motorhead/qemu
1
0
Fork 0
mirror of https://github.com/Motorhead1991/qemu.git synced 2025-08-05 08:43:55 -06:00
Code Issues Projects Releases Packages Wiki Activity Actions

Hexagon (tests/tcg/hexagon) Remove __builtin from scatter_gather

Browse source 
Replace __builtin_* with inline assembly
    The __builtin's are subject to change with different compiler
    releases, so might break
Mark arrays as aligned when accessed as HVX vectors
Clean up comments

Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Reviewed-by: Anton Johansson <anjo@rev.ng>
Message-Id: <20230307025828.1612809-10-tsimpson@quicinc.com>
This commit is contained in:
Taylor Simpson 2023-03-06 18:58:23 -08:00
parent d00d68b504
commit c367938540
1 changed files with 260 additions and 231 deletions
Download patch file Download diff file Expand all files Collapse all files

491
tests/tcg/hexagon/scatter_gather.c
View file

@ -1,5 +1,5 @@
/* /*
* Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved. * Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
* *
* This program is free software; you can redistribute it and/or modify * This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by * it under the terms of the GNU General Public License as published by
@ -40,47 +40,6 @@ typedef long HVX_VectorPair __attribute__((__vector_size__(256)))
typedef long HVX_VectorPred __attribute__((__vector_size__(128))) typedef long HVX_VectorPred __attribute__((__vector_size__(128)))
__attribute__((aligned(128))); __attribute__((aligned(128)));
#define VSCATTER_16(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermh_128B((int)BASE, RGN, OFF, VALS)
#define VSCATTER_16_MASKED(MASK, BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermhq_128B(MASK, (int)BASE, RGN, OFF, VALS)
#define VSCATTER_32(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermw_128B((int)BASE, RGN, OFF, VALS)
#define VSCATTER_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
#define VSCATTER_16_32(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermhw_128B((int)BASE, RGN, OFF, VALS)
#define VSCATTER_16_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermhwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
#define VSCATTER_16_ACC(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermh_add_128B((int)BASE, RGN, OFF, VALS)
#define VSCATTER_32_ACC(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermw_add_128B((int)BASE, RGN, OFF, VALS)
#define VSCATTER_16_32_ACC(BASE, RGN, OFF, VALS) \
__builtin_HEXAGON_V6_vscattermhw_add_128B((int)BASE, RGN, OFF, VALS)
#define VGATHER_16(DSTADDR, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermh_128B(DSTADDR, (int)BASE, RGN, OFF)
#define VGATHER_16_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermhq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
#define VGATHER_32(DSTADDR, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermw_128B(DSTADDR, (int)BASE, RGN, OFF)
#define VGATHER_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
#define VGATHER_16_32(DSTADDR, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermhw_128B(DSTADDR, (int)BASE, RGN, OFF)
#define VGATHER_16_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
__builtin_HEXAGON_V6_vgathermhwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
#define VSHUFF_H(V) \
__builtin_HEXAGON_V6_vshuffh_128B(V)
#define VSPLAT_H(X) \
__builtin_HEXAGON_V6_lvsplath_128B(X)
#define VAND_VAL(PRED, VAL) \
__builtin_HEXAGON_V6_vandvrt_128B(PRED, VAL)
#define VDEAL_H(V) \
__builtin_HEXAGON_V6_vdealh_128B(V)
int err; int err;
/* define the number of rows/cols in a square matrix */ /* define the number of rows/cols in a square matrix */
@ -108,22 +67,22 @@ unsigned short vscatter16_32_ref[SCATTER_BUFFER_SIZE];
unsigned short vgather16_32_ref[MATRIX_SIZE]; unsigned short vgather16_32_ref[MATRIX_SIZE];
/* declare the arrays of offsets */ /* declare the arrays of offsets */
