5b9db462a..ee7726cae0b 100644 --- a/src/compiler/nir/nir_builtin_builder.h +++ b/src/compiler/nir/nir_builtin_builder.h @@ -31,10 +31,61 @@ * Definitions for functions in the C file come first. */ -nir_ssa_def* nir_cross(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_iadd_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_cross3(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_cross4(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_length(nir_builder *b, nir_ssa_def *vec); nir_ssa_def* nir_fast_length(nir_builder *b, nir_ssa_def *vec); +nir_ssa_def* nir_ihadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_uhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_nextafter(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_normalize(nir_builder *b, nir_ssa_def *vec); +nir_ssa_def* nir_irhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_urhadd(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_rotate(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); nir_ssa_def* nir_smoothstep(nir_builder *b, nir_ssa_def *edge0, nir_ssa_def *edge1, nir_ssa_def *x); +nir_ssa_def* nir_isub_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y); +nir_ssa_def* nir_iupsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo); +nir_ssa_def* nir_uupsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo); + +static inline nir_ssa_def * +nir_nan_check2(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *res) +{ + return nir_bcsel(b, nir_fne(b, x, x), x, nir_bcsel(b, nir_fne(b, y, y), y, res)); +} + +static inline nir_ssa_def * +nir_iabs_diff(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *cond = nir_ige(b, x, y); + nir_ssa_def *res0 = nir_isub(b, x, y); + nir_ssa_def *res1 = nir_isub(b, y, x); + return nir_bcsel(b, cond, res0, res1); +} + +static inline nir_ssa_def * +nir_uabs_diff(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *cond = nir_uge(b, x, y); + nir_ssa_def *res0 = nir_isub(b, x, y); + nir_ssa_def *res1 = nir_isub(b, y, x); + return nir_bcsel(b, cond, res0, res1); +} + +static inline nir_ssa_def * +nir_uadd_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *sum = nir_iadd(b, x, y); + nir_ssa_def *cond = nir_ult(b, sum, x); + return nir_bcsel(b, cond, nir_imm_intN_t(b, -1l, x->bit_size), sum); +} + +static inline nir_ssa_def * +nir_bitselect(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *s) +{ + return nir_ior(b, nir_iand(b, nir_inot(b, s), x), nir_iand(b, s, y)); +} static inline nir_ssa_def * nir_fclamp(nir_builder *b, @@ -57,10 +108,40 @@ nir_uclamp(nir_builder *b, return nir_umin(b, nir_umax(b, x, min_val), max_val); } +static inline nir_ssa_def * +nir_copysign(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + uint64_t masks = 1ull << (x->bit_size - 1); + uint64_t maskv = ~masks; + + nir_ssa_def *s = nir_imm_intN_t(b, masks, x->bit_size); + nir_ssa_def *v = nir_imm_intN_t(b, maskv, x->bit_size); + + return nir_ior(b, nir_iand(b, x, v), nir_iand(b, y, s)); +} + static inline nir_ssa_def * nir_degrees(nir_builder *b, nir_ssa_def *val) { - return nir_fmul(b, val, nir_imm_float(b, 57.2957795131)); + nir_ssa_def *c = nir_imm_floatN_t(b, 57.29577951308232, val->bit_size); + return nir_fmul(b, val, c); +} + +static inline nir_ssa_def * +nir_fdim(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *cond = nir_flt(b, y, x); + nir_ssa_def *res = nir_fsub(b, x, y); + nir_ssa_def *zero = nir_imm_floatN_t(b, 0.0, x->bit_size); + + // return NaN if either x or y are NaN, else x-y if x>y, else +0.0 + return nir_nan_check2(b, x, y, nir_bcsel(b, cond, res, zero)); +} + +static inline nir_ssa_def * +nir_distance(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + return nir_length(b, nir_fsub(b, x, y)); } static inline nir_ssa_def * @@ -75,10 +156,73 @@ nir_fast_normalize(nir_builder *b, nir_ssa_def *vec) return nir_fdiv(b, vec, nir_fast_length(b, vec)); } +static