unsigned short half_offsets[MATRIX_SIZE]; unsigned short half_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned int word_offsets[MATRIX_SIZE]; unsigned int word_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
/* declare the arrays of values */ /* declare the arrays of values */
unsigned short half_values[MATRIX_SIZE]; unsigned short half_values[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned short half_values_acc[MATRIX_SIZE]; unsigned short half_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned short half_values_masked[MATRIX_SIZE]; unsigned short half_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned int word_values[MATRIX_SIZE]; unsigned int word_values[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned int word_values_acc[MATRIX_SIZE]; unsigned int word_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned int word_values_masked[MATRIX_SIZE]; unsigned int word_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
/* declare the arrays of predicates */ /* declare the arrays of predicates */
unsigned short half_predicates[MATRIX_SIZE]; unsigned short half_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
unsigned int word_predicates[MATRIX_SIZE]; unsigned int word_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
/* make this big enough for all the intrinsics */ /* make this big enough for all the operations */
const size_t region_len = sizeof(vtcm); const size_t region_len = sizeof(vtcm);
/* optionally add sync instructions */ /* optionally add sync instructions */
@ -261,164 +220,201 @@ void create_offsets_values_preds_16_32(void)
} }
} }
/* scatter the 16 bit elements using intrinsics */ /* scatter the 16 bit elements using HVX */
void vector_scatter_16(void) void vector_scatter_16(void)
{ {
/* copy the offsets and values to vectors */ asm ("m0 = %1\n\t"
HVX_Vector offsets = *(HVX_Vector *)half_offsets; "v0 = vmem(%2 + #0)\n\t"
HVX_Vector values = *(HVX_Vector *)half_values; "v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.h).h = v1\n\t"
VSCATTER_16(&vtcm.vscatter16, region_len, offsets, values); : : "r"(vtcm.vscatter16), "r"(region_len),
"r"(half_offsets), "r"(half_values)
: "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16); sync_scatter(vtcm.vscatter16);
} }
/* scatter-accumulate the 16 bit elements using intrinsics */ /* scatter-accumulate the 16 bit elements using HVX */
void vector_scatter_16_acc(void) void vector_scatter_16_acc(void)
{ {
/* copy the offsets and values to vectors */ asm ("m0 = %1\n\t"
HVX_Vector offsets = *(HVX_Vector *)half_offsets; "v0 = vmem(%2 + #0)\n\t"
HVX_Vector values = *(HVX_Vector *)half_values_acc; "v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.h).h += v1\n\t"
VSCATTER_16_ACC(&vtcm.vscatter16, region_len, offsets, values); : : "r"(vtcm.vscatter16), "r"(region_len),
"r"(half_offsets), "r"(half_values_acc)
: "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16); sync_scatter(vtcm.vscatter16);
} }
/* scatter the 16 bit elements using intrinsics */ /* masked scatter the 16 bit elements using HVX */
void vector_scatter_16_masked(void) void vector_scatter_16_masked(void)
{ {
/* copy the offsets and values to vectors */ asm ("r1 = #-1\n\t"
HVX_Vector offsets = *(HVX_Vector *)half_offsets; "v0 = vmem(%0 + #0)\n\t"
HVX_Vector values = *(HVX_Vector *)half_values_masked; "q0 = vand(v0, r1)\n\t"
HVX_Vector pred_reg = *(HVX_Vector *)half_predicates; "m0 = %2\n\t"
HVX_VectorPred preds = VAND_VAL(pred_reg, ~0); "v0 = vmem(%3 + #0)\n\t"
"v1 = vmem(%4 + #0)\n\t"
VSCATTER_16_MASKED(preds, &vtcm.vscatter16, region_len, offsets, values); "if (q0) vscatter(%1, m0, v0.h).h = v1\n\t"
: : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
"r"(half_offsets), "r"(half_values_masked)
: "r1", "q0", "m0", "q0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16); sync_scatter(vtcm.vscatter16);
} }
/* scatter the 32 bit elements using intrinsics */ /* scatter the 32 bit elements using HVX */
void vector_scatter_32(void) void vector_scatter_32(void)
{ {
/* copy the offsets and values to vectors */ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
HVX_Vector offsetslo = *(HVX_Vector *)word_offsets; HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2]; HVX_Vector *valueslo = (HVX_Vector *)word_values;
HVX_Vector valueslo = *(HVX_Vector *)word_values; HVX_Vector *valueshi = (HVX_Vector *)&word_values[MATRIX_SIZE / 2];
HVX_Vector valueshi = *(HVX_Vector *)&word_values[MATRIX_SIZE / 2];
VSCATTER_32(&vtcm.vscatter32, region_len, offsetslo, valueslo); asm ("m0 = %1\n\t"
VSCATTER_32(&vtcm.vscatter32, region_len, offsetshi, valueshi); "v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.w).w = v1\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetslo), "r"(valueslo)
: "m0", "v0", "v1", "memory");
asm ("m0 = %1\n\t"
"v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.w).w = v1\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetshi), "r"(valueshi)
: "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter32); sync_scatter(vtcm.vscatter32);
} }
/* scatter-acc the 32 bit elements using intrinsics */ /* scatter-accumulate the 32 bit elements using HVX */
void vector_scatter_32_acc(void) void vector_scatter_32_acc(void)
{ {
/* copy the offsets and values to vectors */ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
HVX_Vector offsetslo = *(HVX_Vector *)word_offsets; HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2]; HVX_Vector *valueslo = (HVX_Vector *)word_values_acc;
HVX_Vector valueslo = *(HVX_Vector *)word_values_acc; HVX_Vector *valueshi = (HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
HVX_Vector valueshi = *(HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetslo, valueslo); asm ("m0 = %1\n\t"
VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetshi, valueshi); "v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.w).w += v1\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetslo), "r"(valueslo)
: "m0", "v0", "v1", "memory");
asm ("m0 = %1\n\t"
"v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%3 + #0)\n\t"
"vscatter(%0, m0, v0.w).w += v1\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetshi), "r"(valueshi)
: "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter32); sync_scatter(vtcm.vscatter32);
} }
/* scatter the 32 bit elements using intrinsics */ /* masked scatter the 32 bit elements using HVX */
void vector_scatter_32_masked(void) void vector_scatter_32_masked(void)
{ {
/* copy the offsets and values to vectors */ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
HVX_Vector offsetslo = *(HVX_Vector *)word_offsets; HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2]; HVX_Vector *valueslo = (HVX_Vector *)word_values_masked;
HVX_Vector valueslo = *(HVX_Vector *)word_values_masked; HVX_Vector *valueshi = (HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2];
HVX_Vector valueshi = *(HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2]; HVX_Vector *predslo = (HVX_Vector *)word_predicates;
HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates; HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0);
HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
VSCATTER_32_MASKED(predslo, &vtcm.vscatter32, region_len, offsetslo, asm ("r1 = #-1\n\t"
valueslo); "v0 = vmem(%0 + #0)\n\t"
VSCATTER_32_MASKED(predshi, &vtcm.vscatter32, region_len, offsetshi, "q0 = vand(v0, r1)\n\t"
valueshi); "m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"v1 = vmem(%4 + #0)\n\t"
"if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
: : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetslo), "r"(valueslo)
: "r1", "q0", "m0", "q0", "v0", "v1", "memory");
asm ("r1 = #-1\n\t"
"v0 = vmem(%0 + #0)\n\t"
"q0 = vand(v0, r1)\n\t"
"m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"v1 = vmem(%4 + #0)\n\t"
"if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
: : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetshi), "r"(valueshi)
: "r1", "q0", "m0", "q0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16); sync_scatter(vtcm.vscatter32);
} }