inline nir_ssa_def* +nir_fmad(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *z) +{ + return nir_fadd(b, nir_fmul(b, x, y), z); +} + +static inline nir_ssa_def* +nir_maxmag(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *xabs = nir_fabs(b, x); + nir_ssa_def *yabs = nir_fabs(b, y); + + nir_ssa_def *condy = nir_flt(b, xabs, yabs); + nir_ssa_def *condx = nir_flt(b, yabs, xabs); + + return nir_bcsel(b, condy, y, nir_bcsel(b, condx, x, nir_fmax(b, x, y))); +} + +static inline nir_ssa_def* +nir_minmag(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *xabs = nir_fabs(b, x); + nir_ssa_def *yabs = nir_fabs(b, y); + + nir_ssa_def *condx = nir_flt(b, xabs, yabs); + nir_ssa_def *condy = nir_flt(b, yabs, xabs); + + return nir_bcsel(b, condy, y, nir_bcsel(b, condx, x, nir_fmin(b, x, y))); +} + +static inline nir_ssa_def* +nir_nan(nir_builder *b, nir_ssa_def *x) +{ + nir_ssa_def *nan = nir_imm_floatN_t(b, NAN, x->bit_size); + if (x->num_components == 1) + return nan; + + nir_ssa_def *nans[NIR_MAX_VEC_COMPONENTS]; + for (unsigned i = 0; i < x->num_components; ++i) + nans[i] = nan; + + return nir_vec(b, nans, x->num_components); +} + static inline nir_ssa_def * nir_radians(nir_builder *b, nir_ssa_def *val) { - return nir_fmul(b, val, nir_imm_float(b, 0.01745329251)); + nir_ssa_def *c = nir_imm_floatN_t(b, 0.017453292519943295, val->bit_size); + return nir_fmul(b, val, c); +} + +static inline nir_ssa_def * +nir_select(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *s) +{ + if (s->num_components != 1) { + uint64_t mask = 1ull << (s->bit_size - 1); + s = nir_iand(b, s, nir_imm_intN_t(b, mask, s->bit_size)); + } + return nir_bcsel(b, nir_ieq(b, s, nir_imm_intN_t(b, 0, s->bit_size)), x, y); +} + +static inline nir_ssa_def * +nir_usub_sat(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y) +{ + nir_ssa_def *sum = nir_isub(b, x, y); + nir_ssa_def *cond = nir_ult(b, x, y); + return nir_bcsel(b, cond, nir_imm_intN_t(b, 0, x->bit_size), sum); } #endif /* NIR_BUILTIN_BUILDER_H */ diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c index 96ff09c3659..d72f07dc1f9 100644 --- a/src/compiler/spirv/spirv_to_nir.c +++ b/src/compiler/spirv/spirv_to_nir.c @@ -395,6 +395,8 @@ vtn_handle_extension(struct vtn_builder *b, SpvOp opcode, } else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0) && (b->options && b->options->caps.trinary_minmax)) { val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction; + } else if (strcmp(ext, "OpenCL.std") == 0) { + val->ext_handler = vtn_handle_opencl_instruction; } else { vtn_fail("Unsupported extension: %s", ext); } diff --git a/src/compiler/spirv/vtn_alu.c b/src/compiler/spirv/vtn_alu.c index 6860e7dc090..b1492c1501a 100644 --- a/src/compiler/spirv/vtn_alu.c +++ b/src/compiler/spirv/vtn_alu.c @@ -683,6 +683,21 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode, break; } + case SpvOpSignBitSet: { + unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type); + if (src[0]->num_components == 1) + val->ssa->def = + nir_ushr(&b->nb, src[0], nir_imm_int(&b->nb, src_bit_size - 1)); + else + val->ssa->def = + nir_ishr(&b->nb, src[0], nir_imm_int(&b->nb, src_bit_size - 1)); + + if (src_bit_size != 32) + val->ssa->def = nir_u2u32(&b->nb, val->ssa->def); + + break; + } + default: { bool swap; unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type); diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c index 06a49e48e3f..389e067712b 100644 --- a/src/compiler/spirv/vtn_glsl450.c +++ b/src/compiler/spirv/vtn_glsl450.c @@ -585,7 +585,7 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint, return; case GLSLstd450Cross: { - val->ssa->def = nir_cross(nb, src[0], src[1]); + val->ssa->def = nir_cross3(nb, src[0], src[1]); return; } diff --git a/src/compiler/spirv/vtn_opencl.c b/src/compiler/spirv/vtn_opencl.c new file mode 100644 index 00000000000..089e6168fd8 --- /dev/null +++ b/src/compiler/spirv/vtn_opencl.c @@ -0,0 +1,284 @@ +/* + * Copyright © 2018 Red Hat + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Rob Clark (<a href="mailto:robdclark@gmail.com" target="_blank">robdclark@gmail.com</a>) + */ + +#include "math.h" + +#include "nir/nir_builtin_builder.h" + +#include "vtn_private.h" +#include "OpenCL.std.h" + +typedef nir_ssa_def *(*nir_handler)(struct vtn_builder *b, enum OpenCLstd opcode, + unsigned num_srcs, nir_ssa_def **srcs, + const struct glsl_type *dest_type); + +static void +handle_instr(struct vtn_builder *b, enum OpenCLstd opcode, const uint32_t *w, + unsigned count, nir_handler handler) +{ + const struct glsl_type *dest_type = + vtn_value(b, w[1], vtn_value_type_type)->type->type; + + unsigned num_srcs = count - 5; + nir_ssa_def *srcs[3] = { NULL }; + vtn_assert(num_srcs <= ARRAY_SIZE(srcs)); + for (unsigned i = 0; i < num_srcs; i++) { + srcs[i] = vtn_ssa_value(b, w[i + 5])->def; + } + + nir_ssa_def *result = handler(b, opcode, num_srcs, srcs, dest_type); + if (result) { + struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); + val->ssa = vtn_create_ssa_value(b, dest_type); + val->ssa->def = result; + } else { + vtn_assert(dest_type == glsl_void_type()); + } +} + +static nir_op +nir_alu_op_for_opencl_opcode(struct vtn_builder *b, enum OpenCLstd opcode) +{ + switch (opcode) { + case Fabs: return nir_op_fabs; + case SAbs: return nir_op_iabs; + case Ceil: return nir_op_fceil; + case Cos: return nir_op_fcos; + case Exp2: return nir_op_fexp2; + case Log2: return nir_op_flog2; + case Floor: return nir_op_ffloor; + case Fma: return nir_op_ffma; + case Fmax: return nir_op_fmax; + case SMax: return nir_op_imax; + case UMax: return nir_op_umax; + case Fmin: return nir_op_fmin; + case SMin: return nir_op_imin; + case UMin: return nir_op_umin; + case Fmod: return nir_op_fmod; + case Mix: return nir_op_flrp; + case SMul_hi: return nir_op_imul_high; + case UMul_hi: return nir_op_umul_high; + case Popcount: return nir_op_bit_count; + case Pow: return nir_op_fpow; + case Remainder: return nir_op_frem; + case Rsqrt: return nir_op_frsq; + case Sign: return nir_op_fsign; + case Sin: return nir_op_fsin; + case Sqrt: return nir_op_fsqrt; + case Trunc: return nir_op_ftrunc; + /* uhm... */ + case UAbs: return nir_op_imov; + default: + vtn_fail("No NIR equivalent"); + } +} + +static nir_ssa_def * +handle_alu(struct vtn_builder *b, enum OpenCLstd opcode, unsigned num_srcs, + nir_ssa_def **srcs, const struct glsl_type *dest_type) +{ + return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode), + srcs[0], srcs[1], srcs[2], NULL); +} + +static nir_ssa_def * +handle_special(struct vtn_builder *b, enum OpenCLstd opcode, unsigned num_srcs, + nir_ssa_def **srcs, const struct glsl_type *dest_type) +{ + nir_builder *nb = &b->nb; + + switch (opcode) { + case SAbs_diff: + return nir_iabs_diff(nb, srcs[0], srcs[1]); + case UAbs_diff: + return nir_uabs_diff(nb, srcs[0], srcs[1]); + case SAdd_sat: + return nir_iadd_sat(nb, srcs[0], srcs[1]); + case UAdd_sat: + return nir_uadd_sat(nb, srcs[0], srcs[1]); + case Bitselect: + return nir_bitselect(nb, srcs[0], srcs[1], srcs[2]); + case FClamp: + return nir_fclamp(nb, srcs[0], srcs[1], srcs[2]); + case SClamp: + return nir_iclamp(nb, srcs[0], srcs[1], srcs[2]); + case UClamp: + return nir_uclamp(nb, srcs[0], srcs[1], srcs[2]); + case Copysign: + return nir_copysign(nb, srcs[0], srcs[1]); + case Cross: + if (glsl_get_components(dest_type) == 4) + return nir_cross4(nb, srcs[0], srcs[1]); + return nir_cross3(nb, srcs[0], srcs[1]); + case Degrees: + return nir_degrees(nb, srcs[0]); + case Fdim: + return nir_fdim(nb, srcs[0], srcs[1]); + case Distance: + return nir_distance(nb, srcs[0], srcs[1]); + case Fast_distance: + return nir_fast_distance(nb, srcs[0], srcs[1]); + case Fast_length: + return nir_fast_length(nb, srcs[0]); + case