/* scatter the 16 bit elements with 32 bit offsets using intrinsics */ /* scatter the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32(void) void vector_scatter_16_32(void)
{ {
HVX_VectorPair offsets; asm ("m0 = %1\n\t"
HVX_Vector values; "v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%2 + #1)\n\t"
/* get the word offsets in a vector pair */ "v2 = vmem(%3 + #0)\n\t"
offsets = *(HVX_VectorPair *)word_offsets; "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
"vscatter(%0, m0, v1:0.w).h = v2\n\t"
/* these values need to be shuffled for the scatter */ : : "r"(vtcm.vscatter16_32), "r"(region_len),
values = *(HVX_Vector *)half_values; "r"(word_offsets), "r"(half_values)
values = VSHUFF_H(values); : "m0", "v0", "v1", "v2", "memory");
VSCATTER_16_32(&vtcm.vscatter16_32, region_len, offsets, values);
sync_scatter(vtcm.vscatter16_32); sync_scatter(vtcm.vscatter16_32);
} }
/* scatter-acc the 16 bit elements with 32 bit offsets using intrinsics */ /* scatter-accumulate the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32_acc(void) void vector_scatter_16_32_acc(void)
{ {
HVX_VectorPair offsets; asm ("m0 = %1\n\t"
HVX_Vector values; "v0 = vmem(%2 + #0)\n\t"
"v1 = vmem(%2 + #1)\n\t"
/* get the word offsets in a vector pair */ "v2 = vmem(%3 + #0)\n\t" \
offsets = *(HVX_VectorPair *)word_offsets; "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
"vscatter(%0, m0, v1:0.w).h += v2\n\t"
/* these values need to be shuffled for the scatter */ : : "r"(vtcm.vscatter16_32), "r"(region_len),
values = *(HVX_Vector *)half_values_acc; "r"(word_offsets), "r"(half_values_acc)
values = VSHUFF_H(values); : "m0", "v0", "v1", "v2", "memory");
VSCATTER_16_32_ACC(&vtcm.vscatter16_32, region_len, offsets, values);
sync_scatter(vtcm.vscatter16_32); sync_scatter(vtcm.vscatter16_32);
} }
/* masked scatter the 16 bit elements with 32 bit offsets using intrinsics */ /* masked scatter the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32_masked(void) void vector_scatter_16_32_masked(void)
{ {
HVX_VectorPair offsets; asm ("r1 = #-1\n\t"
HVX_Vector values; "v0 = vmem(%0 + #0)\n\t"
HVX_Vector pred_reg; "v0.h = vshuff(v0.h)\n\t" /* shuffle the predicates */
"q0 = vand(v0, r1)\n\t"
/* get the word offsets in a vector pair */ "m0 = %2\n\t"
offsets = *(HVX_VectorPair *)word_offsets; "v0 = vmem(%3 + #0)\n\t"
"v1 = vmem(%3 + #1)\n\t"
/* these values need to be shuffled for the scatter */ "v2 = vmem(%4 + #0)\n\t" \
values = *(HVX_Vector *)half_values_masked; "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
values = VSHUFF_H(values); "if (q0) vscatter(%1, m0, v1:0.w).h = v2\n\t"
: : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
pred_reg = *(HVX_Vector *)half_predicates; "r"(word_offsets), "r"(half_values_masked)
pred_reg = VSHUFF_H(pred_reg); : "r1", "q0", "m0", "v0", "v1", "v2", "memory");
HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
VSCATTER_16_32_MASKED(preds, &vtcm.vscatter16_32, region_len, offsets,
values);
sync_scatter(vtcm.vscatter16_32); sync_scatter(vtcm.vscatter16_32);
} }
/* gather the elements from the scatter16 buffer */ /* gather the elements from the scatter16 buffer using HVX */
void vector_gather_16(void) void vector_gather_16(void)
{ {
HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16; asm ("m0 = %1\n\t"
HVX_Vector offsets = *(HVX_Vector *)half_offsets; "v0 = vmem(%2 + #0)\n\t"
"{ vtmp.h = vgather(%0, m0, v0.h).h\n\t"
" vmem(%3 + #0) = vtmp.new }\n\t"
: : "r"(vtcm.vscatter16), "r"(region_len),
"r"(half_offsets), "r"(vtcm.vgather16)
: "m0", "v0", "memory");
VGATHER_16(vgather, &vtcm.vscatter16, region_len, offsets); sync_gather(vtcm.vgather16);
sync_gather(vgather);
} }
static unsigned short gather_16_masked_init(void) static unsigned short gather_16_masked_init(void)
@ -427,31 +423,51 @@ static unsigned short gather_16_masked_init(void)
return letter | (letter << 8); return letter | (letter << 8);
} }
/* masked gather the elements from the scatter16 buffer using HVX */
void vector_gather_16_masked(void) void vector_gather_16_masked(void)