Fast_normalize: + return nir_fast_normalize(nb, srcs[0]); + case SHadd: + return nir_ihadd(nb, srcs[0], srcs[1]); + case UHadd: + return nir_uhadd(nb, srcs[0], srcs[1]); + case Length: + return nir_length(nb, srcs[0]); + case Mad: + return nir_fmad(nb, srcs[0], srcs[1], srcs[2]); + case Maxmag: + return nir_maxmag(nb, srcs[0], srcs[1]); + case Minmag: + return nir_minmag(nb, srcs[0], srcs[1]); + case Nan: + return nir_nan(nb, srcs[0]); + case Nextafter: + return nir_nextafter(nb, srcs[0], srcs[1]); + case Normalize: + return nir_normalize(nb, srcs[0]); + case Radians: + return nir_radians(nb, srcs[0]); + case SRhadd: + return nir_irhadd(nb, srcs[0], srcs[1]); + case URhadd: + return nir_urhadd(nb, srcs[0], srcs[1]); + case Rotate: + return nir_rotate(nb, srcs[0], srcs[1]); + case Smoothstep: + return nir_smoothstep(nb, srcs[0], srcs[1], srcs[2]); + case Select: + return nir_select(nb, srcs[0], srcs[1], srcs[2]); + case Step: + return nir_sge(nb, srcs[1], srcs[0]); + case SSub_sat: + return nir_isub_sat(nb, srcs[0], srcs[1]); + case USub_sat: + return nir_usub_sat(nb, srcs[0], srcs[1]); + case S_Upsample: + return nir_iupsample(nb, srcs[0], srcs[1]); + case U_Upsample: + return nir_uupsample(nb, srcs[0], srcs[1]); + default: + vtn_fail("No NIR equivalent"); + return NULL; + } +} + +static nir_ssa_def * +handle_printf(struct vtn_builder *b, enum OpenCLstd opcode, unsigned num_srcs, + nir_ssa_def **srcs, const struct glsl_type *dest_type) +{ + /* hahah, yeah, right.. */ + return nir_imm_int(&b->nb, -1); +} + +bool +vtn_handle_opencl_instruction(struct vtn_builder *b, uint32_t ext_opcode, + const uint32_t *w, unsigned count) +{ + switch (ext_opcode) { + case Fabs: + case SAbs: + case UAbs: + case Ceil: + case Cos: + case Exp2: + case Log2: + case Floor: + case Fma: + case Fmax: + case SMax: + case UMax: + case Fmin: + case SMin: + case UMin: + case Mix: + case Fmod: + case SMul_hi: + case UMul_hi: + case Popcount: + case Pow: + case Remainder: + case Rsqrt: + case Sign: + case Sin: + case Sqrt: + case Trunc: + handle_instr(b, ext_opcode, w, count, handle_alu); + return true; + case SAbs_diff: + case UAbs_diff: + case SAdd_sat: + case UAdd_sat: + case Bitselect: + case FClamp: + case SClamp: + case UClamp: + case Copysign: + case Cross: + case Degrees: + case Fdim: + case Distance: + case Fast_distance: + case Fast_length: + case Fast_normalize: + case SHadd: + case UHadd: + case Length: + case Mad: + case Maxmag: + case Minmag: + case Nan: + case Nextafter: + case Normalize: + case Radians: + case SRhadd: + case URhadd: + case Rotate: + case Select: + case Step: + case Smoothstep: + case SSub_sat: + case USub_sat: + case S_Upsample: + case U_Upsample: + handle_instr(b, ext_opcode, w, count, handle_special); + return true; + case Printf: + handle_instr(b, ext_opcode, w, count, handle_printf); + return true; + case Prefetch: + /* TODO maybe add a nir instruction for this? */ + return true; + default: + vtn_fail("unhandled opencl opc: %u\n", ext_opcode); + return false; + } +} diff --git a/src/compiler/spirv/vtn_private.h b/src/compiler/spirv/vtn_private.h index da7a04ce59f..643a88d1abe 100644 --- a/src/compiler/spirv/vtn_private.h +++ b/src/compiler/spirv/vtn_private.h @@ -745,6 +745,9 @@ void vtn_handle_subgroup(struct vtn_builder *b, SpvOp opcode, bool vtn_handle_glsl450_instruction(struct vtn_builder *b, SpvOp ext_opcode, const uint32_t *words, unsigned count); +bool vtn_handle_opencl_instruction(struct vtn_builder *b, uint32_t ext_opcode, + const uint32_t *words, unsigned count); + struct vtn_builder* vtn_create_builder(const uint32_t *words, size_t word_count, gl_shader_stage stage, const char *entry_point_name, const struct spirv_to_nir_options *options); -- 2.19.1 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org" target="_blank">mesa-dev@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div></div></div>