{ {
HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16; unsigned short init = gather_16_masked_init();
HVX_Vector offsets = *(HVX_Vector *)half_offsets;
HVX_Vector pred_reg = *(HVX_Vector *)half_predicates;
HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
*vgather = VSPLAT_H(gather_16_masked_init()); asm ("v0.h = vsplat(%5)\n\t"
VGATHER_16_MASKED(vgather, preds, &vtcm.vscatter16, region_len, offsets); "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
"r1 = #-1\n\t"
"v0 = vmem(%0 + #0)\n\t"
"q0 = vand(v0, r1)\n\t"
"m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"{ if (q0) vtmp.h = vgather(%1, m0, v0.h).h\n\t"
" vmem(%4 + #0) = vtmp.new }\n\t"
: : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
"r"(half_offsets), "r"(vtcm.vgather16), "r"(init)
: "r1", "q0", "m0", "v0", "memory");
sync_gather(vgather); sync_gather(vtcm.vgather16);
} }
/* gather the elements from the scatter32 buffer */ /* gather the elements from the scatter32 buffer using HVX */
void vector_gather_32(void) void vector_gather_32(void)
{ {
HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32; HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
HVX_Vector *vgatherhi = HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
(HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2)); HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
HVX_Vector offsetslo = *(HVX_Vector *)word_offsets; HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
VGATHER_32(vgatherlo, &vtcm.vscatter32, region_len, offsetslo); asm ("m0 = %1\n\t"
VGATHER_32(vgatherhi, &vtcm.vscatter32, region_len, offsetshi); "v0 = vmem(%2 + #0)\n\t"
"{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
" vmem(%3 + #0) = vtmp.new }\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetslo), "r"(vgatherlo)
: "m0", "v0", "memory");
asm ("m0 = %1\n\t"
"v0 = vmem(%2 + #0)\n\t"
"{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
" vmem(%3 + #0) = vtmp.new }\n\t"
: : "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetshi), "r"(vgatherhi)
: "m0", "v0", "memory");
sync_gather(vgatherlo);
sync_gather(vgatherhi); sync_gather(vgatherhi);
} }
@ -461,79 +477,88 @@ static unsigned int gather_32_masked_init(void)
return letter | (letter << 8) | (letter << 16) | (letter << 24); return letter | (letter << 8) | (letter << 16) | (letter << 24);
} }
/* masked gather the elements from the scatter32 buffer using HVX */
void vector_gather_32_masked(void) void vector_gather_32_masked(void)
{ {
HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32; unsigned int init = gather_32_masked_init();
HVX_Vector *vgatherhi = HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
(HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2)); HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
HVX_Vector offsetslo = *(HVX_Vector *)word_offsets; HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2]; HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates; HVX_Vector *predslo = (HVX_Vector *)word_predicates;
HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0); HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
*vgatherlo = VSPLAT_H(gather_32_masked_init()); asm ("v0.h = vsplat(%5)\n\t"
*vgatherhi = VSPLAT_H(gather_32_masked_init()); "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
VGATHER_32_MASKED(vgatherlo, predslo, &vtcm.vscatter32, region_len, "r1 = #-1\n\t"
offsetslo); "v0 = vmem(%0 + #0)\n\t"
VGATHER_32_MASKED(vgatherhi, predshi, &vtcm.vscatter32, region_len, "q0 = vand(v0, r1)\n\t"
offsetshi); "m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
" vmem(%4 + #0) = vtmp.new }\n\t"
: : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetslo), "r"(vgatherlo), "r"(init)
: "r1", "q0", "m0", "v0", "memory");
asm ("v0.h = vsplat(%5)\n\t"
"vmem(%4 + #0) = v0\n\t" /* initialize the write area */
"r1 = #-1\n\t"
"v0 = vmem(%0 + #0)\n\t"
"q0 = vand(v0, r1)\n\t"
"m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
" vmem(%4 + #0) = vtmp.new }\n\t"
: : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
"r"(offsetshi), "r"(vgatherhi), "r"(init)
: "r1", "q0", "m0", "v0", "memory");
sync_gather(vgatherlo); sync_gather(vgatherlo);
sync_gather(vgatherhi); sync_gather(vgatherhi);
} }
/* gather the elements from the scatter16_32 buffer */ /* gather the elements from the scatter16_32 buffer using HVX */
void vector_gather_16_32(void) void vector_gather_16_32(void)
{ {
HVX_Vector *vgather; asm ("m0 = %1\n\t"
HVX_VectorPair offsets; "v0 = vmem(%2 + #0)\n\t"
HVX_Vector values; "v1 = vmem(%2 + #1)\n\t"
"{ vtmp.h = vgather(%0, m0, v1:0.w).h\n\t"
" vmem(%3 + #0) = vtmp.new }\n\t"
"v0 = vmem(%3 + #0)\n\t"
"v0.h = vdeal(v0.h)\n\t" /* deal the elements to get the order back */
"vmem(%3 + #0) = v0\n\t"
: : "r"(vtcm.vscatter16_32), "r"(region_len),
"r"(word_offsets), "r"(vtcm.vgather16_32)
: "m0", "v0", "v1", "memory");
/* get the vtcm address to gather from */ sync_gather(vtcm.vgather16_32);
vgather = (HVX_Vector *)&vtcm.vgather16_32;
/* get the word offsets in a vector pair */
offsets = *(HVX_VectorPair *)word_offsets;
VGATHER_16_32(vgather, &vtcm.vscatter16_32, region_len, offsets);
/* deal the elements to get the order back */
values = *(HVX_Vector *)vgather;
values = VDEAL_H(values);
/* write it back to vtcm address */
*(HVX_Vector *)vgather = values;
} }
/* masked gather the elements from the scatter16_32 buffer using HVX */
void vector_gather_16_32_masked(void) void vector_gather_16_32_masked(void)
{ {
HVX_Vector *vgather; unsigned short init = gather_16_masked_init();
HVX_VectorPair offsets;
HVX_Vector pred_reg;
HVX_VectorPred preds;
HVX_Vector values;
/* get the vtcm address to gather from */ asm ("v0.h = vsplat(%5)\n\t"
vgather = (HVX_Vector *)&vtcm.vgather16_32; "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
"r1 = #-1\n\t"
"v0 = vmem(%0 + #0)\n\t"
"v0.h = vshuff(v0.h)\n\t" /* shuffle the predicates */
"q0 = vand(v0, r1)\n\t"
"m0 = %2\n\t"
"v0 = vmem(%3 + #0)\n\t"
"v1 = vmem(%3 + #1)\n\t"
"{ if (q0) vtmp.h = vgather(%1, m0, v1:0.w).h\n\t"
" vmem(%4 + #0) = vtmp.new }\n\t"
"v0 = vmem(%4 + #0)\n\t"
"v0.h = vdeal(v0.h)\n\t" /* deal the elements to get the order back */
"vmem(%4 + #0) = v0\n\t"
: : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
"r"(word_offsets), "r"(vtcm.vgather16_32), "r"(init)
: "r1", "q0", "m0", "v0", "v1", "memory");
/* get the word offsets in a vector pair */ sync_gather(vtcm.vgather16_32);
offsets = *(HVX_VectorPair *)word_offsets;
pred_reg = *(HVX_Vector *)half_predicates;
pred_reg = VSHUFF_H(pred_reg);
preds = VAND_VAL(pred_reg, ~0);
*vgather = VSPLAT_H(gather_16_masked_init());
VGATHER_16_32_MASKED(vgather, preds, &vtcm.vscatter16_32, region_len,
offsets);
/* deal the elements to get the order back */
values = *(HVX_Vector *)vgather;
values = VDEAL_H(values);
/* write it back to vtcm address */
*(HVX_Vector *)vgather = values;
} }
static void check_buffer(const char *name, void *c, void *r, size_t size) static void check_buffer(const char *name, void *c, void *r, size_t size)
@ -579,6 +604,7 @@ void scalar_scatter_16_acc(unsigned short *vscatter16)
} }
} }
/* scatter-accumulate the 16 bit elements using C */
void check_scatter_16_acc() void check_scatter_16_acc()
{ {
memset(vscatter16_ref, FILL_CHAR, memset(vscatter16_ref, FILL_CHAR,
@ -589,7 +615,7 @@ void check_scatter_16_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned short)); SCATTER_BUFFER_SIZE * sizeof(unsigned short));
} }
/* scatter the 16 bit elements using C */ /* masked scatter the 16 bit elements using C */
void scalar_scatter_16_masked(unsigned short *vscatter16) void scalar_scatter_16_masked(unsigned short *vscatter16)
{ {
for (int i = 0; i < MATRIX_SIZE; i++) { for (int i = 0; i < MATRIX_SIZE; i++) {
@ -628,7 +654,7 @@ void check_scatter_32()
SCATTER_BUFFER_SIZE * sizeof(unsigned int)); SCATTER_BUFFER_SIZE * sizeof(unsigned int));
} }
/* scatter the 32 bit elements using C */ /* scatter-accumulate the 32 bit elements using C */
void scalar_scatter_32_acc(unsigned int *vscatter32) void scalar_scatter_32_acc(unsigned int *vscatter32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -646,7 +672,7 @@ void check_scatter_32_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned int)); SCATTER_BUFFER_SIZE * sizeof(unsigned int));
} }
/* scatter the 32 bit elements using C */ /* masked scatter the 32 bit elements using C */
void scalar_scatter_32_masked(unsigned int *vscatter32) void scalar_scatter_32_masked(unsigned int *vscatter32)
{ {
for (int i = 0; i < MATRIX_SIZE; i++) { for (int i = 0; i < MATRIX_SIZE; i++) {
@ -667,7 +693,7 @@ void check_scatter_32_masked()
SCATTER_BUFFER_SIZE * sizeof(unsigned int)); SCATTER_BUFFER_SIZE * sizeof(unsigned int));
} }
/* scatter the 32 bit elements using C */ /* scatter the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32(unsigned short *vscatter16_32) void scalar_scatter_16_32(unsigned short *vscatter16_32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -684,7 +710,7 @@ void check_scatter_16_32()
SCATTER_BUFFER_SIZE * sizeof(unsigned short)); SCATTER_BUFFER_SIZE * sizeof(unsigned short));
} }
/* scatter the 32 bit elements using C */ /* scatter-accumulate the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32_acc(unsigned short *vscatter16_32) void scalar_scatter_16_32_acc(unsigned short *vscatter16_32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -702,6 +728,7 @@ void check_scatter_16_32_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned short)); SCATTER_BUFFER_SIZE * sizeof(unsigned short));
} }
/* masked scatter the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32_masked(unsigned short *vscatter16_32) void scalar_scatter_16_32_masked(unsigned short *vscatter16_32)
{ {
for (int i = 0; i < MATRIX_SIZE; i++) { for (int i = 0; i < MATRIX_SIZE; i++) {
@ -738,6 +765,7 @@ void check_gather_16()
MATRIX_SIZE * sizeof(unsigned short)); MATRIX_SIZE * sizeof(unsigned short));
} }
/* masked gather the elements from the scatter buffer using C */
void scalar_gather_16_masked(unsigned short *vgather16) void scalar_gather_16_masked(unsigned short *vgather16)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -756,7 +784,7 @@ void check_gather_16_masked()
MATRIX_SIZE * sizeof(unsigned short)); MATRIX_SIZE * sizeof(unsigned short));
} }
/* gather the elements from the scatter buffer using C */ /* gather the elements from the scatter32 buffer using C */
void scalar_gather_32(unsigned int *vgather32) void scalar_gather_32(unsigned int *vgather32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -772,6 +800,7 @@ void check_gather_32(void)
MATRIX_SIZE * sizeof(unsigned int)); MATRIX_SIZE * sizeof(unsigned int));
} }
/* masked gather the elements from the scatter32 buffer using C */
void scalar_gather_32_masked(unsigned int *vgather32) void scalar_gather_32_masked(unsigned int *vgather32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -781,7 +810,6 @@ void scalar_gather_32_masked(unsigned int *vgather32)
} }
} }
void check_gather_32_masked(void) void check_gather_32_masked(void)
{ {
memset(vgather32_ref, gather_32_masked_init(), memset(vgather32_ref, gather_32_masked_init(),
@ -791,7 +819,7 @@ void check_gather_32_masked(void)
vgather32_ref, MATRIX_SIZE * sizeof(unsigned int)); vgather32_ref, MATRIX_SIZE * sizeof(unsigned int));
} }
/* gather the elements from the scatter buffer using C */ /* gather the elements from the scatter16_32 buffer using C */
void scalar_gather_16_32(unsigned short *vgather16_32) void scalar_gather_16_32(unsigned short *vgather16_32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {
@ -807,6 +835,7 @@ void check_gather_16_32(void)
MATRIX_SIZE * sizeof(unsigned short)); MATRIX_SIZE * sizeof(unsigned short));
} }
/* masked gather the elements from the scatter16_32 buffer using C */
void scalar_gather_16_32_masked(unsigned short *vgather16_32) void scalar_gather_16_32_masked(unsigned short *vgather16_32)
{ {
for (int i = 0; i < MATRIX_SIZE; ++i) { for (int i = 0; i < MATRIX_SIZE; ++i) {